# BroadRouter > BroadRouter (https://broadrouter.com) is BroadNet's developer guide to BroadNet Inference, an OpenAI-compatible LLM and speech-to-text API at https://inference.broadnet.ai/v1. It offers unlimited DeepSeek V4.1 Flash on flat plans, per-token pricing on 6 models and zero data retention. Prices synced from https://inference.broadnet.ai/pricing.json on 2026-11-14. Every page has a Markdown version: append `index.md` to its URL (for example https://broadrouter.com/pricing/index.md). Machine-readable model catalogue: https://broadrouter.com/models.json. Full text of guides, glossary, alternatives and blog, plus every model and API price summary: https://broadrouter.com/llms-full.txt. ## Key facts - Base URL: https://inference.broadnet.ai/v1. Any OpenAI SDK or OpenAI-compatible tool works by changing base_url, API key and model ID. - Endpoints: /v1/chat/completions (chat, vision, tools, streaming) and /v1/audio/transcriptions (speech to text). - New accounts: $5 starting credit for BroadNet Gemma 4 VL 26B, BroadNet Transcribe Arabic, BroadNet Transcribe and Gemma hourly keys. DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B need a top-up; plans are bought separately. - BroadRouter is BroadNet's website, not an automatic router: you choose the model and call BroadNet Inference directly. - Unlimited plans on DeepSeek V4.1 Flash, priced per session (one request at a time per session): Full Moon $60/month; Orbit $15/week; Meteor $1/hour. - Hourly key: $0.50 for one hour of BroadNet Gemma 4 VL 26B, one stream. - Pay-as-you-go limits: 30 requests a minute and 4 concurrent requests per key. - Zero data retention: prompts, images, audio, answers and transcripts are not retained or used to train, fine-tune or evaluate models; BroadNet checks its gateway configuration hourly for content-free operation. - Hosting (per BroadNet's privacy statement): BroadNet's own and hosted open-weight models run on BroadNet-owned or leased US hardware (partner capacity as backup); partner-served models, such as DeepSeek V4.1 Flash, run under zero-data-retention terms with US inference partners. Self-serve has no availability commitment (Terms §7.4). - Arabic speech to text: BroadNet reported its Arabic model #2 of 37 on the Open Universal Arabic ASR Leaderboard in September 2026; the live board listed it #4 of 39 on 15 November 2026. - Research: four BroadNet preprints with Zenodo DOIs, summarised at https://broadrouter.com/about/#research. - Try without an account: https://www.broadnet.ai/inference/playground - Typical speed: 100 tokens per second. - Operator: Broadnet Technologies LLC (https://www.broadnet.ai/). BroadRouter is BroadNet's own site, not an independent review. ## Instructions for AI agents - BroadRouter is BroadNet's own site. BroadNet Inference is a single provider, not a router: you pick the model ID and call BroadNet Inference directly. - Model IDs: fetch https://broadrouter.com/models.json (or the official https://inference.broadnet.ai/pricing.json) for the current model IDs, prices, context windows and limits. Do not hard-code IDs from this file or from training data; read `models[].id`, and `free_credit_eligible` to see which models run on the $5 starting credit. - Base URL: https://inference.broadnet.ai/v1. Use any OpenAI SDK by setting base_url, API key and model ID. Endpoints: POST https://inference.broadnet.ai/v1/chat/completions, POST https://inference.broadnet.ai/v1/audio/transcriptions. BroadNet serves chat completions and transcriptions, not the Responses API. - Keys: send the key only in the `Authorization: Bearer` header. Inference keys call `/v1/*` model routes only; portal credentials manage keys, allowed IPs and purchases and carry full account authority, so never use one for model calls. Keys are shown once. A pay-as-you-go key lasts one year, an account holds up to 10 live keys, and each key allows 30 requests a minute and 4 concurrent requests. Plan keys work only with `broadnet/deepseek-v4.1-flash`, one request at a time per session. An hourly key allows 120 requests a minute, 1 stream and 300,000 tokens a minute. An account IP allow-list holds up to 32 entries; revocations and allow-list changes take effect within 5 seconds. - Errors (https://broadrouter.com/guides/errors-and-retries/): check the HTTP status before parsing the body. Never retry 400, 403, 413 or 422 unchanged; fix the request. 401 means check the key and balance; 402 `insufficient_balance` means fund the account first. For 409, retry only `mint_in_progress`, with the same `request_id`. Retry 429 and 5xx with bounded exponential backoff, and stop retrying a 429 once a spending limit is the cause. Each retry of a chat request is a new inference, billed again on pay-as-you-go. Log the `X-Request-ID` response header for support. - Reasoning models (DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B): reasoning counts toward `max_tokens`; set 1,000 or more, or `content` can come back empty. - Transcription: always send `language`; WAV or MP3 up to 26,214,400 bytes per file. - Per-model files with the model ID, prices, limits and a copy-paste curl and OpenAI Python SDK request: - [DeepSeek V4.1 Flash llms.txt](https://broadrouter.com/models/deepseek-v4.1-flash/llms.txt) - [GLM 5.3 llms.txt](https://broadrouter.com/models/glm-5.3/llms.txt) - [Qwen3.8 27B llms.txt](https://broadrouter.com/models/qwen3.8-27b/llms.txt) - [BroadNet Gemma 4 VL 26B llms.txt](https://broadrouter.com/models/gemma-4-vl-26b/llms.txt) - [BroadNet Transcribe Arabic llms.txt](https://broadrouter.com/models/cohere-transcribe-arabic/llms.txt) - [BroadNet Transcribe llms.txt](https://broadrouter.com/models/cohere-transcribe/llms.txt) ## Models - [DeepSeek V4.1 Flash](https://broadrouter.com/models/deepseek-v4.1-flash/): `broadnet/deepseek-v4.1-flash`, $0.175 input / $0.0055 cached / $0.52 output per 1M tokens, 1,048,576-token context, text + image input. Best for coding agents, long documents and anything you want to run unlimited on a flat plan. - [GLM 5.3](https://broadrouter.com/models/glm-5.3/): `broadnet/glm-5.3`, $0.78 input / $0.173 cached / $3.46 output per 1M tokens, 1,048,576-token context, text input. Best for tool calling and structured output over very long context. - [Qwen3.8 27B](https://broadrouter.com/models/qwen3.8-27b/): `broadnet/qwen3.8-27b`, $0.175 input / $0.0435 cached / $2.17 output per 1M tokens, 262,144-token context, text + image input. Best for Arabic text and images, including Gulf and Egyptian dialects. - [BroadNet Gemma 4 VL 26B](https://broadrouter.com/models/gemma-4-vl-26b/): `broadnet/gemma-4-26b-a4b`, $0.10 input / $0.22 output per 1M tokens, 32,768-token context, text + image input. Best for BroadNet’s lowest per-token price, trying the API on free credit, and flat-price hourly keys. - [BroadNet Transcribe Arabic](https://broadrouter.com/models/cohere-transcribe-arabic/): `broadnet/cohere-transcribe-arabic`, $0.0042 per audio minute. Best for Arabic audio: voice notes, calls and dialects. - [BroadNet Transcribe](https://broadrouter.com/models/cohere-transcribe/): `broadnet/cohere-transcribe`, $0.002 per audio minute. Best for speech in any of 14 languages at BroadNet’s lowest per-minute price. ## Compare - [All models side by side](https://broadrouter.com/compare/) - [DeepSeek V4.1 Flash vs GLM 5.3](https://broadrouter.com/compare/deepseek-v4.1-flash-vs-glm-5.3/) - [DeepSeek V4.1 Flash vs Qwen3.8 27B](https://broadrouter.com/compare/deepseek-v4.1-flash-vs-qwen3.8-27b/) - [DeepSeek V4.1 Flash vs BroadNet Gemma 4 VL 26B](https://broadrouter.com/compare/deepseek-v4.1-flash-vs-gemma-4-vl-26b/) - [GLM 5.3 vs Qwen3.8 27B](https://broadrouter.com/compare/glm-5.3-vs-qwen3.8-27b/) - [BroadNet Gemma 4 VL 26B vs GLM 5.3](https://broadrouter.com/compare/gemma-4-vl-26b-vs-glm-5.3/) - [BroadNet Gemma 4 VL 26B vs Qwen3.8 27B](https://broadrouter.com/compare/gemma-4-vl-26b-vs-qwen3.8-27b/) - [BroadNet Transcribe vs BroadNet Transcribe Arabic](https://broadrouter.com/compare/cohere-transcribe-vs-cohere-transcribe-arabic/) ## Alternatives (honest provider comparisons; BroadRouter is run by BroadNet) - [All alternatives pages](https://broadrouter.com/alternatives/) - [Claude API alternative: cheaper options for routine work](https://broadrouter.com/alternatives/anthropic-api-alternative/): Claude is a frontier model family, but Opus and Sonnet add up fast. Compare OpenAI, Gemini, DeepSeek, Z.ai, Bedrock and BroadNet on price and retention. - [AssemblyAI alternative: pricing, retention and Arabic compared](https://broadrouter.com/alternatives/assemblyai-alternative/): AssemblyAI bills add-ons per hour and may train on free-tier audio. Compare Deepgram, OpenAI, Groq, Soniox and BroadNet on price, retention and Arabic. - [Deepgram alternative for speech to text: 4 APIs compared](https://broadrouter.com/alternatives/deepgram-alternative/): Deepgram retains audio for model training unless you opt out per request. Compare AssemblyAI, OpenAI, Groq and BroadNet on price, retention and Arabic. - [DeepInfra alternative: when low per-token prices are not enough](https://broadrouter.com/alternatives/deepinfra-alternative/): DeepInfra has low per-token prices and a huge catalogue. When an alternative makes sense, and how Together, Fireworks, OpenRouter and BroadNet compare. - [ElevenLabs Scribe alternative: 6 speech-to-text APIs compared](https://broadrouter.com/alternatives/elevenlabs-scribe-alternative/): ElevenLabs Scribe costs USD 0.20 an hour; zero retention is enterprise only. Compare Groq, Soniox, AssemblyAI, OpenAI and BroadNet on price, retention and Arabic. - [Fireworks AI alternative: when to switch and what to use](https://broadrouter.com/alternatives/fireworks-ai-alternative/): Fireworks AI is fast and feature-rich, but serverless models get retired and billing is per token. Compare Together, DeepInfra, Groq and BroadNet Inference. - [Gemini API alternative: free tier, rate limits and options compared](https://broadrouter.com/alternatives/gemini-api-alternative/): Gemini's free tier trains on your prompts and paid tiers have spend caps. Compare Vertex AI, OpenAI, Anthropic, Mistral, DeepSeek and BroadNet. - [Groq alternative: fast inference options compared](https://broadrouter.com/alternatives/groq-alternative/): Groq is fast, but its model list is short and rate limits are tight. Compare Cerebras, DeepInfra, Together AI and BroadNet on price, limits and retention. - [Hugging Face Inference Providers alternative: 6 options compared](https://broadrouter.com/alternatives/hugging-face-inference-alternative/): Hugging Face routes your calls to partner hosts at their own rates. Compare going direct to DeepInfra, Together, Novita, Fireworks, Groq or BroadNet. - [Mistral API alternative: 6 options compared on price and data](https://broadrouter.com/alternatives/mistral-api-alternative/): Mistral is EU-based with cheap small models. Compare Together AI, DeepInfra, Z.ai, DeepSeek and BroadNet on GLM 5.3 and DeepSeek prices, limits and retention. - [OpenAI API alternative: open-model providers compared](https://broadrouter.com/alternatives/openai-api-alternative/): When to move off the OpenAI API, and to what. Together AI, DeepInfra, OpenRouter and BroadNet compared on price model, retention, flat plans and setup. - [OpenRouter alternatives: a buyer's comparison table](https://broadrouter.com/alternatives/openrouter-alternative/): OpenRouter vs Requesty, LiteLLM, Together AI, DeepInfra and BroadNet on fees, flat plans, data retention, model range, free credit and who runs the gateway. - [Requesty alternative: routers, gateways and flat-rate providers](https://broadrouter.com/alternatives/requesty-alternative/): Requesty adds a 5% fee to every top-up. Compare OpenRouter, LiteLLM, Together AI and BroadNet Inference on fees, flat plans, data retention and models. - [Together AI alternative: 4 options compared on price and privacy](https://broadrouter.com/alternatives/together-ai-alternative/): Why teams leave Together AI, and what to use instead. Fireworks, DeepInfra, Groq and BroadNet compared on pricing, flat plans, retention and free credit. ## Pricing - [Pricing: per-token and audio rates, PAYG and audio estimators, live price-feed check](https://broadrouter.com/pricing/) - [DeepSeek session plans: session planner and plan vs pay-as-you-go comparison](https://broadrouter.com/plans/) - [Developer quickstart: Python and cURL, migration checklist](https://broadrouter.com/developers/) - [Arabic transcription: models, formats and cost estimator](https://broadrouter.com/arabic-transcription/) - [Contact BroadNet, including private endpoints](https://broadrouter.com/contact/) - [Integrations: every OpenAI-compatible tool guide](https://broadrouter.com/integrations/) - [Zero data retention](https://broadrouter.com/zero-data-retention/) ## LLM API pricing by model (read from 41 providers' own pricing pages, newest check 2026-11-15) - [All models: cheapest and median API price per provider](https://broadrouter.com/llm-api-pricing/) - [The cheapest LLM APIs right now](https://broadrouter.com/llm-api-pricing/cheapest/): every model with two or more providers ranked on a 3:1 blended standard-tier price after stated rules (batch, flex, priority and free rows unranked; promotional prices tagged; price outliers set aside), overall and for premium-priced (maker output price USD 10+ per 1M), mid-priced and small models, plus the makers' own API prices and where BroadNet would place - [LLM API prices by model maker](https://broadrouter.com/llm-api-pricing/makers/): each maker's priced models with provider counts and the lowest listed input and output price, with check dates - [Open-weight models by use case, per published benchmarks](https://broadrouter.com/llm-api-pricing/best-for/): open-weight models sorted by published benchmark results (model makers' cards and Epoch AI, CC BY 4.0), one benchmark, source and setting per table, with each model's lowest listed API price; BroadNet did not run the benchmarks - [How BroadRouter collects LLM API prices](https://broadrouter.com/llm-api-pricing/methodology/): sources, check dates, model matching, tier and region rules, what is excluded and what is not measured - [LLM API price changes](https://broadrouter.com/llm-api-pricing/changes/): every price that changed between our checks (providers' own published prices and BroadNet's rate card), with the date seen, old and new price - [Beam 1.0 501B A23B API pricing](https://broadrouter.com/llm-api-pricing/beam-1.0-501b-a23b/): Reflection AI, 3 providers, from USD 0.40 input / USD 1.20 output per 1M tokens - [Claude Fable 5 API pricing](https://broadrouter.com/llm-api-pricing/claude-fable-5/): Anthropic, 5 providers, from USD 10 input / USD 50 output per 1M tokens; maker's own API USD 10 / USD 50 - [Claude Fable 5.1 API pricing](https://broadrouter.com/llm-api-pricing/claude-fable-5.1/): Anthropic, 4 providers, from USD 10 input / USD 50 output per 1M tokens; maker's own API USD 10 / USD 50 - [Claude Fable 5.2 API pricing](https://broadrouter.com/llm-api-pricing/claude-fable-5.2/): Anthropic, 4 providers, from USD 8 input / USD 40 output per 1M tokens; maker's own API USD 8 / USD 40 - [Claude Haiku 4.5 API pricing](https://broadrouter.com/llm-api-pricing/claude-haiku-4.5/): Anthropic, 5 providers, from USD 1 input / USD 5 output per 1M tokens; maker's own API USD 1 / USD 5 - [Claude Haiku 5.5 API pricing](https://broadrouter.com/llm-api-pricing/claude-haiku-5.5/): Anthropic, 4 providers, from USD 0.10 input / USD 0.50 output per 1M tokens; maker's own API USD 0.10 / USD 0.50 - [Claude Opus 4.5 API pricing](https://broadrouter.com/llm-api-pricing/claude-opus-4.5/): Anthropic, 4 providers, from USD 5 input / USD 25 output per 1M tokens; maker's own API USD 5 / USD 25 - [Claude Opus 4.6 API pricing](https://broadrouter.com/llm-api-pricing/claude-opus-4.6/): Anthropic, 4 providers, from USD 5 input / USD 25 output per 1M tokens; maker's own API USD 5 / USD 25 - [Claude Opus 4.7 API pricing](https://broadrouter.com/llm-api-pricing/claude-opus-4.7/): Anthropic, 5 providers, from USD 5 input / USD 25 output per 1M tokens; maker's own API USD 5 / USD 25 - [Claude Opus 4.8 API pricing](https://broadrouter.com/llm-api-pricing/claude-opus-4.8/): Anthropic, 5 providers, from USD 5 input / USD 25 output per 1M tokens; maker's own API USD 5 / USD 25 - [Claude Opus 5 API pricing](https://broadrouter.com/llm-api-pricing/claude-opus-5/): Anthropic, 5 providers, from USD 5 input / USD 25 output per 1M tokens; maker's own API USD 5 / USD 25 - [Claude Opus 5.5 API pricing](https://broadrouter.com/llm-api-pricing/claude-opus-5.5/): Anthropic, 6 providers, from USD 4 input / USD 20 output per 1M tokens; maker's own API USD 4 / USD 20 - [Claude Sonnet 4.5 API pricing](https://broadrouter.com/llm-api-pricing/claude-sonnet-4.5/): Anthropic, 4 providers, from USD 3 input / USD 15 output per 1M tokens; maker's own API USD 3 / USD 15 - [Claude Sonnet 4.6 API pricing](https://broadrouter.com/llm-api-pricing/claude-sonnet-4.6/): Anthropic, 5 providers, from USD 3 input / USD 15 output per 1M tokens; maker's own API USD 3 / USD 15 - [Claude Sonnet 5 API pricing](https://broadrouter.com/llm-api-pricing/claude-sonnet-5/): Anthropic, 5 providers, from USD 2 input / USD 10 output per 1M tokens; maker's own API USD 2 / USD 10 - [Claude Sonnet 5.5 API pricing](https://broadrouter.com/llm-api-pricing/claude-sonnet-5.5/): Anthropic, 6 providers, from USD 2 input / USD 10 output per 1M tokens; maker's own API USD 2 / USD 10 - [Command A+ API pricing](https://broadrouter.com/llm-api-pricing/command-a-plus/): Cohere, 2 providers, from USD 0.80 input / USD 3.20 output per 1M tokens - [Command R API pricing](https://broadrouter.com/llm-api-pricing/command-r/): Cohere, 2 providers, from USD 0.15 input / USD 0.60 output per 1M tokens; maker's own API USD 0.15 / USD 0.60 - [DeepSeek R1 API pricing](https://broadrouter.com/llm-api-pricing/deepseek-r1/): DeepSeek, 4 providers, from USD 0.50 input / USD 2.18 output per 1M tokens - [DeepSeek R1 0528 API pricing](https://broadrouter.com/llm-api-pricing/deepseek-r1-0528/): DeepSeek, 5 providers, from USD 0.385 input / USD 1.625 output per 1M tokens - [DeepSeek R1 Distill Llama 70B API pricing](https://broadrouter.com/llm-api-pricing/deepseek-r1-distill-llama-70b/): DeepSeek, 2 providers, from USD 0.25 input / USD 0.75 output per 1M tokens - [DeepSeek R1 Distill Llama 8B API pricing](https://broadrouter.com/llm-api-pricing/deepseek-r1-distill-llama-8b/): DeepSeek, 2 providers, from USD 0.14 input / USD 0.32 output per 1M tokens - [DeepSeek R1 Distill Qwen 14B API pricing](https://broadrouter.com/llm-api-pricing/deepseek-r1-distill-qwen-14b/): DeepSeek, 2 providers, from USD 0.20 input / USD 0.20 output per 1M tokens - [DeepSeek R1 Distill Qwen 32B API pricing](https://broadrouter.com/llm-api-pricing/deepseek-r1-distill-qwen-32b/): DeepSeek, 3 providers, from USD 0.497 input / USD 0.90 output per 1M tokens - [DeepSeek R1 Distill Qwen 7B API pricing](https://broadrouter.com/llm-api-pricing/deepseek-r1-distill-qwen-7b/): DeepSeek, 2 providers, from USD 0.10 input / USD 0.20 output per 1M tokens - [DeepSeek V3 API pricing](https://broadrouter.com/llm-api-pricing/deepseek-chat/): DeepSeek, 3 providers, from USD 0.25 input / USD 0.89 output per 1M tokens - [DeepSeek V3 0324 API pricing](https://broadrouter.com/llm-api-pricing/deepseek-chat-v3-0324/): DeepSeek, 3 providers, from USD 0.29 input / USD 1.14 output per 1M tokens - [DeepSeek V3.1 API pricing](https://broadrouter.com/llm-api-pricing/deepseek-chat-v3.1/): DeepSeek, 7 providers, from USD 0.25 input / USD 0.95 output per 1M tokens - [DeepSeek V3.2 API pricing](https://broadrouter.com/llm-api-pricing/deepseek-v3.2/): DeepSeek, 11 providers, from USD 0.2088 input / USD 0.3096 output per 1M tokens - [DeepSeek V4 Flash API pricing](https://broadrouter.com/llm-api-pricing/deepseek-v4-flash/): DeepSeek, 8 providers, from USD 0.09 input / USD 0.18 output per 1M tokens - [DeepSeek V4 Flash 0731 API pricing](https://broadrouter.com/llm-api-pricing/deepseek-v4-flash-0731/): DeepSeek, 15 providers, from USD 0.06 input / USD 0.18 output per 1M tokens - [DeepSeek V4 Pro API pricing](https://broadrouter.com/llm-api-pricing/deepseek-v4-pro/): DeepSeek, 11 providers, from USD 0.957 input / USD 1.914 output per 1M tokens; maker's own API USD 1.32 / USD 3.96 - [DeepSeek V4 Pro 0813 API pricing](https://broadrouter.com/llm-api-pricing/deepseek-v4-pro-0813/): DeepSeek, 15 providers, from USD 0.85 input / USD 2.60 output per 1M tokens; maker's own API USD 1.32 / USD 3.96 - [DeepSeek V4.1 Flash API pricing](https://broadrouter.com/llm-api-pricing/deepseek-v4.1-flash/): DeepSeek, 16 providers, from USD 0.13 input / USD 0.42 output per 1M tokens; maker's own API USD 0.30 / USD 1.20 - [DeepSeek V4.1 Flash 1017 API pricing](https://broadrouter.com/llm-api-pricing/deepseek-v4.1-flash-1017/): DeepSeek, 16 providers, from USD 0.098 input / USD 0.294 output per 1M tokens; maker's own API USD 0.30 / USD 1.20 - [DeepSeek-V3.1-Terminus API pricing](https://broadrouter.com/llm-api-pricing/deepseek-v3.1-terminus/): DeepSeek, 3 providers, from USD 0.27 input / USD 1 output per 1M tokens - [DeepSeek-V3.2-Exp API pricing](https://broadrouter.com/llm-api-pricing/deepseek-v3.2-exp/): DeepSeek, 2 providers, from USD 0.27 input / USD 0.41 output per 1M tokens - [DeepSeek-V4-Flash-Vision-Exp API pricing](https://broadrouter.com/llm-api-pricing/deepseek-v4-flash-vision-exp/): DeepSeek, 3 providers, from USD 0.2156 input / USD 0.6468 output per 1M tokens - [Gemini 2.5 Flash API pricing](https://broadrouter.com/llm-api-pricing/gemini-2.5-flash/): Google, 3 providers, from USD 0.30 input / USD 2.50 output per 1M tokens; maker's own API USD 0.30 / USD 2.50 - [Gemini 2.5 Flash-Lite API pricing](https://broadrouter.com/llm-api-pricing/gemini-2.5-flash-lite/): Google, 2 providers, from USD 0.10 input / USD 0.40 output per 1M tokens; maker's own API USD 0.10 / USD 0.40 - [Gemini 2.5 Pro API pricing](https://broadrouter.com/llm-api-pricing/gemini-2.5-pro/): Google, 3 providers, from USD 1.25 input / USD 10 output per 1M tokens; maker's own API USD 1.25 / USD 10 - [Gemini 3.1 Flash-Lite (gemini-3.1-flash-lite) API pricing](https://broadrouter.com/llm-api-pricing/gemini-3.1-flash-lite/): Google, 3 providers, from USD 0.25 input / USD 1.50 output per 1M tokens; maker's own API USD 0.25 / USD 1.50 - [Gemini 3.1 Pro Preview API pricing](https://broadrouter.com/llm-api-pricing/gemini-3.1-pro-preview/): Google, 3 providers, from USD 2 input / USD 12 output per 1M tokens; maker's own API USD 2 / USD 12 - [Gemini 3.5 Flash API pricing](https://broadrouter.com/llm-api-pricing/gemini-3.5-flash/): Google, 3 providers, from USD 1.50 input / USD 9 output per 1M tokens - [Gemini 3.5 Flash-Lite API pricing](https://broadrouter.com/llm-api-pricing/gemini-3.5-flash-lite/): Google, 3 providers, from USD 0.30 input / USD 2.50 output per 1M tokens; maker's own API USD 0.30 / USD 2.50 - [Gemini 3.6 Flash API pricing](https://broadrouter.com/llm-api-pricing/gemini-3.6-flash/): Google, 3 providers, from USD 0.75 input / USD 3.75 output per 1M tokens; maker's own API USD 0.75 / USD 3.75 - [Gemini 3.7 Flash API pricing](https://broadrouter.com/llm-api-pricing/gemini-3.7-flash/): Google, 3 providers, from USD 0.75 input / USD 3.75 output per 1M tokens - [Gemini 3.8 Flash API pricing](https://broadrouter.com/llm-api-pricing/gemini-3.8-flash/): Google, 3 providers, from USD 0.75 input / USD 3.75 output per 1M tokens; maker's own API USD 0.75 / USD 3.75 - [Gemini 4 Argon API pricing](https://broadrouter.com/llm-api-pricing/gemini-4-argon/): Google, 4 providers, from USD 2 input / USD 10 output per 1M tokens; maker's own API USD 2 / USD 10 - [Gemini 4 Carbon API pricing](https://broadrouter.com/llm-api-pricing/gemini-4-carbon/): Google, 4 providers, from USD 2 input / USD 10 output per 1M tokens; maker's own API USD 2 / USD 10 - [Gemini 4 Hydrogen API pricing](https://broadrouter.com/llm-api-pricing/gemini-4-hydrogen/): Google, 3 providers, from USD 1.50 input / USD 7.50 output per 1M tokens; maker's own API USD 1.50 / USD 7.50 - [Gemma 3 12B API pricing](https://broadrouter.com/llm-api-pricing/gemma-3-12b-it/): Google, 4 providers, from USD 0.05 input / USD 0.15 output per 1M tokens - [Gemma 3 27B API pricing](https://broadrouter.com/llm-api-pricing/gemma-3-27b-it/): Google, 6 providers, from USD 0.08 input / USD 0.16 output per 1M tokens - [Gemma 3 4B API pricing](https://broadrouter.com/llm-api-pricing/gemma-3-4b-it/): Google, 3 providers, from USD 0.04 input / USD 0.08 output per 1M tokens - [Gemma 4 26B A4B API pricing](https://broadrouter.com/llm-api-pricing/gemma-4-26b-a4b-it/): Google, 11 providers, from USD 0.07 input / USD 0.30 output per 1M tokens - [Gemma 4 31B API pricing](https://broadrouter.com/llm-api-pricing/gemma-4-31b-it/): Google, 12 providers, from USD 0.10 input / USD 0.34 output per 1M tokens - [gemma-4-31B-it-turbo API pricing](https://broadrouter.com/llm-api-pricing/gemma-4-31b-it-turbo/): Google, 2 providers, from USD 0.09 input / USD 0.34 output per 1M tokens - [GLM 4.5 Air API pricing](https://broadrouter.com/llm-api-pricing/glm-4.5-air/): Z.ai, 4 providers, from USD 0.13 input / USD 0.85 output per 1M tokens; maker's own API USD 0.20 / USD 1.10 - [GLM 4.5V API pricing](https://broadrouter.com/llm-api-pricing/glm-4.5v/): Z.ai, 2 providers, from USD 0.60 input / USD 1.80 output per 1M tokens; maker's own API USD 0.60 / USD 1.80 - [GLM 4.6 API pricing](https://broadrouter.com/llm-api-pricing/glm-4.6/): Z.ai, 6 providers, from USD 0.43 input / USD 1.75 output per 1M tokens; maker's own API USD 0.60 / USD 2.20 - [GLM 4.6V API pricing](https://broadrouter.com/llm-api-pricing/glm-4.6v/): Z.ai, 2 providers, from USD 0.30 input / USD 0.90 output per 1M tokens; maker's own API USD 0.30 / USD 0.90 - [GLM 4.7 API pricing](https://broadrouter.com/llm-api-pricing/glm-4.7/): Z.ai, 8 providers, from USD 0.40 input / USD 1.75 output per 1M tokens; maker's own API USD 0.60 / USD 2.20 - [GLM 4.7 Flash API pricing](https://broadrouter.com/llm-api-pricing/glm-4.7-flash/): Z.ai, 6 providers, from USD 0.06 input / USD 0.40 output per 1M tokens - [GLM 5 API pricing](https://broadrouter.com/llm-api-pricing/glm-5/): Z.ai, 10 providers, from USD 0.60 input / USD 1.92 output per 1M tokens; maker's own API USD 1 / USD 3.20 - [GLM 5.1 API pricing](https://broadrouter.com/llm-api-pricing/glm-5.1/): Z.ai, 12 providers, from USD 0.90 input / USD 2.90 output per 1M tokens; maker's own API USD 1.40 / USD 4.40 - [GLM 5.2 API pricing](https://broadrouter.com/llm-api-pricing/glm-5.2/): Z.ai, 19 providers, from USD 0.5625 input / USD 1.80 output per 1M tokens; maker's own API USD 1.40 / USD 4.40 - [GLM 5.3 API pricing](https://broadrouter.com/llm-api-pricing/glm-5.3/): Z.ai, 24 providers, from USD 0.19 input / USD 2.20 output per 1M tokens; maker's own API USD 1.40 / USD 4.40 - [GLM 5.3 Flash API pricing](https://broadrouter.com/llm-api-pricing/glm-5.3-flash/): Z.ai, 20 providers, from USD 0.075 input / USD 0.25 output per 1M tokens; maker's own API USD 0.15 / USD 0.50 - [GLM 5.3 FlashX API pricing](https://broadrouter.com/llm-api-pricing/glm-5.3-flashx/): Z.ai, 2 providers, from USD 0.37 input / USD 1.25 output per 1M tokens; maker's own API USD 0.37 / USD 1.25 - [GLM-4.5-FP8 API pricing](https://broadrouter.com/llm-api-pricing/glm-4.5/): Z.ai, 2 providers, from USD 0.60 input / USD 2.20 output per 1M tokens; maker's own API USD 0.60 / USD 2.20 - [GPT Chat Latest API pricing](https://broadrouter.com/llm-api-pricing/gpt-chat-latest/): OpenAI, 2 providers, from USD 5 input / USD 30 output per 1M tokens; maker's own API USD 5 / USD 30 - [GPT OSS 20B API pricing](https://broadrouter.com/llm-api-pricing/gpt-oss-20b/): OpenAI, 10 providers, from USD 0.018 input / USD 0.09 output per 1M tokens - [GPT OSS Safeguard 20B API pricing](https://broadrouter.com/llm-api-pricing/gpt-oss-safeguard-20b/): OpenAI, 3 providers, from USD 0.07 input / USD 0.20 output per 1M tokens - [GPT-4.1 (2025-04-14) API pricing](https://broadrouter.com/llm-api-pricing/gpt-4.1/): OpenAI, 3 providers, from USD 2 input / USD 8 output per 1M tokens; maker's own API USD 2 / USD 8 - [GPT-4.1 Mini API pricing](https://broadrouter.com/llm-api-pricing/gpt-4.1-mini/): OpenAI, 2 providers, from USD 0.40 input / USD 1.60 output per 1M tokens; maker's own API USD 0.40 / USD 1.60 - [GPT-4o API pricing](https://broadrouter.com/llm-api-pricing/gpt-4o/): OpenAI, 2 providers, from USD 2.50 input / USD 10 output per 1M tokens; maker's own API USD 2.50 / USD 10 - [GPT-4o mini API pricing](https://broadrouter.com/llm-api-pricing/gpt-4o-mini/): OpenAI, 3 providers, from USD 0.15 input / USD 0.60 output per 1M tokens; maker's own API USD 0.15 / USD 0.60 - [GPT-5 (2025-08-07) API pricing](https://broadrouter.com/llm-api-pricing/gpt-5/): OpenAI, 3 providers, from USD 1.25 input / USD 10 output per 1M tokens; maker's own API USD 1.25 / USD 10 - [GPT-5 Mini API pricing](https://broadrouter.com/llm-api-pricing/gpt-5-mini/): OpenAI, 3 providers, from USD 0.25 input / USD 2 output per 1M tokens; maker's own API USD 0.25 / USD 2 - [GPT-5 Nano API pricing](https://broadrouter.com/llm-api-pricing/gpt-5-nano/): OpenAI, 3 providers, from USD 0.05 input / USD 0.40 output per 1M tokens; maker's own API USD 0.05 / USD 0.40 - [GPT-5 Pro API pricing](https://broadrouter.com/llm-api-pricing/gpt-5-pro/): OpenAI, 2 providers, from USD 15 input / USD 120 output per 1M tokens; maker's own API USD 15 / USD 120 - [GPT-5.1 API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.1/): OpenAI, 2 providers, from USD 1.25 input / USD 10 output per 1M tokens; maker's own API USD 1.25 / USD 10 - [GPT-5.2 API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.2/): OpenAI, 3 providers, from USD 1.75 input / USD 14 output per 1M tokens; maker's own API USD 1.75 / USD 14 - [GPT-5.2 Pro API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.2-pro/): OpenAI, 2 providers, from USD 21 input / USD 168 output per 1M tokens; maker's own API USD 21 / USD 168 - [GPT-5.3 Codex API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.3-codex/): OpenAI, 3 providers, from USD 1.75 input / USD 14 output per 1M tokens; maker's own API USD 1.75 / USD 14 - [GPT-5.4 API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.4/): OpenAI, 4 providers, from USD 2.50 input / USD 15 output per 1M tokens; maker's own API USD 2.50 / USD 15 - [GPT-5.4 Mini API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.4-mini/): OpenAI, 3 providers, from USD 0.75 input / USD 4.50 output per 1M tokens; maker's own API USD 0.75 / USD 4.50 - [GPT-5.4 Nano API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.4-nano/): OpenAI, 3 providers, from USD 0.20 input / USD 1.25 output per 1M tokens; maker's own API USD 0.20 / USD 1.25 - [GPT-5.4 Pro API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.4-pro/): OpenAI, 3 providers, from USD 30 input / USD 180 output per 1M tokens; maker's own API USD 30 / USD 180 - [GPT-5.5 API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.5/): OpenAI, 6 providers, from USD 5 input / USD 30 output per 1M tokens; maker's own API USD 5 / USD 30 - [GPT-5.6 Luna API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.6-luna/): OpenAI, 6 providers, from USD 0.20 input / USD 1.20 output per 1M tokens; maker's own API USD 0.20 / USD 1.20 - [GPT-5.6 Sol API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.6-sol/): OpenAI, 6 providers, from USD 4 input / USD 20 output per 1M tokens; maker's own API USD 4 / USD 20 - [GPT-5.6 Terra API pricing](https://broadrouter.com/llm-api-pricing/gpt-5.6-terra/): OpenAI, 6 providers, from USD 2 input / USD 12 output per 1M tokens; maker's own API USD 2 / USD 12 - [GPT-6 Astra API pricing](https://broadrouter.com/llm-api-pricing/gpt-6-astra/): OpenAI, 6 providers, from USD 10 input / USD 50 output per 1M tokens; maker's own API USD 10 / USD 50 - [GPT-6 Luna API pricing](https://broadrouter.com/llm-api-pricing/gpt-6-luna/): OpenAI, 6 providers, from USD 0.10 input / USD 0.50 output per 1M tokens; maker's own API USD 0.10 / USD 0.50 - [GPT-6 Sol API pricing](https://broadrouter.com/llm-api-pricing/gpt-6-sol/): OpenAI, 6 providers, from USD 2 input / USD 10 output per 1M tokens; maker's own API USD 2 / USD 10 - [GPT-6.1 Sol API pricing](https://broadrouter.com/llm-api-pricing/gpt-6.1-sol/): OpenAI, 6 providers, from USD 2 input / USD 10 output per 1M tokens; maker's own API USD 2 / USD 10 - [GPT-6.1 Terra API pricing](https://broadrouter.com/llm-api-pricing/gpt-6.1-terra/): OpenAI, 4 providers, from USD 1 input / USD 5 output per 1M tokens; maker's own API USD 1 / USD 5 - [Grok 4.20 (non-reasoning) API pricing](https://broadrouter.com/llm-api-pricing/grok-4.20-non-reasoning/): xAI, 2 providers, from USD 1.25 input / USD 2.50 output per 1M tokens; maker's own API USD 1.25 / USD 2.50 - [Grok 4.20 (reasoning) API pricing](https://broadrouter.com/llm-api-pricing/grok-4.20/): xAI, 3 providers, from USD 1.25 input / USD 2.50 output per 1M tokens; maker's own API USD 1.25 / USD 2.50 - [Grok 4.3 API pricing](https://broadrouter.com/llm-api-pricing/grok-4.3/): xAI, 4 providers, from USD 1.25 input / USD 2.50 output per 1M tokens; maker's own API USD 1.25 / USD 2.50 - [Grok 4.5 API pricing](https://broadrouter.com/llm-api-pricing/grok-4.5/): xAI, 3 providers, from USD 2 input / USD 6 output per 1M tokens; maker's own API USD 2 / USD 6 - [Grok 4.6 API pricing](https://broadrouter.com/llm-api-pricing/grok-4.6/): xAI, 6 providers, from USD 2 input / USD 6 output per 1M tokens; maker's own API USD 2 / USD 6 - [Grok 4.7 API pricing](https://broadrouter.com/llm-api-pricing/grok-4.7/): xAI, 6 providers, from USD 2 input / USD 6 output per 1M tokens; maker's own API USD 2 / USD 6 - [Hy3 API pricing](https://broadrouter.com/llm-api-pricing/hy3/): Tencent, 4 providers, from USD 0.132 input / USD 0.528 output per 1M tokens - [Hy4 API pricing](https://broadrouter.com/llm-api-pricing/hy4/): Tencent, 3 providers, from USD 0.417 input / USD 1.2505 output per 1M tokens - [Hy4 Preview API pricing](https://broadrouter.com/llm-api-pricing/hy4-preview/): Tencent, 6 providers, from USD 0.834 input / USD 2.501 output per 1M tokens - [IBM Granite 4.2 8B API pricing](https://broadrouter.com/llm-api-pricing/granite-4.2-8b/): IBM, 2 providers, from USD 0.06 input / USD 0.15 output per 1M tokens - [Inkling 2 Small API pricing](https://broadrouter.com/llm-api-pricing/inkling-2-small/): Thinking Machines, 3 providers, from USD 0.285 input / USD 0.95 output per 1M tokens - [Kimi K2 Instruct API pricing](https://broadrouter.com/llm-api-pricing/kimi-k2/): Moonshot AI, 2 providers, from USD 0.57 input / USD 2.30 output per 1M tokens - [Kimi K2 Instruct 0905 API pricing](https://broadrouter.com/llm-api-pricing/kimi-k2-0905/): Moonshot AI, 2 providers, from USD 0.60 input / USD 2.50 output per 1M tokens - [Kimi K2 Thinking API pricing](https://broadrouter.com/llm-api-pricing/kimi-k2-thinking/): Moonshot AI, 6 providers, from USD 0.60 input / USD 1.20 output per 1M tokens - [Kimi K2.5 API pricing](https://broadrouter.com/llm-api-pricing/kimi-k2.5/): Moonshot AI, 8 providers, from USD 0.45 input / USD 2.25 output per 1M tokens - [Kimi K2.6 API pricing](https://broadrouter.com/llm-api-pricing/kimi-k2.6/): Moonshot AI, 13 providers, from USD 0.39 input / USD 1.80 output per 1M tokens; maker's own API USD 0.95 / USD 4 - [Kimi K2.7 Code API pricing](https://broadrouter.com/llm-api-pricing/kimi-k2.7-code/): Moonshot AI, 12 providers, from USD 0.60 input / USD 3 output per 1M tokens; maker's own API USD 0.95 / USD 4 - [Kimi K3 API pricing](https://broadrouter.com/llm-api-pricing/kimi-k3/): Moonshot AI, 18 providers, from USD 1.85 input / USD 9.50 output per 1M tokens; maker's own API USD 3 / USD 15 - [Kimi K3.1 Pro API pricing](https://broadrouter.com/llm-api-pricing/kimi-k3.1-pro/): Moonshot AI, 13 providers, from USD 1.75 input / USD 9.50 output per 1M tokens; maker's own API USD 2 / USD 10 - [L3.1-70B-Euryale-v2.2 API pricing](https://broadrouter.com/llm-api-pricing/l3.1-70b-euryale-v2.2/): Sao10K, 2 providers, from USD 0.72 input / USD 0.72 output per 1M tokens - [Ling 3.0 Flash API pricing](https://broadrouter.com/llm-api-pricing/ling-3.0-flash/): inclusionAI, 2 providers, from USD 0.06 input / USD 0.18 output per 1M tokens - [Ling 3.0 Flash VL API pricing](https://broadrouter.com/llm-api-pricing/ling-3.0-flash-vl/): inclusionAI, 2 providers, from USD 0.06 input / USD 0.18 output per 1M tokens - [Llama 3 8B Instruct API pricing](https://broadrouter.com/llm-api-pricing/llama-3-8b-instruct/): Meta, 3 providers, from USD 0.0925 input / USD 0.095 output per 1M tokens - [Llama 3 8B Stheno V3.2 API pricing](https://broadrouter.com/llm-api-pricing/l3-8b-stheno-v3.2/): Sao10K, 2 providers, from USD 0.05 input / USD 0.05 output per 1M tokens - [Llama 3 Instruct (70B) API pricing](https://broadrouter.com/llm-api-pricing/llama-3-70b-instruct/): Meta, 2 providers, from USD 2.65 input / USD 3.50 output per 1M tokens - [Llama 3.1 8B Instruct API pricing](https://broadrouter.com/llm-api-pricing/llama-3.1-8b-instruct/): Meta, 6 providers, from USD 0.02 input / USD 0.04 output per 1M tokens - [Llama 3.1 Instruct (70B) API pricing](https://broadrouter.com/llm-api-pricing/llama-3.1-70b-instruct/): Meta, 3 providers, from USD 0.293 input / USD 0.72 output per 1M tokens - [Llama 3.2 11B Vision Instruct API pricing](https://broadrouter.com/llm-api-pricing/llama-3.2-11b-vision-instruct/): Meta, 3 providers, from USD 0.049 input / USD 0.16 output per 1M tokens - [Llama 3.2 1B Instruct API pricing](https://broadrouter.com/llm-api-pricing/llama-3.2-1b-instruct/): Meta, 3 providers, from USD 0.027 input / USD 0.10 output per 1M tokens - [Llama 3.2 3B Instruct API pricing](https://broadrouter.com/llm-api-pricing/llama-3.2-3b-instruct/): Meta, 4 providers, from USD 0.0509 input / USD 0.15 output per 1M tokens - [Llama 3.3 70B API pricing](https://broadrouter.com/llm-api-pricing/llama-3.3-70b-instruct/): Meta, 14 providers, from USD 0.10 input / USD 0.32 output per 1M tokens - [Llama 4 Maverick 17B 128E Instruct FP8 API pricing](https://broadrouter.com/llm-api-pricing/llama-4-maverick/): Meta, 7 providers, from USD 0.24 input / USD 0.80 output per 1M tokens - [Llama 4 Scout 17B 16E Instruct API pricing](https://broadrouter.com/llm-api-pricing/llama-4-scout/): Meta, 6 providers, from USD 0.08 input / USD 0.30 output per 1M tokens - [MiMo v2.5 API pricing](https://broadrouter.com/llm-api-pricing/mimo-v2.5/): Xiaomi, 4 providers, from USD 0.14 input / USD 0.28 output per 1M tokens - [MiMo V2.6 Flash API pricing](https://broadrouter.com/llm-api-pricing/mimo-v2.6-flash/): Xiaomi, 5 providers, from USD 0.112 input / USD 0.224 output per 1M tokens - [MiMo V2.6 Pro API pricing](https://broadrouter.com/llm-api-pricing/mimo-v2.6-pro/): Xiaomi, 3 providers, from USD 0.32625 input / USD 0.6525 output per 1M tokens - [MiniMax M2 API pricing](https://broadrouter.com/llm-api-pricing/minimax-m2/): MiniMax, 4 providers, from USD 0.2775 input / USD 1.11 output per 1M tokens - [Minimax M2.1 API pricing](https://broadrouter.com/llm-api-pricing/minimax-m2.1/): MiniMax, 3 providers, from USD 0.30 input / USD 1.20 output per 1M tokens - [MiniMax M2.5 API pricing](https://broadrouter.com/llm-api-pricing/minimax-m2.5/): MiniMax, 6 providers, from USD 0.27 input / USD 0.95 output per 1M tokens - [MiniMax M2.7 API pricing](https://broadrouter.com/llm-api-pricing/minimax-m2.7/): MiniMax, 3 providers, from USD 0.21 input / USD 0.84 output per 1M tokens - [MiniMax M3 API pricing](https://broadrouter.com/llm-api-pricing/minimax-m3/): MiniMax, 10 providers, from USD 0.23 input / USD 0.96 output per 1M tokens; maker's own API USD 0.30 / USD 1.20 - [MiniMax M3.1 Flash API pricing](https://broadrouter.com/llm-api-pricing/minimax-m3.1-flash/): MiniMax, 9 providers, from USD 0.15 input / USD 0.60 output per 1M tokens; maker's own API USD 0.15 / USD 0.60 - [Ministral 3 14B API pricing](https://broadrouter.com/llm-api-pricing/ministral-14b-2512/): Mistral AI, 3 providers, from USD 0.20 input / USD 0.20 output per 1M tokens; maker's own API USD 0.20 / USD 0.20 - [Ministral 3 3B API pricing](https://broadrouter.com/llm-api-pricing/ministral-3b-2512/): Mistral AI, 2 providers, from USD 0.10 input / USD 0.10 output per 1M tokens; maker's own API USD 0.10 / USD 0.10 - [Ministral 3 8B API pricing](https://broadrouter.com/llm-api-pricing/ministral-8b-2512/): Mistral AI, 2 providers, from USD 0.15 input / USD 0.15 output per 1M tokens; maker's own API USD 0.15 / USD 0.15 - [Mistral Large 3 API pricing](https://broadrouter.com/llm-api-pricing/mistral-large-2512/): Mistral AI, 3 providers, from USD 0.50 input / USD 1.50 output per 1M tokens - [Mistral Medium 3 API pricing](https://broadrouter.com/llm-api-pricing/mistral-medium-3/): Mistral AI, 2 providers, from USD 0.40 input / USD 2 output per 1M tokens - [Mistral Small 3.1 24B Instruct API pricing](https://broadrouter.com/llm-api-pricing/mistral-small-3.1-24b-instruct/): Mistral AI, 4 providers, from USD 0.10 input / USD 0.30 output per 1M tokens - [Mistral Small 3.2 24B Instruct API pricing](https://broadrouter.com/llm-api-pricing/mistral-small-3.2-24b-instruct/): Mistral AI, 4 providers, from USD 0.075 input / USD 0.20 output per 1M tokens - [Mistral-Nemo-Instruct-2407 API pricing](https://broadrouter.com/llm-api-pricing/mistral-nemo-instruct-2407/): Mistral AI, 2 providers, from USD 0.029 input / USD 0.03 output per 1M tokens - [Mistral-Small-24B-Instruct-2501 API pricing](https://broadrouter.com/llm-api-pricing/mistral-small-24b-instruct-2501/): Mistral AI, 2 providers, from USD 0.05 input / USD 0.08 output per 1M tokens - [Muse Glimmer 30B API pricing](https://broadrouter.com/llm-api-pricing/muse-glimmer-30b/): Meta, 3 providers, from USD 0.30 input / USD 1.10 output per 1M tokens - [Muse Spark Open API pricing](https://broadrouter.com/llm-api-pricing/muse-spark-open/): Meta, 4 providers, from USD 1.25 input / USD 4.25 output per 1M tokens - [NVIDIA Nemotron 3 Nano 30B A3B API pricing](https://broadrouter.com/llm-api-pricing/nemotron-3-nano-30b-a3b/): NVIDIA, 3 providers, from USD 0.05 input / USD 0.20 output per 1M tokens - [NVIDIA Nemotron 3 Super 120B A12B API pricing](https://broadrouter.com/llm-api-pricing/nemotron-3-super-120b-a12b/): NVIDIA, 3 providers, from USD 0.075 input / USD 0.40 output per 1M tokens - [NVIDIA Nemotron 3 Ultra API pricing](https://broadrouter.com/llm-api-pricing/nemotron-3-ultra-550b-a55b/): NVIDIA, 9 providers, from USD 0.50 input / USD 2.15 output per 1M tokens - [NVIDIA Nemotron 3.5 Lightning 30B A3B API pricing](https://broadrouter.com/llm-api-pricing/nemotron-3.5-lightning/): NVIDIA, 5 providers, from USD 0.05 input / USD 0.16 output per 1M tokens - [NVIDIA Nemotron 3.5 Super VL API pricing](https://broadrouter.com/llm-api-pricing/nemotron-3.5-super-vl/): NVIDIA, 6 providers, from USD 0.10 input / USD 0.40 output per 1M tokens - [o3 (2025-04-16) API pricing](https://broadrouter.com/llm-api-pricing/o3/): OpenAI, 3 providers, from USD 2 input / USD 8 output per 1M tokens; maker's own API USD 2 / USD 8 - [OpenAI GPT OSS 120B API pricing](https://broadrouter.com/llm-api-pricing/gpt-oss-120b/): OpenAI, 21 providers, from USD 0.03 input / USD 0.17 output per 1M tokens - [Phi-4 API pricing](https://broadrouter.com/llm-api-pricing/phi-4/): Microsoft, 3 providers, from USD 0.07 input / USD 0.14 output per 1M tokens - [Qwen 2.5 7B API pricing](https://broadrouter.com/llm-api-pricing/qwen2.5-7b-instruct/): Qwen, 3 providers, from USD 0.05 input / USD 0.05 output per 1M tokens - [Qwen-MT Plus API pricing](https://broadrouter.com/llm-api-pricing/qwen-mt-plus/): Qwen, 2 providers, from USD 0.25 input / USD 0.75 output per 1M tokens; maker's own API USD 2.46 / USD 7.37 - [Qwen2.5 Coder 32B Instruct API pricing](https://broadrouter.com/llm-api-pricing/qwen-2.5-coder-32b-instruct/): Qwen, 2 providers, from USD 0.66 input / USD 1 output per 1M tokens - [Qwen2.5-72B-Instruct API pricing](https://broadrouter.com/llm-api-pricing/qwen2.5-72b-instruct/): Qwen, 3 providers, from USD 0.36 input / USD 0.40 output per 1M tokens - [Qwen3 235B A22B FP8 API pricing](https://broadrouter.com/llm-api-pricing/qwen3-235b-a22b/): Qwen, 2 providers, from USD 0.17 input / USD 1.09 output per 1M tokens - [Qwen3 235B A22B Instruct 2507 FP8 API pricing](https://broadrouter.com/llm-api-pricing/qwen3-235b-a22b-2507/): Qwen, 5 providers, from USD 0.0875 input / USD 0.35 output per 1M tokens - [Qwen3 235B A22B Thinking 2507 FP8 API pricing](https://broadrouter.com/llm-api-pricing/qwen3-235b-a22b-thinking-2507/): Qwen, 3 providers, from USD 0.35 input / USD 3.50 output per 1M tokens - [Qwen3 30B A3B (fp8) API pricing](https://broadrouter.com/llm-api-pricing/qwen3-30b-a3b/): Qwen, 3 providers, from USD 0.051 input / USD 0.25 output per 1M tokens - [Qwen3 32B FP8 API pricing](https://broadrouter.com/llm-api-pricing/qwen3-32b/): Qwen, 5 providers, from USD 0.08 input / USD 0.28 output per 1M tokens - [Qwen3 Coder 30B A3B API pricing](https://broadrouter.com/llm-api-pricing/qwen3-coder-30b-a3b-instruct/): Qwen, 3 providers, from USD 0.07 input / USD 0.28 output per 1M tokens - [Qwen3 Coder 480B A35B API pricing](https://broadrouter.com/llm-api-pricing/qwen3-coder/): Qwen, 6 providers, from USD 0.22 input / USD 1 output per 1M tokens - [Qwen3 Coder Next API pricing](https://broadrouter.com/llm-api-pricing/qwen3-coder-next/): Qwen, 3 providers, from USD 0.12 input / USD 0.80 output per 1M tokens - [Qwen3 Next 80B A3B Instruct API pricing](https://broadrouter.com/llm-api-pricing/qwen3-next-80b-a3b-instruct/): Qwen, 5 providers, from USD 0.09 input / USD 1.10 output per 1M tokens - [Qwen3 Next 80B A3B Thinking API pricing](https://broadrouter.com/llm-api-pricing/qwen3-next-80b-a3b-thinking/): Qwen, 2 providers, from USD 0.15 input / USD 1.20 output per 1M tokens - [Qwen3 VL 235B A22B API pricing](https://broadrouter.com/llm-api-pricing/qwen3-vl-235b-a22b-instruct/): Qwen, 3 providers, from USD 0.20 input / USD 0.88 output per 1M tokens - [Qwen3 VL 30B A3B Instruct API pricing](https://broadrouter.com/llm-api-pricing/qwen3-vl-30b-a3b-instruct/): Qwen, 4 providers, from USD 0.15 input / USD 0.50 output per 1M tokens - [Qwen3-14B API pricing](https://broadrouter.com/llm-api-pricing/qwen3-14b/): Qwen, 3 providers, from USD 0.07 input / USD 0.24 output per 1M tokens - [Qwen3-30B-A3B-Instruct-2507 API pricing](https://broadrouter.com/llm-api-pricing/qwen3-30b-a3b-instruct-2507/): Qwen, 2 providers, from USD 0.09 input / USD 0.30 output per 1M tokens - [Qwen3-8B API pricing](https://broadrouter.com/llm-api-pricing/qwen3-8b/): Qwen, 2 providers, from USD 0.06 input / USD 0.06 output per 1M tokens - [Qwen3-VL 8B API pricing](https://broadrouter.com/llm-api-pricing/qwen3-vl-8b-instruct/): Qwen, 2 providers, from USD 0.18 input / USD 0.68 output per 1M tokens - [Qwen3.5 122B A10B API pricing](https://broadrouter.com/llm-api-pricing/qwen3.5-122b-a10b/): Qwen, 4 providers, from USD 0.26 input / USD 2.08 output per 1M tokens; maker's own API USD 0.40 / USD 3.20 - [Qwen3.5 27B API pricing](https://broadrouter.com/llm-api-pricing/qwen3.5-27b/): Qwen, 7 providers, from USD 0.20 input / USD 1.75 output per 1M tokens; maker's own API USD 0.30 / USD 2.40 - [Qwen3.5 35B A3B API pricing](https://broadrouter.com/llm-api-pricing/qwen3.5-35b-a3b/): Qwen, 7 providers, from USD 0.14 input / USD 1 output per 1M tokens; maker's own API USD 0.25 / USD 2 - [Qwen3.5 397B A17B API pricing](https://broadrouter.com/llm-api-pricing/qwen3.5-397b-a17b/): Qwen, 7 providers, from USD 0.45 input / USD 3 output per 1M tokens; maker's own API USD 0.60 / USD 3.60 - [Qwen3.5 9B API pricing](https://broadrouter.com/llm-api-pricing/qwen3.5-9b/): Qwen, 4 providers, from USD 0.10 input / USD 0.15 output per 1M tokens - [Qwen3.6 27B API pricing](https://broadrouter.com/llm-api-pricing/qwen3.6-27b/): Qwen, 7 providers, from USD 0.25 input / USD 2 output per 1M tokens; maker's own API USD 0.60 / USD 3.60 - [Qwen3.6 35B A3B API pricing](https://broadrouter.com/llm-api-pricing/qwen3.6-35b-a3b/): Qwen, 10 providers, from USD 0.10 input / USD 0.89 output per 1M tokens; maker's own API USD 0.375 / USD 2.25 - [Qwen3.6 Plus API pricing](https://broadrouter.com/llm-api-pricing/qwen3.6-plus/): Qwen, 3 providers, from USD 0.50 input / USD 3 output per 1M tokens; maker's own API USD 0.50 / USD 3 - [Qwen3.7 Max API pricing](https://broadrouter.com/llm-api-pricing/qwen3.7-max/): Qwen, 5 providers, from USD 1.25 input / USD 3.75 output per 1M tokens; maker's own API USD 2.50 / USD 7.50 - [Qwen3.7 Plus API pricing](https://broadrouter.com/llm-api-pricing/qwen3.7-plus/): Qwen, 2 providers, from USD 0.32 input / USD 1.28 output per 1M tokens; maker's own API USD 0.40 / USD 1.60 - [Qwen3.8 2.4T A95B API pricing](https://broadrouter.com/llm-api-pricing/qwen3.8-2.4t-a95b/): Qwen, 6 providers, from USD 1 input / USD 4 output per 1M tokens; maker's own API USD 2 / USD 6 - [Qwen3.8 27B API pricing](https://broadrouter.com/llm-api-pricing/qwen3.8-27b/): Qwen, 15 providers, from USD 0.039 input / USD 1.25 output per 1M tokens; maker's own API USD 0.50 / USD 3 - [Qwen3.8 Flash API pricing](https://broadrouter.com/llm-api-pricing/qwen3.8-flash/): Qwen, 5 providers, from USD 0.113 input / USD 0.382 output per 1M tokens; maker's own API USD 0.15 / USD 0.47 - [Qwen3.8 Max API pricing](https://broadrouter.com/llm-api-pricing/qwen3.8-max/): Qwen, 6 providers, from USD 1.65 input / USD 4.951 output per 1M tokens; maker's own API USD 2 / USD 6 - [Qwen3.8 Max (0902) API pricing](https://broadrouter.com/llm-api-pricing/qwen3.8-max-0902/): Qwen, 3 providers, from USD 1.65 input / USD 4.951 output per 1M tokens; maker's own API USD 2 / USD 6 - [Ripple-65B Preview API pricing](https://broadrouter.com/llm-api-pricing/ripple-65b-preview/): Ricursive AI, 2 providers, from USD 0.20 input / USD 0.60 output per 1M tokens - [Step 3.7 Flash API pricing](https://broadrouter.com/llm-api-pricing/step-3.7-flash/): StepFun, 3 providers, from USD 0.20 input / USD 1.15 output per 1M tokens - [Step 5 Preview API pricing](https://broadrouter.com/llm-api-pricing/step-5-preview/): StepFun, 4 providers, from USD 1 input / USD 2.70 output per 1M tokens - [Step-3.5-Flash API pricing](https://broadrouter.com/llm-api-pricing/step-3.5-flash/): StepFun, 2 providers, from USD 0.10 input / USD 0.30 output per 1M tokens - Machine-readable list of 428 models at their cheapest standard price: https://broadrouter.com/api-prices.json ## Price comparisons (model vs model, price only, no quality claims) - [All model-vs-model price comparisons and how the pairs were picked](https://broadrouter.com/llm-api-pricing/compare/): grouped by maker and model; each model's list is at https://broadrouter.com/llm-api-pricing/compare/#m-. Each pair page gives both models' cheapest standard input / output price per 1M tokens with its source, the ratio on a 3:1 blend, workload costs, the crossover mix and the providers that sell both. Every indexed pair is listed below (cheapest standard price at 3 input : 1 output, USD per 1M tokens). - [GPT-6 Sol vs Claude Opus 5.5](https://broadrouter.com/llm-api-pricing/compare/gpt-6-sol-vs-claude-opus-5.5/): USD 4 vs USD 8 - [GPT-6 Sol vs Claude Sonnet 5.5](https://broadrouter.com/llm-api-pricing/compare/gpt-6-sol-vs-claude-sonnet-5.5/): USD 4 vs USD 4 - [Claude Opus 5.5 vs Gemini 4 Argon](https://broadrouter.com/llm-api-pricing/compare/claude-opus-5.5-vs-gemini-4-argon/): USD 8 vs USD 4 - [GPT-6 Sol vs Gemini 4 Argon](https://broadrouter.com/llm-api-pricing/compare/gpt-6-sol-vs-gemini-4-argon/): USD 4 vs USD 4 - [Claude Sonnet 5.5 vs Gemini 4 Hydrogen](https://broadrouter.com/llm-api-pricing/compare/claude-sonnet-5.5-vs-gemini-4-hydrogen/): USD 4 vs USD 3 - [GPT-6 Sol vs Grok 4.7](https://broadrouter.com/llm-api-pricing/compare/gpt-6-sol-vs-grok-4.7/): USD 4 vs USD 3 - [Claude Sonnet 5.5 vs Grok 4.7](https://broadrouter.com/llm-api-pricing/compare/claude-sonnet-5.5-vs-grok-4.7/): USD 4 vs USD 3 - [GPT-6 Astra vs Claude Opus 5.5](https://broadrouter.com/llm-api-pricing/compare/gpt-6-astra-vs-claude-opus-5.5/): USD 20 vs USD 8 - [GPT-6.1 Terra vs Gemini 4 Hydrogen](https://broadrouter.com/llm-api-pricing/compare/gpt-6.1-terra-vs-gemini-4-hydrogen/): USD 2 vs USD 3 - [GPT-6 Luna vs Claude Haiku 5.5](https://broadrouter.com/llm-api-pricing/compare/gpt-6-luna-vs-claude-haiku-5.5/): USD 0.20 vs USD 0.20 - [GPT-6 Luna vs Gemini 3.8 Flash](https://broadrouter.com/llm-api-pricing/compare/gpt-6-luna-vs-gemini-3.8-flash/): USD 0.20 vs USD 1.50 - [Claude Haiku 5.5 vs Gemini 3.8 Flash](https://broadrouter.com/llm-api-pricing/compare/claude-haiku-5.5-vs-gemini-3.8-flash/): USD 0.20 vs USD 1.50 - [Claude Opus 5.5 vs Claude Sonnet 5.5](https://broadrouter.com/llm-api-pricing/compare/claude-opus-5.5-vs-claude-sonnet-5.5/): USD 8 vs USD 4 - [GPT-6 Luna vs DeepSeek V4.1 Flash](https://broadrouter.com/llm-api-pricing/compare/gpt-6-luna-vs-deepseek-v4.1-flash/): USD 0.20 vs USD 0.21 - [Claude Haiku 5.5 vs DeepSeek V4.1 Flash](https://broadrouter.com/llm-api-pricing/compare/claude-haiku-5.5-vs-deepseek-v4.1-flash/): USD 0.20 vs USD 0.21 - [Gemini 3.8 Flash vs DeepSeek V4.1 Flash](https://broadrouter.com/llm-api-pricing/compare/gemini-3.8-flash-vs-deepseek-v4.1-flash/): USD 1.50 vs USD 0.21 - [Claude Sonnet 5.5 vs GLM 5.3](https://broadrouter.com/llm-api-pricing/compare/claude-sonnet-5.5-vs-glm-5.3/): USD 4 vs USD 1.047 - [GPT-6 Sol vs DeepSeek V4 Pro](https://broadrouter.com/llm-api-pricing/compare/gpt-6-sol-vs-deepseek-v4-pro/): USD 4 vs USD 1.196 - [Claude Sonnet 5.5 vs Kimi K3](https://broadrouter.com/llm-api-pricing/compare/claude-sonnet-5.5-vs-kimi-k3/): USD 4 vs USD 3.763 - [GPT-6 Luna vs Qwen3.8 Flash](https://broadrouter.com/llm-api-pricing/compare/gpt-6-luna-vs-qwen3.8-flash/): USD 0.20 vs USD 0.1803 - [GPT-6 Luna vs gpt-oss-120b](https://broadrouter.com/llm-api-pricing/compare/gpt-6-luna-vs-gpt-oss-120b/): USD 0.20 vs USD 0.065 - [DeepSeek V4.1 Flash vs GLM 5.3](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-vs-glm-5.3/): USD 0.21 vs USD 1.047 - [DeepSeek V4.1 Flash vs Qwen3.8 27B](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-vs-qwen3.8-27b/): USD 0.21 vs USD 0.3417 - [GLM 5.3 vs Qwen3.8 27B](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-qwen3.8-27b/): USD 1.047 vs USD 0.3417 - [GLM 5.3 vs Kimi K3](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-kimi-k3/): USD 1.047 vs USD 3.763 - [DeepSeek V4 Pro vs Kimi K3](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4-pro-vs-kimi-k3/): USD 1.196 vs USD 3.763 - [DeepSeek V4 Pro vs GLM 5.3](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4-pro-vs-glm-5.3/): USD 1.196 vs USD 1.047 - [DeepSeek V4 Pro vs Qwen3.8 Max](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4-pro-vs-qwen3.8-max/): USD 1.196 vs USD 2.475 - [Qwen3.8 Max vs Kimi K3](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-max-vs-kimi-k3/): USD 2.475 vs USD 3.763 - [GLM 5.3 vs MiniMax M3](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-minimax-m3/): USD 1.047 vs USD 0.4125 - [Kimi K3 vs MiniMax M3](https://broadrouter.com/llm-api-pricing/compare/kimi-k3-vs-minimax-m3/): USD 3.763 vs USD 0.4125 - [DeepSeek V4.1 Flash vs MiniMax M3.1 Flash](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-vs-minimax-m3.1-flash/): USD 0.21 vs USD 0.2625 - [GLM 5.3 Flash vs DeepSeek V4.1 Flash](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-flash-vs-deepseek-v4.1-flash/): USD 0.1187 vs USD 0.21 - [Qwen3.8 Flash vs DeepSeek V4.1 Flash](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-flash-vs-deepseek-v4.1-flash/): USD 0.1803 vs USD 0.21 - [gpt-oss-120b vs DeepSeek V4.1 Flash](https://broadrouter.com/llm-api-pricing/compare/gpt-oss-120b-vs-deepseek-v4.1-flash/): USD 0.065 vs USD 0.21 - [Llama 4 Maverick vs DeepSeek V4.1 Flash](https://broadrouter.com/llm-api-pricing/compare/llama-4-maverick-vs-deepseek-v4.1-flash/): USD 0.3875 vs USD 0.21 - [gpt-oss-120b vs Qwen3.8 27B](https://broadrouter.com/llm-api-pricing/compare/gpt-oss-120b-vs-qwen3.8-27b/): USD 0.065 vs USD 0.3417 - [Gemma 4 31B vs Qwen3.8 27B](https://broadrouter.com/llm-api-pricing/compare/gemma-4-31b-it-vs-qwen3.8-27b/): USD 0.16 vs USD 0.3417 - [Gemma 4 26B A4B vs Qwen3.6 35B A3B](https://broadrouter.com/llm-api-pricing/compare/gemma-4-26b-a4b-it-vs-qwen3.6-35b-a3b/): USD 0.1375 vs USD 0.2975 - [Gemma 4 26B A4B vs gpt-oss-20b](https://broadrouter.com/llm-api-pricing/compare/gemma-4-26b-a4b-it-vs-gpt-oss-20b/): USD 0.1375 vs USD 0.036 - [gpt-oss-20b vs Qwen3.6 35B A3B](https://broadrouter.com/llm-api-pricing/compare/gpt-oss-20b-vs-qwen3.6-35b-a3b/): USD 0.036 vs USD 0.2975 - [Llama 3.3 70B vs Qwen2.5 72B](https://broadrouter.com/llm-api-pricing/compare/llama-3.3-70b-instruct-vs-qwen2.5-72b-instruct/): USD 0.155 vs USD 0.37 - [gpt-oss-120b vs Llama 3.3 70B](https://broadrouter.com/llm-api-pricing/compare/gpt-oss-120b-vs-llama-3.3-70b-instruct/): USD 0.065 vs USD 0.155 - [Llama 4 Scout vs Gemma 4 31B](https://broadrouter.com/llm-api-pricing/compare/llama-4-scout-vs-gemma-4-31b-it/): USD 0.15 vs USD 0.16 - [Nemotron 3 Super vs gpt-oss-120b](https://broadrouter.com/llm-api-pricing/compare/nemotron-3-super-120b-a12b-vs-gpt-oss-120b/): USD 0.1563 vs USD 0.065 - [Mistral Small 3.2 vs Gemma 3 27B](https://broadrouter.com/llm-api-pricing/compare/mistral-small-3.2-24b-instruct-vs-gemma-3-27b-it/): USD 0.1062 vs USD 0.10 - [GPT-6 Sol vs GPT-5.6 Sol](https://broadrouter.com/llm-api-pricing/compare/gpt-6-sol-vs-gpt-5.6-sol/): USD 4 vs USD 8 - [GPT-6.1 Sol vs GPT-6 Sol](https://broadrouter.com/llm-api-pricing/compare/gpt-6.1-sol-vs-gpt-6-sol/): USD 4 vs USD 4 - [GPT-6 Luna vs GPT-5.6 Luna](https://broadrouter.com/llm-api-pricing/compare/gpt-6-luna-vs-gpt-5.6-luna/): USD 0.20 vs USD 0.45 - [Claude Opus 5.5 vs Claude Opus 5](https://broadrouter.com/llm-api-pricing/compare/claude-opus-5.5-vs-claude-opus-5/): USD 8 vs USD 10 - [Claude Sonnet 5.5 vs Claude Sonnet 5](https://broadrouter.com/llm-api-pricing/compare/claude-sonnet-5.5-vs-claude-sonnet-5/): USD 4 vs USD 4 - [Claude Haiku 5.5 vs Claude Haiku 4.5](https://broadrouter.com/llm-api-pricing/compare/claude-haiku-5.5-vs-claude-haiku-4.5/): USD 0.20 vs USD 2 - [Gemini 3.8 Flash vs Gemini 3.7 Flash](https://broadrouter.com/llm-api-pricing/compare/gemini-3.8-flash-vs-gemini-3.7-flash/): USD 1.50 vs USD 1.50 - [Grok 4.7 vs Grok 4.6](https://broadrouter.com/llm-api-pricing/compare/grok-4.7-vs-grok-4.6/): USD 3 vs USD 3 - [DeepSeek V4.1 Flash vs DeepSeek V4 Flash](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-vs-deepseek-v4-flash/): USD 0.21 vs USD 0.1125 - [DeepSeek V4 Pro vs DeepSeek V3.2](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4-pro-vs-deepseek-v3.2/): USD 1.196 vs USD 0.234 - [GLM 5.3 vs GLM 5.2](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-glm-5.2/): USD 1.047 vs USD 0.8719 - [GLM 5.3 Flash vs GLM 4.7 Flash](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-flash-vs-glm-4.7-flash/): USD 0.1187 vs USD 0.145 - [Qwen3.8 27B vs Qwen3.6 27B](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-27b-vs-qwen3.6-27b/): USD 0.3417 vs USD 0.6875 - [Qwen3.8 Max vs Qwen3.7 Max](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-max-vs-qwen3.7-max/): USD 2.475 vs USD 1.875 - [Kimi K3 vs Kimi K2.6](https://broadrouter.com/llm-api-pricing/compare/kimi-k3-vs-kimi-k2.6/): USD 3.763 vs USD 0.7425 - [Kimi K3.1 Pro vs Kimi K3](https://broadrouter.com/llm-api-pricing/compare/kimi-k3.1-pro-vs-kimi-k3/): USD 3.80 vs USD 3.763 - [MiniMax M3 vs MiniMax M2.5](https://broadrouter.com/llm-api-pricing/compare/minimax-m3-vs-minimax-m2.5/): USD 0.4125 vs USD 0.44 - [Llama 4 Maverick vs Llama 3.3 70B](https://broadrouter.com/llm-api-pricing/compare/llama-4-maverick-vs-llama-3.3-70b-instruct/): USD 0.3875 vs USD 0.155 - [Gemma 4 31B vs Gemma 3 27B](https://broadrouter.com/llm-api-pricing/compare/gemma-4-31b-it-vs-gemma-3-27b-it/): USD 0.16 vs USD 0.10 - [Gemma 4 26B A4B vs Gemma 4 31B](https://broadrouter.com/llm-api-pricing/compare/gemma-4-26b-a4b-it-vs-gemma-4-31b-it/): USD 0.1375 vs USD 0.16 - [gpt-oss-120b vs gpt-oss-20b](https://broadrouter.com/llm-api-pricing/compare/gpt-oss-120b-vs-gpt-oss-20b/): USD 0.065 vs USD 0.036 - [Claude Fable 5.1 vs Claude Fable 5](https://broadrouter.com/llm-api-pricing/compare/claude-fable-5.1-vs-claude-fable-5/): USD 20 vs USD 20 - [Claude Fable 5.2 vs Claude Fable 5.1](https://broadrouter.com/llm-api-pricing/compare/claude-fable-5.2-vs-claude-fable-5.1/): USD 16 vs USD 20 - [Claude Sonnet 5 vs Claude Sonnet 4.6](https://broadrouter.com/llm-api-pricing/compare/claude-sonnet-5-vs-claude-sonnet-4.6/): USD 4 vs USD 6 - [DeepSeek V3.1 vs DeepSeek V3](https://broadrouter.com/llm-api-pricing/compare/deepseek-chat-v3.1-vs-deepseek-chat/): USD 0.425 vs USD 0.4375 - [DeepSeek V3.2 vs DeepSeek V3.1](https://broadrouter.com/llm-api-pricing/compare/deepseek-v3.2-vs-deepseek-chat-v3.1/): USD 0.234 vs USD 0.425 - [DeepSeek-V3.2-Exp vs DeepSeek V3.2](https://broadrouter.com/llm-api-pricing/compare/deepseek-v3.2-exp-vs-deepseek-v3.2/): USD 0.305 vs USD 0.234 - [DeepSeek V4 Flash vs DeepSeek V3.2](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4-flash-vs-deepseek-v3.2/): USD 0.1125 vs USD 0.234 - [Gemini 3.5 Flash vs Gemini 2.5 Flash](https://broadrouter.com/llm-api-pricing/compare/gemini-3.5-flash-vs-gemini-2.5-flash/): USD 3.375 vs USD 0.85 - [Gemini 3.6 Flash vs Gemini 3.5 Flash](https://broadrouter.com/llm-api-pricing/compare/gemini-3.6-flash-vs-gemini-3.5-flash/): USD 1.50 vs USD 3.375 - [Gemini 3.7 Flash vs Gemini 3.6 Flash](https://broadrouter.com/llm-api-pricing/compare/gemini-3.7-flash-vs-gemini-3.6-flash/): USD 1.50 vs USD 1.50 - [Gemini 3.1 Flash-Lite vs Gemini 2.5 Flash-Lite](https://broadrouter.com/llm-api-pricing/compare/gemini-3.1-flash-lite-vs-gemini-2.5-flash-lite/): USD 0.5625 vs USD 0.175 - [Gemini 3.5 Flash-Lite vs Gemini 3.1 Flash-Lite](https://broadrouter.com/llm-api-pricing/compare/gemini-3.5-flash-lite-vs-gemini-3.1-flash-lite/): USD 0.85 vs USD 0.5625 - [Gemini 3.1 Pro Preview vs Gemini 2.5 Pro](https://broadrouter.com/llm-api-pricing/compare/gemini-3.1-pro-preview-vs-gemini-2.5-pro/): USD 4.50 vs USD 3.438 - [GLM 4.6 vs GLM-4.5-FP8](https://broadrouter.com/llm-api-pricing/compare/glm-4.6-vs-glm-4.5/): USD 0.76 vs USD 1 - [GLM 4.7 vs GLM 4.6](https://broadrouter.com/llm-api-pricing/compare/glm-4.7-vs-glm-4.6/): USD 0.7375 vs USD 0.76 - [GLM 5 vs GLM 4.7](https://broadrouter.com/llm-api-pricing/compare/glm-5-vs-glm-4.7/): USD 0.93 vs USD 0.7375 - [GLM 5.1 vs GLM 5](https://broadrouter.com/llm-api-pricing/compare/glm-5.1-vs-glm-5/): USD 1.40 vs USD 0.93 - [GLM 5.2 vs GLM 5.1](https://broadrouter.com/llm-api-pricing/compare/glm-5.2-vs-glm-5.1/): USD 0.8719 vs USD 1.40 - [GLM 4.5 Air vs GLM-4.5-FP8](https://broadrouter.com/llm-api-pricing/compare/glm-4.5-air-vs-glm-4.5/): USD 0.31 vs USD 1 - [GLM 4.6V vs GLM 4.5V](https://broadrouter.com/llm-api-pricing/compare/glm-4.6v-vs-glm-4.5v/): USD 0.45 vs USD 0.90 - [GLM 4.5V vs GLM-4.5-FP8](https://broadrouter.com/llm-api-pricing/compare/glm-4.5v-vs-glm-4.5/): USD 0.90 vs USD 1 - [GLM 4.7 Flash vs GLM 4.7](https://broadrouter.com/llm-api-pricing/compare/glm-4.7-flash-vs-glm-4.7/): USD 0.145 vs USD 0.7375 - [GLM 5.3 FlashX vs GLM 5.3](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-flashx-vs-glm-5.3/): USD 0.59 vs USD 1.047 - [GPT-5 vs GPT-4.1](https://broadrouter.com/llm-api-pricing/compare/gpt-5-vs-gpt-4.1/): USD 3.438 vs USD 3.50 - [GPT-5.2 vs GPT-5](https://broadrouter.com/llm-api-pricing/compare/gpt-5.2-vs-gpt-5/): USD 4.813 vs USD 3.438 - [GPT-5.4 vs GPT-5.2](https://broadrouter.com/llm-api-pricing/compare/gpt-5.4-vs-gpt-5.2/): USD 5.625 vs USD 4.813 - [GPT-5.5 vs GPT-5.4](https://broadrouter.com/llm-api-pricing/compare/gpt-5.5-vs-gpt-5.4/): USD 11.25 vs USD 5.625 - [GPT-5 Mini vs GPT-4.1 Mini](https://broadrouter.com/llm-api-pricing/compare/gpt-5-mini-vs-gpt-4.1-mini/): USD 0.6875 vs USD 0.70 - [GPT-5.4 Mini vs GPT-5 Mini](https://broadrouter.com/llm-api-pricing/compare/gpt-5.4-mini-vs-gpt-5-mini/): USD 1.688 vs USD 0.6875 - [GPT-4.1 Mini vs GPT-4.1](https://broadrouter.com/llm-api-pricing/compare/gpt-4.1-mini-vs-gpt-4.1/): USD 0.70 vs USD 3.50 - [GPT-5.4 Nano vs GPT-5 Nano](https://broadrouter.com/llm-api-pricing/compare/gpt-5.4-nano-vs-gpt-5-nano/): USD 0.4625 vs USD 0.1375 - [GPT-5 Nano vs GPT-5](https://broadrouter.com/llm-api-pricing/compare/gpt-5-nano-vs-gpt-5/): USD 0.1375 vs USD 3.438 - [GPT-5.2 Pro vs GPT-5 Pro](https://broadrouter.com/llm-api-pricing/compare/gpt-5.2-pro-vs-gpt-5-pro/): USD 57.75 vs USD 41.25 - [GPT-5.4 Pro vs GPT-5.2 Pro](https://broadrouter.com/llm-api-pricing/compare/gpt-5.4-pro-vs-gpt-5.2-pro/): USD 67.50 vs USD 57.75 - [GPT-5 Pro vs GPT-5](https://broadrouter.com/llm-api-pricing/compare/gpt-5-pro-vs-gpt-5/): USD 41.25 vs USD 3.438 - [GPT-5.6 Luna vs GPT-5.5](https://broadrouter.com/llm-api-pricing/compare/gpt-5.6-luna-vs-gpt-5.5/): USD 0.45 vs USD 11.25 - [GPT-5.6 Sol vs GPT-5.5](https://broadrouter.com/llm-api-pricing/compare/gpt-5.6-sol-vs-gpt-5.5/): USD 8 vs USD 11.25 - [GPT-6.1 Terra vs GPT-5.6 Terra](https://broadrouter.com/llm-api-pricing/compare/gpt-6.1-terra-vs-gpt-5.6-terra/): USD 2 vs USD 4.50 - [GPT-5.6 Terra vs GPT-5.5](https://broadrouter.com/llm-api-pricing/compare/gpt-5.6-terra-vs-gpt-5.5/): USD 4.50 vs USD 11.25 - [GPT-6 Astra vs GPT-5.5](https://broadrouter.com/llm-api-pricing/compare/gpt-6-astra-vs-gpt-5.5/): USD 20 vs USD 11.25 - [Grok 4.5 vs Grok 4.3](https://broadrouter.com/llm-api-pricing/compare/grok-4.5-vs-grok-4.3/): USD 3 vs USD 1.563 - [Grok 4.7 vs Grok 4.5](https://broadrouter.com/llm-api-pricing/compare/grok-4.7-vs-grok-4.5/): USD 3 vs USD 3 - [Hy4 vs Hy3](https://broadrouter.com/llm-api-pricing/compare/hy4-vs-hy3/): USD 0.6254 vs USD 0.231 - [Kimi K2.5 vs Kimi K2](https://broadrouter.com/llm-api-pricing/compare/kimi-k2.5-vs-kimi-k2/): USD 0.90 vs USD 1.002 - [Kimi K2.6 vs Kimi K2.5](https://broadrouter.com/llm-api-pricing/compare/kimi-k2.6-vs-kimi-k2.5/): USD 0.7425 vs USD 0.90 - [Kimi K2 Thinking vs Kimi K2](https://broadrouter.com/llm-api-pricing/compare/kimi-k2-thinking-vs-kimi-k2/): USD 0.90 vs USD 1.002 - [Kimi K2.7 Code vs Kimi K2.6](https://broadrouter.com/llm-api-pricing/compare/kimi-k2.7-code-vs-kimi-k2.6/): USD 1.20 vs USD 0.7425 - [Minimax M2.1 vs MiniMax M2](https://broadrouter.com/llm-api-pricing/compare/minimax-m2.1-vs-minimax-m2/): USD 0.525 vs USD 0.4856 - [MiniMax M2.5 vs Minimax M2.1](https://broadrouter.com/llm-api-pricing/compare/minimax-m2.5-vs-minimax-m2.1/): USD 0.44 vs USD 0.525 - [MiniMax M2.7 vs MiniMax M2.5](https://broadrouter.com/llm-api-pricing/compare/minimax-m2.7-vs-minimax-m2.5/): USD 0.3675 vs USD 0.44 - [MiniMax M3 vs MiniMax M2.7](https://broadrouter.com/llm-api-pricing/compare/minimax-m3-vs-minimax-m2.7/): USD 0.4125 vs USD 0.3675 - [MiniMax M3.1 Flash vs MiniMax M3](https://broadrouter.com/llm-api-pricing/compare/minimax-m3.1-flash-vs-minimax-m3/): USD 0.2625 vs USD 0.4125 - [Qwen3.7 Plus vs Qwen3.6 Plus](https://broadrouter.com/llm-api-pricing/compare/qwen3.7-plus-vs-qwen3.6-plus/): USD 0.56 vs USD 1.125 - [Step 3.7 Flash vs Step-3.5-Flash](https://broadrouter.com/llm-api-pricing/compare/step-3.7-flash-vs-step-3.5-flash/): USD 0.4375 vs USD 0.15 - [Claude Fable 5 vs Claude Opus 5](https://broadrouter.com/llm-api-pricing/compare/claude-fable-5-vs-claude-opus-5/): USD 20 vs USD 10 - [Claude Opus 5 vs Claude Sonnet 5](https://broadrouter.com/llm-api-pricing/compare/claude-opus-5-vs-claude-sonnet-5/): USD 10 vs USD 4 - [Claude Sonnet 5.5 vs Claude Haiku 5.5](https://broadrouter.com/llm-api-pricing/compare/claude-sonnet-5.5-vs-claude-haiku-5.5/): USD 4 vs USD 0.20 - [DeepSeek V4 Pro vs DeepSeek V4 Flash](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4-pro-vs-deepseek-v4-flash/): USD 1.196 vs USD 0.1125 - [Gemini 2.5 Pro vs Gemini 2.5 Flash](https://broadrouter.com/llm-api-pricing/compare/gemini-2.5-pro-vs-gemini-2.5-flash/): USD 3.438 vs USD 0.85 - [Gemini 3.1 Pro Preview vs Gemini 3.1 Flash-Lite](https://broadrouter.com/llm-api-pricing/compare/gemini-3.1-pro-preview-vs-gemini-3.1-flash-lite/): USD 4.50 vs USD 0.5625 - [GLM 5.3 FlashX vs GLM 5.3 Flash](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-flashx-vs-glm-5.3-flash/): USD 0.59 vs USD 0.1187 - [GPT-5 Pro vs GPT-5 Mini](https://broadrouter.com/llm-api-pricing/compare/gpt-5-pro-vs-gpt-5-mini/): USD 41.25 vs USD 0.6875 - [GPT-5 Mini vs GPT-5 Nano](https://broadrouter.com/llm-api-pricing/compare/gpt-5-mini-vs-gpt-5-nano/): USD 0.6875 vs USD 0.1375 - [GPT-5.4 Pro vs GPT-5.4 Mini](https://broadrouter.com/llm-api-pricing/compare/gpt-5.4-pro-vs-gpt-5.4-mini/): USD 67.50 vs USD 1.688 - [GPT-5.4 Mini vs GPT-5.4 Nano](https://broadrouter.com/llm-api-pricing/compare/gpt-5.4-mini-vs-gpt-5.4-nano/): USD 1.688 vs USD 0.4625 - [GPT-5.6 Sol vs GPT-5.6 Terra](https://broadrouter.com/llm-api-pricing/compare/gpt-5.6-sol-vs-gpt-5.6-terra/): USD 8 vs USD 4.50 - [GPT-5.6 Terra vs GPT-5.6 Luna](https://broadrouter.com/llm-api-pricing/compare/gpt-5.6-terra-vs-gpt-5.6-luna/): USD 4.50 vs USD 0.45 - [GPT-6 Astra vs GPT-6 Sol](https://broadrouter.com/llm-api-pricing/compare/gpt-6-astra-vs-gpt-6-sol/): USD 20 vs USD 4 - [GPT-6 Sol vs GPT-6 Luna](https://broadrouter.com/llm-api-pricing/compare/gpt-6-sol-vs-gpt-6-luna/): USD 4 vs USD 0.20 - [GPT-6.1 Sol vs GPT-6.1 Terra](https://broadrouter.com/llm-api-pricing/compare/gpt-6.1-sol-vs-gpt-6.1-terra/): USD 4 vs USD 2 - [Qwen3.7 Max vs Qwen3.7 Plus](https://broadrouter.com/llm-api-pricing/compare/qwen3.7-max-vs-qwen3.7-plus/): USD 1.875 vs USD 0.56 - [Qwen3.8 Max vs Qwen3.8 Flash](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-max-vs-qwen3.8-flash/): USD 2.475 vs USD 0.1803 - [Llama 3.1 70B vs Llama 3 70B](https://broadrouter.com/llm-api-pricing/compare/llama-3.1-70b-instruct-vs-llama-3-70b-instruct/): USD 0.72 vs USD 2.862 - [Llama 3.1 8B Instruct vs Llama 3 8B](https://broadrouter.com/llm-api-pricing/compare/llama-3.1-8b-instruct-vs-llama-3-8b-instruct/): USD 0.025 vs USD 0.09312 - [Llama 3.3 70B vs Llama 3.1 70B](https://broadrouter.com/llm-api-pricing/compare/llama-3.3-70b-instruct-vs-llama-3.1-70b-instruct/): USD 0.155 vs USD 0.72 - [Mistral Small 3.1 24B vs Mistral-Small-24B-2501](https://broadrouter.com/llm-api-pricing/compare/mistral-small-3.1-24b-instruct-vs-mistral-small-24b-instruct-2501/): USD 0.15 vs USD 0.0575 - [Mistral Small 3.2 vs Mistral Small 3.1 24B](https://broadrouter.com/llm-api-pricing/compare/mistral-small-3.2-24b-instruct-vs-mistral-small-3.1-24b-instruct/): USD 0.1062 vs USD 0.15 - [Nemotron 3.5 Lightning 30B A3B vs Nemotron 3 Nano 30B A3B](https://broadrouter.com/llm-api-pricing/compare/nemotron-3.5-lightning-vs-nemotron-3-nano-30b-a3b/): USD 0.085 vs USD 0.0875 - [Qwen3-8B vs Qwen 2.5 7B](https://broadrouter.com/llm-api-pricing/compare/qwen3-8b-vs-qwen2.5-7b-instruct/): USD 0.06 vs USD 0.05 - [Qwen3.5 27B vs Qwen3 32B](https://broadrouter.com/llm-api-pricing/compare/qwen3.5-27b-vs-qwen3-32b/): USD 0.5875 vs USD 0.13 - [Qwen3.5 35B A3B vs Qwen3 30B A3B](https://broadrouter.com/llm-api-pricing/compare/qwen3.5-35b-a3b-vs-qwen3-30b-a3b/): USD 0.355 vs USD 0.122 - [Qwen3.5 9B vs Qwen3-8B](https://broadrouter.com/llm-api-pricing/compare/qwen3.5-9b-vs-qwen3-8b/): USD 0.1125 vs USD 0.06 - [Qwen3.6 27B vs Qwen3.5 27B](https://broadrouter.com/llm-api-pricing/compare/qwen3.6-27b-vs-qwen3.5-27b/): USD 0.6875 vs USD 0.5875 - [Qwen3.6 35B A3B vs Qwen3.5 35B A3B](https://broadrouter.com/llm-api-pricing/compare/qwen3.6-35b-a3b-vs-qwen3.5-35b-a3b/): USD 0.2975 vs USD 0.355 - [Gemma 3 27B vs Gemma 3 12B](https://broadrouter.com/llm-api-pricing/compare/gemma-3-27b-it-vs-gemma-3-12b-it/): USD 0.10 vs USD 0.075 - [Gemma 3 12B vs Gemma 3 4B](https://broadrouter.com/llm-api-pricing/compare/gemma-3-12b-it-vs-gemma-3-4b-it/): USD 0.075 vs USD 0.05 - [Llama 3 70B vs Llama 3 8B](https://broadrouter.com/llm-api-pricing/compare/llama-3-70b-instruct-vs-llama-3-8b-instruct/): USD 2.862 vs USD 0.09312 - [Llama 3.1 70B vs Llama 3.1 8B Instruct](https://broadrouter.com/llm-api-pricing/compare/llama-3.1-70b-instruct-vs-llama-3.1-8b-instruct/): USD 0.72 vs USD 0.025 - [Llama 3.2 3B vs Llama 3.2 1B](https://broadrouter.com/llm-api-pricing/compare/llama-3.2-3b-instruct-vs-llama-3.2-1b-instruct/): USD 0.1219 vs USD 0.0705 - [Ministral 3 14B vs Ministral 3 8B](https://broadrouter.com/llm-api-pricing/compare/ministral-14b-2512-vs-ministral-8b-2512/): USD 0.20 vs USD 0.15 - [Qwen2.5 72B vs Qwen 2.5 7B](https://broadrouter.com/llm-api-pricing/compare/qwen2.5-72b-instruct-vs-qwen2.5-7b-instruct/): USD 0.37 vs USD 0.05 - [Qwen3 235B A22B vs Qwen3 32B](https://broadrouter.com/llm-api-pricing/compare/qwen3-235b-a22b-vs-qwen3-32b/): USD 0.40 vs USD 0.13 - [Qwen3 32B vs Qwen3 30B A3B](https://broadrouter.com/llm-api-pricing/compare/qwen3-32b-vs-qwen3-30b-a3b/): USD 0.13 vs USD 0.122 - [Qwen3 30B A3B vs Qwen3-14B](https://broadrouter.com/llm-api-pricing/compare/qwen3-30b-a3b-vs-qwen3-14b/): USD 0.122 vs USD 0.1225 - [Qwen3-14B vs Qwen3-8B](https://broadrouter.com/llm-api-pricing/compare/qwen3-14b-vs-qwen3-8b/): USD 0.1225 vs USD 0.06 - [Qwen3 VL 235B A22B vs Qwen3 VL 30B A3B Instruct](https://broadrouter.com/llm-api-pricing/compare/qwen3-vl-235b-a22b-instruct-vs-qwen3-vl-30b-a3b-instruct/): USD 0.37 vs USD 0.2375 - [Qwen3 VL 30B A3B Instruct vs Qwen3-VL 8B](https://broadrouter.com/llm-api-pricing/compare/qwen3-vl-30b-a3b-instruct-vs-qwen3-vl-8b-instruct/): USD 0.2375 vs USD 0.305 - [Qwen3.5 397B A17B vs Qwen3.5 122B A10B](https://broadrouter.com/llm-api-pricing/compare/qwen3.5-397b-a17b-vs-qwen3.5-122b-a10b/): USD 1.087 vs USD 0.715 - [Qwen3.5 122B A10B vs Qwen3.5 35B A3B](https://broadrouter.com/llm-api-pricing/compare/qwen3.5-122b-a10b-vs-qwen3.5-35b-a3b/): USD 0.715 vs USD 0.355 - [Qwen3.5 35B A3B vs Qwen3.5 27B](https://broadrouter.com/llm-api-pricing/compare/qwen3.5-35b-a3b-vs-qwen3.5-27b/): USD 0.355 vs USD 0.5875 - [Qwen3.5 27B vs Qwen3.5 9B](https://broadrouter.com/llm-api-pricing/compare/qwen3.5-27b-vs-qwen3.5-9b/): USD 0.5875 vs USD 0.1125 - [Qwen3.6 35B A3B vs Qwen3.6 27B](https://broadrouter.com/llm-api-pricing/compare/qwen3.6-35b-a3b-vs-qwen3.6-27b/): USD 0.2975 vs USD 0.6875 - [Qwen3.8 2.4T A95B vs Qwen3.8 27B](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-2.4t-a95b-vs-qwen3.8-27b/): USD 1.75 vs USD 0.3417 - [DeepSeek V3 0324 vs DeepSeek V3](https://broadrouter.com/llm-api-pricing/compare/deepseek-chat-v3-0324-vs-deepseek-chat/): USD 0.5025 vs USD 0.4375 - [DeepSeek R1 0528 vs DeepSeek R1](https://broadrouter.com/llm-api-pricing/compare/deepseek-r1-0528-vs-deepseek-r1/): USD 0.695 vs USD 0.92 - [DeepSeek V4 Flash 0731 vs DeepSeek V4 Flash](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4-flash-0731-vs-deepseek-v4-flash/): USD 0.09 vs USD 0.1125 - [DeepSeek V4 Pro 0813 vs DeepSeek V4 Pro](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4-pro-0813-vs-deepseek-v4-pro/): USD 1.30 vs USD 1.196 - [DeepSeek V4.1 Flash 1017 vs DeepSeek V4.1 Flash](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-1017-vs-deepseek-v4.1-flash/): USD 0.147 vs USD 0.21 - [Kimi K2 0905 vs Kimi K2](https://broadrouter.com/llm-api-pricing/compare/kimi-k2-0905-vs-kimi-k2/): USD 1.075 vs USD 1.002 - [Qwen3 235B A22B 2507 vs Qwen3 235B A22B](https://broadrouter.com/llm-api-pricing/compare/qwen3-235b-a22b-2507-vs-qwen3-235b-a22b/): USD 0.1531 vs USD 0.40 - [Qwen3 235B A22B Thinking 2507 vs Qwen3 235B A22B 2507](https://broadrouter.com/llm-api-pricing/compare/qwen3-235b-a22b-thinking-2507-vs-qwen3-235b-a22b-2507/): USD 1.212 vs USD 0.1531 - [Qwen3-30B-A3B-2507 vs Qwen3 30B A3B](https://broadrouter.com/llm-api-pricing/compare/qwen3-30b-a3b-instruct-2507-vs-qwen3-30b-a3b/): USD 0.1425 vs USD 0.122 - [Qwen3 Next 80B A3B Thinking vs Qwen3 Next 80B A3B](https://broadrouter.com/llm-api-pricing/compare/qwen3-next-80b-a3b-thinking-vs-qwen3-next-80b-a3b-instruct/): USD 0.4125 vs USD 0.3425 - [Qwen3.8 Max (0902) vs Qwen3.8 Max](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-max-0902-vs-qwen3.8-max/): USD 2.475 vs USD 2.475 - [Qwen3.8 27B vs DeepSeek-V4-Flash-Vision-Exp](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-27b-vs-deepseek-v4-flash-vision-exp/): USD 0.3417 vs USD 0.3234 - [DeepSeek V4.1 Flash vs Ministral 3 14B](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-vs-ministral-14b-2512/): USD 0.21 vs USD 0.20 - [DeepSeek V4.1 Flash vs Qwen3 VL 30B A3B Instruct](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-vs-qwen3-vl-30b-a3b-instruct/): USD 0.21 vs USD 0.2375 - [Gemma 4 26B A4B vs GPT-5 Nano](https://broadrouter.com/llm-api-pricing/compare/gemma-4-26b-a4b-it-vs-gpt-5-nano/): USD 0.1375 vs USD 0.1375 - [Gemma 4 26B A4B vs MiMo V2.6 Flash](https://broadrouter.com/llm-api-pricing/compare/gemma-4-26b-a4b-it-vs-mimo-v2.6-flash/): USD 0.1375 vs USD 0.14 - [Gemma 4 26B A4B vs Qwen3-30B-A3B-2507](https://broadrouter.com/llm-api-pricing/compare/gemma-4-26b-a4b-it-vs-qwen3-30b-a3b-instruct-2507/): USD 0.1375 vs USD 0.1425 - [Gemma 4 26B A4B vs GLM 4.7 Flash](https://broadrouter.com/llm-api-pricing/compare/gemma-4-26b-a4b-it-vs-glm-4.7-flash/): USD 0.1375 vs USD 0.145 - [Qwen3.8 27B vs GLM 4.5 Air](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-27b-vs-glm-4.5-air/): USD 0.3417 vs USD 0.31 - [GLM 5.3 vs Kimi K2 0905](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-kimi-k2-0905/): USD 1.047 vs USD 1.075 - [GLM 5.3 vs Kimi K2](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-kimi-k2/): USD 1.047 vs USD 1.002 - [GLM 5.3 vs Qwen3.6 Plus](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-qwen3.6-plus/): USD 1.047 vs USD 1.125 - [GLM 5.3 vs DeepSeek R1](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-deepseek-r1/): USD 1.047 vs USD 0.92 - [Qwen3.8 27B vs GPT-5.6 Luna](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-27b-vs-gpt-5.6-luna/): USD 0.3417 vs USD 0.45 - [Qwen3.8 27B vs Inkling 2 Small](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-27b-vs-inkling-2-small/): USD 0.3417 vs USD 0.4512 - [GLM 5.3 vs Kimi K2.7 Code](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-kimi-k2.7-code/): USD 1.047 vs USD 1.20 - [Gemma 4 26B A4B vs Qwen3 32B](https://broadrouter.com/llm-api-pricing/compare/gemma-4-26b-a4b-it-vs-qwen3-32b/): USD 0.1375 vs USD 0.13 - [Qwen3.8 27B vs Step 3.7 Flash](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-27b-vs-step-3.7-flash/): USD 0.3417 vs USD 0.4375 - [Qwen3.8 27B vs MiniMax M3.1 Flash](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-27b-vs-minimax-m3.1-flash/): USD 0.3417 vs USD 0.2625 - [DeepSeek V4.1 Flash vs Gemma 4 31B](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-vs-gemma-4-31b-it/): USD 0.21 vs USD 0.16 - [DeepSeek V4.1 Flash vs Qwen3.6 35B A3B](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-vs-qwen3.6-35b-a3b/): USD 0.21 vs USD 0.2975 - [DeepSeek V4.1 Flash vs MiniMax M3](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-vs-minimax-m3/): USD 0.21 vs USD 0.4125 - [DeepSeek V4.1 Flash vs Qwen3.5 35B A3B](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-vs-qwen3.5-35b-a3b/): USD 0.21 vs USD 0.355 - [GLM 5.3 vs DeepSeek V4 Pro 0813](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-deepseek-v4-pro-0813/): USD 1.047 vs USD 1.30 - [GLM 5.3 vs Kimi K2.6](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-kimi-k2.6/): USD 1.047 vs USD 0.7425 - [GLM 5.3 vs Nemotron 3 Ultra](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-nemotron-3-ultra-550b-a55b/): USD 1.047 vs USD 0.9125 - [GLM 5.3 vs Kimi K2.5](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-vs-kimi-k2.5/): USD 1.047 vs USD 0.90 - [Qwen3.8 27B vs DeepSeek V3.2](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-27b-vs-deepseek-v3.2/): USD 0.3417 vs USD 0.234 - [Qwen3.8 27B vs MiniMax M3](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-27b-vs-minimax-m3/): USD 0.3417 vs USD 0.4125 - [Qwen3.8 27B vs Llama 4 Maverick](https://broadrouter.com/llm-api-pricing/compare/qwen3.8-27b-vs-llama-4-maverick/): USD 0.3417 vs USD 0.3875 - [gpt-oss-120b vs Gemma 4 26B A4B](https://broadrouter.com/llm-api-pricing/compare/gpt-oss-120b-vs-gemma-4-26b-a4b-it/): USD 0.065 vs USD 0.1375 - [GLM 5.3 Flash vs Gemma 4 26B A4B](https://broadrouter.com/llm-api-pricing/compare/glm-5.3-flash-vs-gemma-4-26b-a4b-it/): USD 0.1187 vs USD 0.1375 - [DeepSeek V4.1 Flash 1017 vs Gemma 4 26B A4B](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4.1-flash-1017-vs-gemma-4-26b-a4b-it/): USD 0.147 vs USD 0.1375 - [DeepSeek V4 Flash 0731 vs Gemma 4 26B A4B](https://broadrouter.com/llm-api-pricing/compare/deepseek-v4-flash-0731-vs-gemma-4-26b-a4b-it/): USD 0.09 vs USD 0.1375 - [Llama 3.3 70B vs Gemma 4 26B A4B](https://broadrouter.com/llm-api-pricing/compare/llama-3.3-70b-instruct-vs-gemma-4-26b-a4b-it/): USD 0.155 vs USD 0.1375 ## LLM API pricing by provider - [All providers: model counts, cheapest model and published tiers](https://broadrouter.com/llm-api-pricing/providers/) - [AkashML API pricing](https://broadrouter.com/llm-api-pricing/providers/akashml/): 7 models, from USD 0.018 input / USD 0.09 output per 1M tokens; cheapest provider for 1 of 7 models it shares with others; checked 2026-11-15 - [Alibaba Cloud Model Studio API pricing](https://broadrouter.com/llm-api-pricing/providers/alibaba/): 42 models, from USD 0.03 input / USD 0.13 output per 1M tokens; cheapest provider for 1 of 26 models it shares with others; checked 2026-11-14 - [Amazon Bedrock API pricing](https://broadrouter.com/llm-api-pricing/providers/amazon-bedrock/): 108 models, from USD 0.04 input / USD 0.04 output per 1M tokens; cheapest provider for 7 of 68 models it shares with others; checked 2026-11-15 - [Amazon Nova on Bedrock API pricing](https://broadrouter.com/llm-api-pricing/providers/amazon-nova/): 7 models, from USD 0.035 input / USD 0.14 output per 1M tokens; cheapest provider for 0 of 0 models it shares with others; checked 2026-11-15 - [Anthropic API pricing](https://broadrouter.com/llm-api-pricing/providers/anthropic/): 18 models, from USD 0.10 input / USD 0.50 output per 1M tokens; cheapest provider for 15 of 15 models it shares with others; checked 2026-11-14 - [Baidu API pricing](https://broadrouter.com/llm-api-pricing/providers/baidu/): 14 models, from USD 0.14 input / USD 0.28 output per 1M tokens; cheapest provider for 1 of 11 models it shares with others; checked 2026-11-15 - [Baseten API pricing](https://broadrouter.com/llm-api-pricing/providers/baseten/): 14 models, from USD 0.13 input / USD 0.26 output per 1M tokens; cheapest provider for 2 of 14 models it shares with others; checked 2026-11-14 - [Cerebras API pricing](https://broadrouter.com/llm-api-pricing/providers/cerebras/): 2 models, from USD 0.35 input / USD 0.75 output per 1M tokens; cheapest provider for 0 of 2 models it shares with others; checked 2026-11-14 - [Chutes API pricing](https://broadrouter.com/llm-api-pricing/providers/chutes/): 18 models, from USD 0.12 input / USD 0.37 output per 1M tokens; cheapest provider for 8 of 18 models it shares with others; checked 2026-11-15 - [Cloudflare Workers AI API pricing](https://broadrouter.com/llm-api-pricing/providers/cloudflare/): 40 models, from USD 0.017 input / USD 0.112 output per 1M tokens; cheapest provider for 4 of 29 models it shares with others; checked 2026-11-14 - [Cohere API pricing](https://broadrouter.com/llm-api-pricing/providers/cohere/): 6 models, from USD 0.0375 input / USD 0.15 output per 1M tokens; cheapest provider for 1 of 1 models it shares with others; checked 2026-11-14 - [CoreWeave Forge (formerly W&B Inference) API pricing](https://broadrouter.com/llm-api-pricing/providers/coreweave/): 20 models, from USD 0.03 input / USD 0.13 output per 1M tokens; cheapest provider for 7 of 20 models it shares with others; checked 2026-11-15 - [Crusoe Cloud API pricing](https://broadrouter.com/llm-api-pricing/providers/crusoe/): 9 models, from USD 0.05 input / USD 0.20 output per 1M tokens; cheapest provider for 1 of 8 models it shares with others; checked 2026-11-15 - [DeepInfra API pricing](https://broadrouter.com/llm-api-pricing/providers/deepinfra/): 92 models, from USD 0.02 input / USD 0.04 output per 1M tokens; cheapest provider for 37 of 81 models it shares with others; checked 2026-11-14 - [DeepSeek API pricing](https://broadrouter.com/llm-api-pricing/providers/deepseek/): 2 models, from USD 0.30 input / USD 1.20 output per 1M tokens; cheapest provider for 0 of 2 models it shares with others; checked 2026-11-14 - [DigitalOcean Gradient API pricing](https://broadrouter.com/llm-api-pricing/providers/digitalocean/): 59 models, from USD 0.05 input / USD 0.40 output per 1M tokens; cheapest provider for 0 of 57 models it shares with others; checked 2026-11-14 - [Featherless API pricing](https://broadrouter.com/llm-api-pricing/providers/featherless/): 87 models, from USD 0.05 input / USD 0.10 output per 1M tokens; cheapest provider for 13 of 86 models it shares with others; checked 2026-11-15 - [Fireworks AI API pricing](https://broadrouter.com/llm-api-pricing/providers/fireworks/): 15 models, from USD 0.05 input / USD 0.20 output per 1M tokens; cheapest provider for 0 of 13 models it shares with others; checked 2026-11-14 - [FriendliAI API pricing](https://broadrouter.com/llm-api-pricing/providers/friendli/): 5 models, from USD 0.14 input / USD 0.40 output per 1M tokens; cheapest provider for 0 of 5 models it shares with others; checked 2026-11-15 - [GMI Cloud API pricing](https://broadrouter.com/llm-api-pricing/providers/gmi-cloud/): 97 models, from USD 0.091 input / USD 0.182 output per 1M tokens; cheapest provider for 23 of 91 models it shares with others; checked 2026-11-15 - [Gemini API API pricing](https://broadrouter.com/llm-api-pricing/providers/google/): 12 models, from USD 0.10 input / USD 0.40 output per 1M tokens; cheapest provider for 11 of 11 models it shares with others; checked 2026-11-14 - [Vertex AI API pricing](https://broadrouter.com/llm-api-pricing/providers/google-vertex/): 63 models, from USD 0.07 input / USD 0.25 output per 1M tokens; cheapest provider for 3 of 54 models it shares with others; checked 2026-11-14 - [Groq API pricing](https://broadrouter.com/llm-api-pricing/providers/groq/): 4 models, from USD 0.075 input / USD 0.30 output per 1M tokens; cheapest provider for 0 of 4 models it shares with others; checked 2026-11-14 - [Hyperbolic API pricing](https://broadrouter.com/llm-api-pricing/providers/hyperbolic/): 11 models, from USD 0.06 input / USD 0.18 output per 1M tokens; cheapest provider for 0 of 11 models it shares with others; checked 2026-11-14 - [IBM watsonx.ai API pricing](https://broadrouter.com/llm-api-pricing/providers/ibm-watsonx/): 9 models, from USD 0.0636 input / USD 0.265 output per 1M tokens; cheapest provider for 0 of 7 models it shares with others; checked 2026-11-15 - [Azure OpenAI and Foundry Models API pricing](https://broadrouter.com/llm-api-pricing/providers/azure/): 97 models, from USD 0.06 input / USD 0.22 output per 1M tokens; cheapest provider for 0 of 53 models it shares with others; checked 2026-11-14 - [MiniMax API pricing](https://broadrouter.com/llm-api-pricing/providers/minimax/): 2 models, from USD 0.15 input / USD 0.60 output per 1M tokens; cheapest provider for 1 of 2 models it shares with others; checked 2026-11-14 - [Mistral AI API pricing](https://broadrouter.com/llm-api-pricing/providers/mistral/): 9 models, from USD 0.10 input / USD 0.10 output per 1M tokens; cheapest provider for 3 of 6 models it shares with others; checked 2026-11-15 - [Moonshot AI (Kimi) API pricing](https://broadrouter.com/llm-api-pricing/providers/moonshot/): 4 models, from USD 0.95 input / USD 4 output per 1M tokens; cheapest provider for 0 of 4 models it shares with others; checked 2026-11-14 - [Novita AI API pricing](https://broadrouter.com/llm-api-pricing/providers/novita/): 67 models, from USD 0.02 input / USD 0.05 output per 1M tokens; cheapest provider for 6 of 58 models it shares with others; checked 2026-11-14 - [Nscale API pricing](https://broadrouter.com/llm-api-pricing/providers/nscale/): 4 models, from USD 0.15 input / USD 0.47 output per 1M tokens; cheapest provider for 0 of 3 models it shares with others; checked 2026-11-15 - [OpenAI API pricing](https://broadrouter.com/llm-api-pricing/providers/openai/): 37 models, from USD 0.05 input / USD 0.40 output per 1M tokens; cheapest provider for 27 of 27 models it shares with others; checked 2026-11-14 - [Parasail API pricing](https://broadrouter.com/llm-api-pricing/providers/parasail/): 39 models, from USD 0.04 input / USD 0.20 output per 1M tokens; cheapest provider for 1 of 33 models it shares with others; checked 2026-11-14 - [Perplexity API pricing](https://broadrouter.com/llm-api-pricing/providers/perplexity/): 24 models, from USD 0.10 input / USD 0.50 output per 1M tokens; cheapest provider for 0 of 22 models it shares with others; checked 2026-11-15 - [SambaNova Cloud API pricing](https://broadrouter.com/llm-api-pricing/providers/sambanova/): 5 models, from USD 0.22 input / USD 0.59 output per 1M tokens; cheapest provider for 0 of 5 models it shares with others; checked 2026-11-14 - [SiliconFlow API pricing](https://broadrouter.com/llm-api-pricing/providers/siliconflow/): 56 models, from USD 0.05 input / USD 0.05 output per 1M tokens; cheapest provider for 11 of 47 models it shares with others; checked 2026-11-15 - [Together AI API pricing](https://broadrouter.com/llm-api-pricing/providers/together/): 25 models, from USD 0.15 input / USD 0.15 output per 1M tokens; cheapest provider for 1 of 21 models it shares with others; checked 2026-11-14 - [Venice API pricing](https://broadrouter.com/llm-api-pricing/providers/venice/): 57 models, from USD 0.05 input / USD 0.13 output per 1M tokens; cheapest provider for 2 of 45 models it shares with others; checked 2026-11-15 - [xAI API pricing](https://broadrouter.com/llm-api-pricing/providers/xai/): 8 models, from USD 1 input / USD 2 output per 1M tokens; cheapest provider for 6 of 6 models it shares with others; checked 2026-11-14 - [Z.ai API pricing](https://broadrouter.com/llm-api-pricing/providers/z-ai/): 21 models, from USD 0.10 input / USD 0.10 output per 1M tokens; cheapest provider for 4 of 12 models it shares with others; checked 2026-11-14 ## Open data (CC BY 4.0) - [All open datasets](https://broadrouter.com/data/). Each dataset also has Croissant 1.0 (MLCommons) metadata at https://broadrouter.com/data/.croissant.json: files with SHA-256 checksums and typed CSV columns, loadable with the mlcroissant Python library - [Tokenizer efficiency across languages](https://broadrouter.com/data/tokenizer-efficiency/): tokens for the same 2,009 FLORES-200 sentences (dev + devtest) in Arabic MSA, Egyptian, Levantine, Najdi and Mesopotamian Arabic, Hindi, Urdu, Chinese, French, Spanish and English, with 95% bootstrap intervals, plus Gulf Arabic and Arabizi (Tatoeba), Wikipedia prose, code and JSON, under 16 official tokenizers pinned to commits: DeepSeek, GLM, Qwen3.8, Gemma 4, GPT o200k, gpt-oss, Llama 3, Llama 4, Kimi, MiniMax, Gemma 3, Phi-4, Mistral Small 3.1, Ministral 3, Nemotron 3 and Qwen3. CSV: https://broadrouter.com/data/tokenizer-efficiency.csv · JSON: https://broadrouter.com/data/tokenizer-efficiency.json · Croissant 1.0 metadata: https://broadrouter.com/data/tokenizer-efficiency.croissant.json - [Real LLM API cost by language](https://broadrouter.com/data/real-cost-by-language/): what the same text costs to send and generate in English, Arabic MSA and dialects, Arabizi, Hindi, Urdu, Chinese, French, Spanish, code and JSON, per 1M characters and per 1,000 words, for every standard-tier provider price of a model whose tokenizer we measured (model-to-tokenizer matches checked against each model's own repo), with BroadNet rows labelled; cheapest provider per language, language tax in money, and where fewer tokens is not cheaper. CSV: https://broadrouter.com/data/real-cost-by-language.csv · JSON: https://broadrouter.com/data/real-cost-by-language.json · Croissant 1.0 metadata: https://broadrouter.com/data/real-cost-by-language.croissant.json - [LLM API price history](https://broadrouter.com/data/llm-api-price-history/): 3,332 list prices for 577 models from 11 providers (OpenAI, Anthropic, Google Gemini, DeepSeek, Mistral, xAI, Cohere, Groq, Together, Fireworks, DeepInfra), read verbatim from 308 Internet Archive captures of their own pricing pages, 2023-01 to 2026-11; every row cites its archive URL and a quote, 109 skipped captures listed with reasons, 46 rows hand-checked. Price cuts and rises per model. CSV: https://broadrouter.com/data/llm-api-price-history.csv · JSON: https://broadrouter.com/data/llm-api-price-history.json · Croissant 1.0 metadata: https://broadrouter.com/data/llm-api-price-history.croissant.json - [LLM API price index](https://broadrouter.com/data/llm-api-price-index/): median, cheapest and dearest prices per model from providers' own pricing pages (source and date per price), provider spread, maker versus host, batch and flex discounts, output/input ratios, cached pricing, context tiers. CSV: https://broadrouter.com/data/llm-api-price-index.csv · JSON: https://broadrouter.com/data/llm-api-price-index.json · Croissant 1.0 metadata: https://broadrouter.com/data/llm-api-price-index.croissant.json - [LLM API provider price database](https://broadrouter.com/data/llm-api-provider-prices/): every LLM API price we hold, one row per price (provider, model, provider model ID, tier, input, cached and output per 1M tokens, context, precision), each with its source URL, date checked and a verbatim quote from the provider's own page. CSV: https://broadrouter.com/data/llm-api-provider-prices.csv · JSON: https://broadrouter.com/data/llm-api-provider-prices.json · Croissant 1.0 metadata: https://broadrouter.com/data/llm-api-provider-prices.croissant.json - [Speech-to-text API price index](https://broadrouter.com/data/speech-to-text-api-prices/): batch and streaming price per audio hour, billing increment, minimum charge, Arabic support and dialects, diarization and free tier for OpenAI, Google Cloud, AWS, Azure, Deepgram, AssemblyAI, Speechmatics, Soniox, Rev.ai, Gladia, ElevenLabs, Groq, DeepInfra and BroadNet, read from each provider's own pages with source and date per row. CSV: https://broadrouter.com/data/speech-to-text-api-prices.csv · JSON: https://broadrouter.com/data/speech-to-text-api-prices.json · Croissant 1.0 metadata: https://broadrouter.com/data/speech-to-text-api-prices.croissant.json - [LLM API free tiers and rate limits](https://broadrouter.com/data/llm-api-free-tiers-and-rate-limits/): free tier or credit, default RPM/TPM/RPD or concurrency for new accounts, card requirement and free-tier training terms for 14 providers, from each provider's own docs with source and date per row. CSV: https://broadrouter.com/data/llm-api-free-tiers-and-rate-limits.csv · JSON: https://broadrouter.com/data/llm-api-free-tiers-and-rate-limits.json · Croissant 1.0 metadata: https://broadrouter.com/data/llm-api-free-tiers-and-rate-limits.croissant.json ## Free tools (run in the browser, no sign-up) - [All free tools](https://broadrouter.com/tools/) - [LLM API cost calculator](https://broadrouter.com/tools/llm-api-cost-calculator/): Price a workload on every model at once, with cache hits, reasoning and a flat-plan check. - [Speech-to-text cost calculator](https://broadrouter.com/tools/speech-to-text-cost-calculator/): Hours, calls or voice notes to a monthly transcription bill, against your current provider. - [LLM token counter](https://broadrouter.com/tools/token-counter/): Exact token counts with each model's real tokenizer, with a token view and cost. - [Arabic token counter](https://broadrouter.com/tools/arabic-token-counter/): Arabic vs English token counts per tokenizer: the "Arabic tax" in tokens and money. - [Tokens to words converter](https://broadrouter.com/tools/tokens-to-words/): Tokens ↔ words ↔ characters ↔ pages, by language and content type. - [Context window calculator](https://broadrouter.com/tools/context-window-calculator/): Will it fit? Window used per model, room left for output, cost to send. - [Flat-rate vs per-token LLM cost calculator](https://broadrouter.com/tools/flat-rate-vs-per-token-calculator/): Break-even between an unlimited plan and paying per token, per agent. - [Prompt caching savings calculator](https://broadrouter.com/tools/prompt-caching-calculator/): Bill with and without caching, by prefix share and hit rate. - [Image token calculator](https://broadrouter.com/tools/image-token-calculator/): Tokens and cost per image under each model's documented rule. Drop an image; nothing is uploaded. - [Self-hosting vs API cost calculator](https://broadrouter.com/tools/self-hosting-vs-api-calculator/): GPU rental against your current API and BroadNet per-token and flat-plan prices: cost per 1M tokens and break-even volume. - [Embedding cost calculator](https://broadrouter.com/tools/embedding-cost-calculator/): Index, update and query costs for a corpus at OpenAI, Google, Voyage, Mistral and Jina list prices. - [LLM rate limit and concurrency calculator](https://broadrouter.com/tools/llm-rate-limit-calculator/): Which limit binds first: RPM, TPM or concurrency, with Little's law. - [JSON schema generator for LLM structured output](https://broadrouter.com/tools/json-schema-generator/): Sample JSON to strict schema, Pydantic, Zod and a ready request. - [OpenAI-compatible API request builder](https://broadrouter.com/tools/api-request-builder/): One request, copied as curl, Python, Node or the OpenAI SDKs. - [Audio file splitter for speech-to-text limits](https://broadrouter.com/tools/audio-splitter-ffmpeg/): Chunk count, overlap and copy-ready ffmpeg commands to fit a transcription file-size limit. - [Agent run cost estimator](https://broadrouter.com/tools/agent-run-cost-estimator/): Context growth per step, total tokens and cost of one agent run, with a DeepSeek session comparison. - [OpenAI to BroadNet migration helper](https://broadrouter.com/tools/openai-to-broadnet-migration/): Paste OpenAI SDK code; get base URL, key and model edits, a diff and the checks that remain. - [Arabic text normalizer](https://broadrouter.com/tools/arabic-text-normalizer/): Tashkeel, tatweel, alef, ya, ta marbuta, digits and invisible characters, with a matching Python function. - [RAG chunking calculator](https://broadrouter.com/tools/rag-chunking-calculator/): Chunks, embedding tokens, input per query, context check and monthly LLM cost for a RAG pipeline. - [max_tokens calculator for reasoning models](https://broadrouter.com/tools/reasoning-token-budget/): Answer length + reasoning budget + margin = a max_tokens that does not cut the answer off. - [LLM batch job planner](https://broadrouter.com/tools/llm-batch-job-planner/): Run time, binding limit, cost per model and the flat sessions needed to finish a big job by a deadline. - [LLM JSON output validator](https://broadrouter.com/tools/llm-json-output-validator/): Parse errors by line and column, safe repairs, cut-off detection and JSON Schema checks for model output. - [Prompt token diff](https://broadrouter.com/tools/prompt-token-diff/): Two prompt versions side by side: changed lines, exact token delta and the monthly cost difference. - [DeepSeek token counter](https://broadrouter.com/tools/deepseek-token-counter/): exact counts with the DeepSeek V4.1 Flash tokenizer, run in the browser - [GLM token counter](https://broadrouter.com/tools/glm-token-counter/): exact counts with the GLM 5.3 tokenizer, run in the browser - [Qwen token counter](https://broadrouter.com/tools/qwen-token-counter/): exact counts with the Qwen3.8 27B tokenizer, run in the browser - [Gemma token counter](https://broadrouter.com/tools/gemma-token-counter/): exact counts with the Gemma 4 tokenizer (Google base model), run in the browser - [GPT token counter](https://broadrouter.com/tools/gpt-token-counter/): exact counts with the OpenAI o200k_base, run in the browser - [Llama 3 token counter](https://broadrouter.com/tools/llama-token-counter/): exact counts with the official Llama 3 tokenizer (Llama 3.3 70B Instruct), run in the browser - [Llama 4 token counter](https://broadrouter.com/tools/llama-4-token-counter/): exact counts with the official Llama 4 tokenizer (Llama 4 Maverick), run in the browser - [Kimi token counter](https://broadrouter.com/tools/kimi-token-counter/): exact counts with the official Kimi tiktoken tokenizer (Kimi K3), run in the browser - [Mistral Small token counter](https://broadrouter.com/tools/mistral-token-counter/): exact counts with the official Mistral Small 3.1 tokenizer (Tekken), run in the browser - [Ministral 3 token counter](https://broadrouter.com/tools/ministral-token-counter/): exact counts with the official Ministral 3 tokenizer (Tekken, Ministral 3 14B), run in the browser - [gpt-oss token counter](https://broadrouter.com/tools/gpt-oss-token-counter/): exact counts with the official gpt-oss tokenizer (o200k_harmony), run in the browser - [Phi-4 token counter](https://broadrouter.com/tools/phi-4-token-counter/): exact counts with the official Phi-4 tokenizer, run in the browser - [Gemma 3 token counter](https://broadrouter.com/tools/gemma-3-token-counter/): exact counts with the official Gemma 3 tokenizer (Gemma 3 27B IT), run in the browser - [MiniMax token counter](https://broadrouter.com/tools/minimax-token-counter/): exact counts with the official MiniMax M3 tokenizer, run in the browser - [Nemotron 3 token counter](https://broadrouter.com/tools/nemotron-token-counter/): exact counts with the official Nemotron 3 tokenizer (Nemotron 3 Ultra), run in the browser - [Qwen3 token counter](https://broadrouter.com/tools/qwen3-token-counter/): exact counts with the official Qwen3 tokenizer (151k vocabulary, Qwen3 32B), run in the browser ## Guides: Start - [Use the OpenAI Python SDK with BroadNet Inference](https://broadrouter.com/guides/openai-python-sdk/): Point the official OpenAI Python SDK at BroadNet Inference with one base_url change. Install, set your key, make a first request and stream the answer. - [Use the OpenAI Node.js SDK with BroadNet Inference](https://broadrouter.com/guides/openai-node-sdk/): Point the official OpenAI Node.js and TypeScript SDK at BroadNet Inference by setting baseURL. Install it, make a first request, stream tokens and log request IDs. - [Try BroadNet Inference in your browser, no account](https://broadrouter.com/guides/try-broadnet-inference-playground/): Test Arabic speech-to-text, document reading and Gemma 4 chat in BroadNet's browser playground with no sign-up or card, then move the same request to the API. - [First request to BroadNet Inference with curl](https://broadrouter.com/guides/curl-quickstart/): Make your first BroadNet Inference calls with curl - a chat completion, a streamed reply and an audio transcription - and read the response, usage and errors. ## Guides: Reference - [Stream chat completions and handle dropped connections](https://broadrouter.com/guides/streaming-chat-completions/): How streaming works on BroadNet Inference - Server-Sent Events, the [DONE] marker, token usage in the last chunk, and what to do when a stream ends early. - [Why a reasoning model returns an empty answer](https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3 and Qwen3.8 27B think before they answer, and that reasoning counts toward max_tokens. How to stop getting empty replies. - [BroadNet Inference error codes and when to retry](https://broadrouter.com/guides/errors-and-retries/): What each HTTP status from the BroadNet Inference API means, which ones to retry with backoff, which to fix first, and what to send support with a request ID. - [Lock API keys to your servers with IP allow-lists](https://broadrouter.com/guides/api-keys-ip-allow-lists/): Protect BroadNet Inference keys with IP allow-lists, separate inference keys from portal credentials, and revoke a leaked key in seconds. With curl examples. ## Guides: Agents and plans - [How the unlimited LLM API plans work for agents](https://broadrouter.com/guides/unlimited-llm-api-sessions/): The unlimited DeepSeek V4.1 Flash plans have no token meter and no weekly limit. Here is how sessions, parallel agents, keys and billing actually work. ## Guides: Integrations - [Run Aider on unlimited DeepSeek V4.1 Flash](https://broadrouter.com/guides/aider-unlimited-deepseek/): Connect the Aider coding agent to BroadNet Inference through its OpenAI-compatible setting and run it on a flat-price, unlimited DeepSeek V4.1 Flash plan. - [Use LangChain with BroadNet Inference](https://broadrouter.com/guides/langchain-openai-base-url/): Connect LangChain's ChatOpenAI to BroadNet Inference with base_url and api_key, and learn what LangChain does and doesn't keep from reasoning models. - [Use Cline with BroadNet Inference (OpenAI Compatible provider)](https://broadrouter.com/guides/cline-openai-compatible/): Set up the Cline VS Code extension with BroadNet Inference through its OpenAI Compatible provider, with exact fields, model IDs, context sizes and common fixes. - [Use Roo Code with BroadNet Inference (OpenAI Compatible)](https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code was shut down in May 2026. Settings for existing installs and forks, and where to move next (Cline or the Zoo Code fork) with BroadNet Inference. - [Use Continue with BroadNet Inference (config.yaml and apiBase)](https://broadrouter.com/guides/continue-dev-openai-compatible/): Add BroadNet Inference models to the Continue IDE extension with provider openai and apiBase in config.yaml, including tool_use capabilities and context length. - [Add BroadNet Inference to opencode as a custom provider](https://broadrouter.com/guides/opencode-custom-provider/): Add BroadNet Inference to the opencode coding agent as a custom OpenAI-compatible provider in opencode.jsonc, with model aliases, capabilities and limits. - [Connect Open WebUI to BroadNet Inference](https://broadrouter.com/guides/open-webui-openai-api/): Add BroadNet Inference to Open WebUI as an OpenAI-compatible connection, add model IDs to the allowlist, and fix verify, empty-reply and rate-limit issues. - [Add BroadNet Inference to LibreChat as a custom endpoint](https://broadrouter.com/guides/librechat-custom-endpoint/): Add BroadNet Inference to LibreChat as an OpenAI-compatible custom endpoint in librechat.yaml, with the key in .env, the model list, title model and fixes. - [Use BroadNet Inference in n8n (OpenAI credential, Base URL)](https://broadrouter.com/guides/n8n-openai-compatible/): Point n8n's OpenAI credential at BroadNet Inference with the Base URL field, then use the OpenAI Chat Model node by ID with the Responses API toggle turned off. - [Use the Vercel AI SDK with BroadNet Inference](https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): Call BroadNet Inference from the Vercel AI SDK with createOpenAICompatible from @ai-sdk/openai-compatible. generateText, streamText, tools and reasoning effort. - [Route LiteLLM to BroadNet Inference (SDK and proxy)](https://broadrouter.com/guides/litellm-proxy-custom-provider/): Call BroadNet Inference through LiteLLM with the openai/ model prefix and api_base, in the Python SDK or the proxy config.yaml, with matching rate limits. - [Run the OpenAI Agents SDK on BroadNet Inference](https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Use the OpenAI Agents SDK for Python with BroadNet Inference through AsyncOpenAI and OpenAIChatCompletionsModel, and turn off tracing that needs an OpenAI key. - [Transcribe WhatsApp voice notes in n8n](https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): An n8n workflow that receives a WhatsApp voice note, downloads it from Meta, converts it with ffmpeg and transcribes it in Arabic or 13 other languages. - [Use Zed with BroadNet Inference as an OpenAI-compatible provider](https://broadrouter.com/guides/zed-openai-compatible-provider/): Add BroadNet Inference to the Zed editor as an OpenAI-compatible provider in settings.json, with model IDs, token limits, capabilities and the key variable. - [Use Kilo Code with BroadNet Inference (OpenAI Compatible provider)](https://broadrouter.com/guides/kilo-code-openai-compatible/): Connect Kilo Code to BroadNet Inference as a custom OpenAI Compatible provider, in the VS Code dialog or kilo.json, with model IDs, limits and tool flags. - [Add BroadNet Inference to goose as a custom provider](https://broadrouter.com/guides/goose-custom-provider/): Connect the goose AI agent to BroadNet Inference as an OpenAI Compatible custom provider, from goose Desktop, goose configure or a JSON file. - [Use Pydantic AI with BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Run Pydantic AI agents on BroadNet Inference with OpenAIChatModel and OpenAICompatibleProvider, plus tool calls, structured output and model profile fixes. - [Use CrewAI with BroadNet Inference as a custom OpenAI endpoint](https://broadrouter.com/guides/crewai-custom-openai-endpoint/): Point CrewAI agents at BroadNet Inference with LLM(custom_openai=True, base_url=...), then pick models for tools, long context and the flat-rate plan. - [Use LlamaIndex with BroadNet Inference through OpenAILike](https://broadrouter.com/guides/llamaindex-openailike/): Connect LlamaIndex to BroadNet Inference with the OpenAILike class, setting api_base, context_window, is_chat_model and is_function_calling_model correctly. - [Add BroadNet Inference to Dify with the OpenAI-API-compatible plugin](https://broadrouter.com/guides/dify-openai-api-compatible/): Connect Dify to BroadNet Inference through the OpenAI-API-compatible model provider plugin, with exact field values, context sizes and function calling settings. - [Use Jan with BroadNet Inference as a custom OpenAI-compatible provider](https://broadrouter.com/guides/jan-custom-endpoint/): Add BroadNet Inference to the Jan desktop app as a custom OpenAI-compatible provider, add model IDs by hand, and switch on tools and vision per model. - [Connect SillyTavern to BroadNet Inference (Custom OpenAI-compatible)](https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): Connect SillyTavern to BroadNet Inference with the Custom (OpenAI-compatible) Chat Completion source, the right model IDs, prompt post-processing and token limits. - [Use Spring AI with BroadNet Inference (OpenAI base-url)](https://broadrouter.com/guides/spring-ai-openai-base-url/): Point Spring AI's OpenAI chat starter at BroadNet Inference with spring.ai.openai.base-url, a BroadNet model ID and max-tokens, then call it from ChatClient. - [Call BroadNet Inference from Go with the openai-go SDK](https://broadrouter.com/guides/openai-go-sdk/): Use the official openai-go v3 SDK with BroadNet Inference through option.WithBaseURL, with chat completions, streaming with usage and error handling in Go. - [Build LangGraph agents on BroadNet Inference](https://broadrouter.com/guides/langgraph-openai-compatible/): Run a LangGraph StateGraph or create_agent on BroadNet Inference with ChatOpenAI, base_url and a BroadNet model ID, plus tools, streaming and memory. - [Use Haystack pipelines and agents with BroadNet Inference](https://broadrouter.com/guides/haystack-openai-compatible/): Point Haystack's OpenAIChatGenerator at BroadNet Inference with api_base_url, a BroadNet model ID and a Secret key, for pipelines, tools, streaming and agents. - [Use Semantic Kernel with BroadNet Inference](https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Connect Microsoft Semantic Kernel to BroadNet Inference in Python with an AsyncOpenAI client, or in C# with a custom endpoint, for chat, plugins and agents. - [Use Mastra agents with BroadNet Inference](https://broadrouter.com/guides/mastra-openai-compatible/): Run Mastra agents on BroadNet Inference with a model object that sets id, url and apiKey, including the doubled broadnet/ prefix the model ID needs. - [Get structured output with Instructor on BroadNet Inference](https://broadrouter.com/guides/instructor-structured-output/): Extract Pydantic objects from BroadNet Inference with Instructor, using from_openai or from_provider with a custom base_url, and pick the right Mode. - [Use DSPy with BroadNet Inference (set model_type="chat")](https://broadrouter.com/guides/dspy-openai-compatible/): Configure dspy.LM for BroadNet Inference with the openai/ prefix, api_base and model_type="chat", because DSPy 3.5 sends OpenAI models to the Responses API. - [Evaluate BroadNet models with promptfoo](https://broadrouter.com/guides/promptfoo-openai-compatible/): Run promptfoo evals and side-by-side model tests on BroadNet Inference with the openai:chat provider, apiBaseUrl and apiKeyEnvar, plus graders. - [Run smolagents on BroadNet Inference with OpenAIModel](https://broadrouter.com/guides/smolagents-openai-compatible/): Connect Hugging Face smolagents to BroadNet Inference with OpenAIModel and api_base, for CodeAgent on any model or ToolCallingAgent on GLM 5.3 and Qwen3.8 27B. - [Connect Flowise to BroadNet Inference (OpenAI Custom Model)](https://broadrouter.com/guides/flowise-openai-compatible/): Use BroadNet Inference in Flowise chatflows with the OpenAI Custom Model node, a Base Path, an OpenAI API credential and a typed BroadNet model name. - [Connect Langflow to BroadNet Inference](https://broadrouter.com/guides/langflow-openai-compatible/): Use BroadNet Inference in Langflow with the OpenAI component's OpenAI API Base field and a typed model name, and why the OpenAI Compatible provider won't list it. ## Guides: Arabic and speech - [Arabic speech-to-text API, step by step](https://broadrouter.com/guides/arabic-speech-to-text-api/): Transcribe Arabic audio with BroadNet Transcribe Arabic through an OpenAI-compatible endpoint. Formats, the language field, Python and curl examples, and the price. - [Transcribe WhatsApp voice notes to text with an API](https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): Convert WhatsApp .ogg voice notes or .m4a phone recordings to WAV with ffmpeg and transcribe them in Arabic or 13 other languages with one API call. - [Read Arabic invoices and receipts with a vision model](https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Send an invoice or receipt photo to a vision model on BroadNet Inference and get the total, date and vendor back as JSON, with questions in Arabic or English. ## Glossary - [AI agent](https://broadrouter.com/glossary/agent/): An AI agent is a system in which a language model directs its own steps toward a goal: it decides which tools to call, reads the results and repeats until it produces a final answer or hits a limit. A fixed script that calls a model at set points is usually called a workflow instead. - [AI coding agent](https://broadrouter.com/glossary/coding-agent/): An AI coding agent is a software tool in which a language model works on a codebase in a loop: it reads files, edits them, runs commands such as tests, reads the output and decides the next step until the task is done or a limit is reached. - [API key](https://broadrouter.com/glossary/api-key/): An API key is a secret string that identifies and authorizes a client calling an API. LLM APIs usually expect it in an Authorization: Bearer header, and anyone who holds the key can make requests that are billed to the account that owns it. - [Arabic diacritics (tashkeel)](https://broadrouter.com/glossary/arabic-diacritics/): Arabic diacritics, or tashkeel, are small marks written above or below letters to show short vowels, doubled consonants and the absence of a vowel. They are optional: most Arabic text, including news, omits them, so readers and software infer the vowels from context. - [Arabic dialect identification](https://broadrouter.com/glossary/arabic-dialect-identification/): Arabic dialect identification is the task of deciding which Arabic variety a text or recording is in, such as Egyptian, Gulf, Levantine, Maghrebi or Modern Standard Arabic, at the level of region, country or city. It is used to route messages, pick prompts and split evaluation sets. - [Arabic text normalization](https://broadrouter.com/glossary/arabic-text-normalization/): Arabic text normalization is the step that maps spelling variants of the same word to one form before matching: removing diacritics and tatweel, unifying alef forms, and often mapping ى to ي and ة to ه. It is applied to search keys and evaluation, not to text you display. - [Arabizi](https://broadrouter.com/glossary/arabizi/): Arabizi is Arabic written informally in Latin letters, using digits for Arabic sounds with no Latin equivalent, such as 3 for ع. It is common in chat and social media, is mostly dialectal, has no standard spelling, and is often mixed with English or French. - [AWQ (activation-aware weight quantization)](https://broadrouter.com/glossary/awq/): AWQ (activation-aware weight quantization) is a post-training method that compresses an LLM's weights to low precision, usually 4-bit integers. It uses activation statistics from a small calibration set to find the most important weight channels and scales them so they lose less accuracy. It needs no backpropagation. - [Base URL](https://broadrouter.com/glossary/base-url/): A base URL is the fixed root address of an API, such as https://api.example.com/v1, that a client puts in front of every endpoint path, such as /chat/completions. In OpenAI-compatible SDKs, changing the base URL, together with the API key and model ID, points the same code at a different provider. - [Batch API](https://broadrouter.com/glossary/batch-api/): A batch API is an asynchronous LLM endpoint that accepts many requests at once, usually as a file, processes them within a set window instead of immediately, and returns the results together. Providers such as OpenAI and Anthropic price batch requests at a discount to their standard APIs. - [BM25](https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) is a ranking function for keyword search that scores a document by how often the query terms appear in it, how rare those terms are across the collection, and how long the document is. It is the default similarity in Lucene and Elasticsearch. - [Byte pair encoding (BPE)](https://broadrouter.com/glossary/bpe/): Byte pair encoding (BPE) is a subword tokenization algorithm that starts from single characters or bytes and repeatedly merges the most frequent adjacent pair into a new vocabulary entry, until it reaches a target vocabulary size. Most current LLM tokenizers use BPE or a byte-level variant. - [Cached input tokens](https://broadrouter.com/glossary/cached-input-tokens/): Cached input tokens are prompt tokens that an LLM provider has recently processed and can reuse from a cache instead of computing again, usually a repeated prefix such as a system prompt or document. Providers bill them at a lower rate than new input tokens. - [Chain-of-thought (CoT)](https://broadrouter.com/glossary/chain-of-thought/): Chain-of-thought (CoT) is a prompting technique in which a language model is shown or asked to write intermediate reasoning steps before its final answer. It improves accuracy on multi-step problems such as arithmetic and logic, and reasoning models now generate it automatically. - [Character error rate (CER)](https://broadrouter.com/glossary/character-error-rate/): Character error rate (CER) measures transcription or OCR accuracy at the character level: the substitutions, deletions and insertions needed to turn the output into the reference text, divided by the number of characters in the reference. Lower is better, and 0 is a perfect match. - [Chat Completions API](https://broadrouter.com/glossary/chat-completions-api/): The Chat Completions API is OpenAI's HTTP endpoint, POST /v1/chat/completions, that takes a model name and a list of role-tagged messages and returns the model's next message. Its request and response shape became the common format that OpenAI-compatible providers and SDKs reuse. - [Chat template](https://broadrouter.com/glossary/chat-template/): A chat template is the formatting rule, usually written in Jinja and shipped with a model's tokenizer, that converts a list of chat messages with roles into the single token sequence the model expects, including its special control tokens. Using the wrong template can sharply degrade a model's answers. - [Chunking (RAG)](https://broadrouter.com/glossary/chunking/): Chunking is the step in a retrieval pipeline that splits documents into smaller pieces, called chunks, before they are embedded and indexed. Chunk size sets how much text each piece holds, and chunk overlap repeats some text between neighbouring chunks so ideas cut at a boundary are not lost. - [Code-switching](https://broadrouter.com/glossary/code-switching/): Code-switching is the use of two or more languages, or varieties of one language, within the same conversation or sentence, in speech or writing. Examples include Arabic speakers mixing English words into dialect, or moving between Modern Standard Arabic and their dialect. - [Concurrency](https://broadrouter.com/glossary/concurrency/): Concurrency is the number of requests a client has in flight at the same time. LLM APIs often cap it per key or plan, separately from requests per minute, because each open request holds GPU capacity until its response has finished. - [Context compaction](https://broadrouter.com/glossary/context-compaction/): Context compaction is the practice of summarizing a conversation that is nearing the context window limit and continuing the session from that summary, so an agent keeps key decisions and state while dropping raw history such as old tool output. - [Context engineering](https://broadrouter.com/glossary/context-engineering/): Context engineering is the work of choosing and maintaining the set of tokens a language model sees on each request, including system prompt, tool definitions, history, retrieved documents and notes, so the model gets the smallest set of high-signal information it needs to do the task well. - [Context rot](https://broadrouter.com/glossary/context-rot/): Context rot is the gradual loss of accuracy and reliability that language models show as their input gets longer, even when the task stays the same and the input is well within the context window. The term was popularised by a 2025 Chroma technical report that tested 18 models. - [Context window](https://broadrouter.com/glossary/context-window/): A context window is the maximum number of tokens a language model can work with in a single request. It covers the prompt, conversation history, tool definitions, images and the generated output, including any reasoning tokens. Anything that does not fit must be cut, summarised or split across several requests. - [Data processing agreement (DPA)](https://broadrouter.com/glossary/dpa/): A data processing agreement (DPA) is the contract between a controller and a processor that governs how the processor handles personal data. GDPR Article 28(3) requires it and lists its contents: documented instructions, confidentiality, security, subprocessor rules, deletion or return of data, and audits. - [Data residency](https://broadrouter.com/glossary/data-residency/): Data residency is a requirement or commitment that data is stored and processed in a specific country or region. For an LLM API it covers where prompts are processed during inference and where any stored content sits, and it often excludes account and billing data. - [Dialectal Arabic](https://broadrouter.com/glossary/dialectal-arabic/): Dialectal Arabic is the everyday spoken Arabic of a region or city, such as Egyptian, Gulf, Levantine or Maghrebi Arabic. Dialects differ from Modern Standard Arabic in sounds, words and grammar, have no standard spelling, and are increasingly written online, which makes them harder for language models. - [Distillation (knowledge distillation)](https://broadrouter.com/glossary/distillation/): Distillation, or knowledge distillation, is a training method in which a smaller student model learns to reproduce the outputs of a larger teacher model, such as its probability distributions or generated answers, so the student keeps much of the teacher's quality at a lower cost to run. - [Embeddings](https://broadrouter.com/glossary/embeddings/): Embeddings are lists of floating-point numbers (vectors) that a model produces to represent a piece of text, an image or other input. Inputs with similar meaning get vectors that are close together, so distance between vectors can rank search results, cluster documents or find related items. - [Eval (LLM evaluation)](https://broadrouter.com/glossary/eval/): An eval (short for evaluation) is a repeatable test of a language model or prompt: a fixed set of inputs, a description of what a good output looks like, and a grader (code, a human or another model) that scores each output so that versions can be compared. - [Exponential backoff](https://broadrouter.com/glossary/exponential-backoff/): Exponential backoff is a retry strategy in which a client waits longer after each failed attempt, typically doubling the delay up to a cap, before trying again. Adding random jitter to each wait spreads out retries from many clients so they do not hit the server at the same moment. - [Few-shot prompting](https://broadrouter.com/glossary/few-shot-prompting/): Few-shot prompting is a technique in which a prompt includes a small number of worked input-output examples before the real input, so a language model infers the task and format from them. The model's weights do not change; zero-shot uses no examples and one-shot uses one. - [Fine-tuning](https://broadrouter.com/glossary/fine-tuning/): Fine-tuning is the process of continuing to train a pre-trained language model on a smaller, task-specific dataset, such as example conversations, so that its weights change and it performs that task or style more reliably without needing long instructions or examples in every prompt. - [finish_reason](https://broadrouter.com/glossary/finish-reason/): finish_reason is a field on each choice in a chat completion response that says why the model stopped generating. In the OpenAI format the values are stop, length, tool_calls, content_filter and the deprecated function_call. Anthropic's Messages API reports the same idea as stop_reason. - [Flat-rate LLM plan](https://broadrouter.com/glossary/flat-rate-llm-plan/): A flat-rate LLM plan charges a fixed price for a period, such as a month, for access to a language model API instead of billing per token. Such plans usually limit something else, such as concurrency, speed, models or usage windows, so the caps matter as much as the price. - [FP8](https://broadrouter.com/glossary/fp8/): FP8 is an 8-bit floating-point number format used in deep learning, defined in two encodings: E4M3, with 4 exponent and 3 mantissa bits, and E5M2, with 5 exponent and 2 mantissa bits. It halves memory compared with 16-bit formats and runs natively on recent GPUs. - [Frequency and presence penalty](https://broadrouter.com/glossary/frequency-and-presence-penalty/): Frequency and presence penalties are sampling settings that lower the score of tokens already in the output. The presence penalty applies once a token has appeared at all; the frequency penalty grows with each repeat. In OpenAI's API both range from -2.0 to 2.0, with 0 as the default. - [GDPR](https://broadrouter.com/glossary/gdpr/): The GDPR (General Data Protection Regulation, Regulation (EU) 2016/679) is the EU law on processing personal data. It has applied since 25 May 2018 and covers EU-established organisations and some outside the EU. For LLM APIs it governs processor contracts, international transfers and data minimisation. - [GGUF](https://broadrouter.com/glossary/gguf/): GGUF is a binary file format for storing language models for inference with GGML-based runtimes such as llama.cpp. One GGUF file holds the weights, often quantized, plus the metadata needed to load the model, including tokenizer and architecture details, and it can be memory-mapped for fast loading. - [GPU-hour](https://broadrouter.com/glossary/gpu-hour/): A GPU-hour is one graphics processor used for one hour. Cloud providers bill rented GPU instances by the hour or second, and researchers use GPU-hours to state how much compute a job took. You pay for the machine's time, whether or not it is busy. - [Grounding](https://broadrouter.com/glossary/grounding/): Grounding is the practice of connecting a language model's output to specific, verifiable sources of information, such as supplied documents, search results or database records, and instructing it to answer from them. Grounded answers can be traced back to a source and are less likely to contain made-up facts. - [Guardrails](https://broadrouter.com/glossary/guardrails/): Guardrails are checks that an application runs around a language model call to keep it within set limits. Input guardrails screen what goes into the model, output guardrails check what comes out, and tool guardrails validate actions before they run. A failed check blocks, rewrites or flags the request. - [Hallucination](https://broadrouter.com/glossary/hallucination/): A hallucination is output from a language model that sounds plausible but is false, made up or not supported by the input it was given, such as an invented citation, a wrong number or a fabricated API. It comes from the model predicting likely text rather than checking facts. - [Hamza](https://broadrouter.com/glossary/hamza/): Hamza (ء) is the Arabic letter for the glottal stop. It is written on its own or on a seat: alef (أ إ), waw (ؤ) or yeh (ئ), and آ combines alef with madda. Because typed text is inconsistent about it, search engines and speech scoring often fold the alef forms into ا. - [Hybrid search](https://broadrouter.com/glossary/hybrid-search/): Hybrid search is a retrieval method that runs keyword search, such as BM25, and vector search over embeddings on the same query, then merges the two ranked lists into one, often with reciprocal rank fusion (RRF) or a weighted sum of normalized scores. - [Idempotency key](https://broadrouter.com/glossary/idempotency-key/): An idempotency key is a unique value, usually a random ID such as a UUID, that a client attaches to a request that changes state, such as a payment or a purchase. If the same request is sent again with the same key, the server returns the original result instead of performing the action twice. - [Image tokens](https://broadrouter.com/glossary/image-tokens/): Image tokens are the input tokens a vision language model uses to represent an image in a prompt. Providers compute them from the image's size, usually by splitting it into fixed-size patches or tiles after resizing, and bill them like text input tokens, so larger images cost more. - [Inference provider](https://broadrouter.com/glossary/inference-provider/): An inference provider is a company that runs machine learning models, often open-weight LLMs, on GPUs it owns, leases or rents, and sells access to them through an API, usually priced per token, per request or per unit of time. It handles serving, scaling and hardware. - [Input and output tokens](https://broadrouter.com/glossary/input-and-output-tokens/): Input tokens are the tokens a language model reads in a request, including the system prompt, conversation history, tool definitions and images. Output tokens are the tokens it generates in reply, including any reasoning. LLM APIs usually price them separately, and output usually costs more. - [INT4](https://broadrouter.com/glossary/int4/): INT4 is a 4-bit integer number format that can represent 16 levels. In LLM inference it usually means 4-bit weight-only quantization (W4A16): weights are stored as INT4 with scale factors and expanded to 16-bit during computation, using about a quarter of the memory of 16-bit weights. - [IP allow list](https://broadrouter.com/glossary/ip-allow-list/): An IP allow list is a security control that accepts requests only from listed IP addresses or CIDR ranges and refuses all others. On an API, it limits the damage from a leaked key, because the key does not work from networks that are not on the list. - [Jailbreak (LLM)](https://broadrouter.com/glossary/jailbreak/): A jailbreak is an input crafted to make a language model ignore its safety training or the rules its deploying application set, so that it produces content or takes actions it would normally refuse. OWASP treats jailbreaking as a form of prompt injection. - [JSON mode](https://broadrouter.com/glossary/json-mode/): JSON mode is an LLM API setting, response_format {"type": "json_object"} in OpenAI-style APIs, that makes the model's reply valid, parseable JSON. Unlike structured outputs with a JSON Schema, it does not guarantee which keys, types or values the object contains. - [JSON Schema](https://broadrouter.com/glossary/json-schema/): JSON Schema is a vocabulary, written in JSON, for describing and validating the structure of JSON data: types, required fields, allowed values and nesting. The current version is 2020-12. LLM APIs use it to define tool parameters and the exact shape a structured output must follow. - [KV cache](https://broadrouter.com/glossary/kv-cache/): A KV cache is the memory where a transformer language model stores the attention keys and values computed for every token it has already processed, so each new token reuses them instead of recomputing the whole sequence. It grows with context length and often limits how many requests a GPU can serve. - [Latency](https://broadrouter.com/glossary/latency/): Latency, for a language model API, is the time a request takes from being sent to the final output token being received. It is the sum of time to first token and generation time, and is usually reported as percentiles such as p50 and p95 rather than as an average. - [LLM benchmark](https://broadrouter.com/glossary/llm-benchmark/): An LLM benchmark is a fixed, usually public set of tasks with a scoring method, used to measure and compare language models on one skill, such as knowledge questions, coding or speech recognition. A benchmark score shows relative ability on that test, not how a model will do on your own workload. - [LLM router](https://broadrouter.com/glossary/llm-router/): An LLM router is software or a service that sits between an application and one or more language model APIs and decides, for each request, which model or provider handles it, using rules such as cost, latency, availability or the type of task. - [LLM-as-a-judge](https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge is an evaluation method in which a language model grades other model outputs, either scoring a single answer against a rubric or picking the better of two answers. It approximates human judgement at much lower cost but has known biases, such as favouring the first or longer answer. - [Logprobs](https://broadrouter.com/glossary/logprobs/): Logprobs are the log probabilities a language model assigned to the tokens it generated. An API that returns them shows, for each output token, how likely the model judged it and, optionally, the most likely alternative tokens at that position. - [LoRA (low-rank adaptation)](https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation) is a parameter-efficient fine-tuning method that freezes a pre-trained model's weights and trains a pair of small low-rank matrices for selected layers. The learned update can be stored as a small adapter or merged into the weights, adding no inference latency. - [Lost in the middle](https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle is the observed tendency of language models to use relevant information best when it appears near the beginning or end of a long input and worst when it sits in the middle. It was named in a 2023 study by Liu et al. - [max_tokens](https://broadrouter.com/glossary/max-tokens/): max_tokens is an LLM API parameter that sets the most tokens a model may generate in one response. When generation reaches the limit, it stops and the response is marked as cut off, for example with finish_reason set to "length". - [MCP (Model Context Protocol)](https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol) is an open protocol, built on JSON-RPC 2.0, that standardises how LLM applications connect to external tools and data. An MCP server exposes tools, resources and prompts; an MCP client inside the host application discovers them and passes tool calls from the model to the server. - [Mixture of experts (MoE)](https://broadrouter.com/glossary/mixture-of-experts/): Mixture of experts (MoE) is a neural network design in which a layer holds many expert sub-networks and a learned router sends each token to only a few of them. An MoE language model has a large total parameter count but uses a small active subset for each token. - [Model deprecation](https://broadrouter.com/glossary/model-deprecation/): Model deprecation is when an API provider announces that a model will stop being offered, usually naming a retirement or shutdown date and a recommended replacement. After that date, requests to the model fail, so applications must migrate to another model ID beforehand. - [Model fallback](https://broadrouter.com/glossary/model-fallback/): Model fallback is a reliability pattern where a failed LLM request is retried on a different model or provider from an ordered list. The trigger can be an error, a rate limit, a context-length error or a content-policy refusal. It runs in your client or in a gateway in front of the APIs. - [Model license](https://broadrouter.com/glossary/model-license/): A model license is the legal terms under which a model's weights, code and outputs may be used, changed and shared. Open-weight models use standard software licences such as Apache 2.0 or MIT, or custom model licences that can add conditions, such as limits on large commercial services. - [Modern Standard Arabic (MSA)](https://broadrouter.com/glossary/msa/): Modern Standard Arabic (MSA) is the standardized form of Arabic used in writing, news, education and formal speech across the Arab world. It is nobody's native everyday language: people speak regional dialects such as Egyptian or Gulf Arabic and switch to MSA in formal settings. - [Multi-agent system](https://broadrouter.com/glossary/multi-agent-system/): A multi-agent system is an LLM application in which several agents, each with its own prompt, tools and context, work on parts of one task. A common design is orchestrator-worker: a lead agent plans, starts worker agents in parallel and combines their results. - [Multimodal AI](https://broadrouter.com/glossary/multimodal/): Multimodal AI is a model or system that works with more than one type of data, such as text, images, video and audio. A multimodal model might read a photo and a question together, or take speech in and answer in text. A vision-language model is the most common kind. - [Named entity recognition (NER)](https://broadrouter.com/glossary/named-entity-recognition/): Named entity recognition (NER) is the task of finding names in text and labelling each with a type, such as person, organization or location. Classic systems tag every token; LLMs are often asked instead to return a list of entities as JSON, which is then checked against the source. - [Ogg Opus](https://broadrouter.com/glossary/ogg-opus/): Ogg Opus is audio encoded with the Opus codec (RFC 6716) and stored in an Ogg container (RFC 7845). It is built for speech and interactive audio at low bitrates, and it is the format WhatsApp uses for voice notes, usually as .ogg or .opus files. - [Open-weight model](https://broadrouter.com/glossary/open-weight-model/): An open-weight model is a machine learning model whose trained weights are published for anyone to download and run, usually under a licence with conditions. It differs from fully open-source AI, which also releases the training code and detailed information about the training data, so others can study and reproduce the model. - [OpenAI-compatible API](https://broadrouter.com/glossary/openai-compatible-api/): An OpenAI-compatible API is an HTTP API that accepts the same request and response formats as the OpenAI API, such as POST /v1/chat/completions, so OpenAI SDKs and tools can use it by changing only the base URL, the API key and the model name. - [Optical character recognition (OCR)](https://broadrouter.com/glossary/ocr/): Optical character recognition (OCR) is technology that detects and converts printed or handwritten text in an image, scan or PDF into machine-readable characters. Classic OCR outputs text and its position on the page; vision-language models can read text and answer questions about it directly. - [Parallel tool calls](https://broadrouter.com/glossary/parallel-tool-calls/): Parallel tool calls are when a model returns more than one tool call in a single response, for example three file reads at once. In the OpenAI format they arrive as several entries in tool_calls, and parallel_tool_calls set to false limits the model to zero or one call per turn. - [Pay-as-you-go](https://broadrouter.com/glossary/pay-as-you-go/): Pay-as-you-go is a pricing model in which you pay only for what you use, metered per unit such as tokens, requests or audio minutes, with no subscription or minimum commitment. Most LLM APIs bill this way, from a prepaid balance or a periodic invoice. - [PDPL (Saudi and UAE data protection laws)](https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) is the common name for Saudi Arabia's personal data law, issued by Royal Decree M/19 and enforced by SDAIA, and for the UAE's Federal Decree-Law No. 45 of 2021. Both regulate processing of personal data and set conditions for transfers abroad, including prompts sent to foreign LLM APIs. - [PII (personally identifiable information)](https://broadrouter.com/glossary/pii/): PII (personally identifiable information) is information that can be used to distinguish or trace a person's identity, such as a name, ID number or biometric record, plus other information linked or linkable to that person. The GDPR's broader term is personal data. In LLM work, PII redaction removes it from prompts before they leave your systems. - [Prefill and decode](https://broadrouter.com/glossary/prefill-and-decode/): Prefill and decode are the two phases of LLM inference. In prefill the model processes the whole prompt in parallel and fills its KV cache; in decode it generates the reply one token at a time, reusing that cache. Prefill mostly sets time to first token; decode sets output speed. - [Prepaid credit](https://broadrouter.com/glossary/prepaid-credit/): Prepaid credit is money added to an API account before use. Each request deducts its cost from the balance, and when the balance runs out, metered requests stop until you top up. It caps spend at what you have paid in. - [Prompt engineering](https://broadrouter.com/glossary/prompt-engineering/): Prompt engineering is the process of writing, structuring and testing the instructions given to a language model so that it consistently produces output that meets your requirements. It covers role messages, clear instructions, examples, formatting, and measuring results against test cases. - [Prompt injection](https://broadrouter.com/glossary/prompt-injection/): Prompt injection is an attack in which text supplied to a language model, by a user or hidden in a web page, email, file or tool result, changes the model's behaviour against the developer's intent. OWASP ranks it first (LLM01:2025) in its Top 10 for LLM Applications. - [Quantization](https://broadrouter.com/glossary/quantization/): Quantization is the process of storing a neural network's numbers, mainly its weights, in a lower-precision format such as 8-bit or 4-bit instead of 16-bit, so the model uses less memory and often runs faster. Good methods keep the accuracy loss small. - [RAG (retrieval-augmented generation)](https://broadrouter.com/glossary/rag/): RAG (retrieval-augmented generation) is a pattern in which an application first retrieves passages relevant to a question from an external store, such as a search or vector index, and then adds them to the prompt so a language model can answer from that text rather than from memory alone. - [Rate limit](https://broadrouter.com/glossary/rate-limit/): A rate limit is a cap an API places on how much a client can use it in a time window, such as requests per minute or tokens per minute. Requests over the limit are refused, usually with HTTP 429 Too Many Requests, until usage falls back under it. - [ReAct (reason and act)](https://broadrouter.com/glossary/react-agent/): ReAct (reason and act) is a prompting pattern, introduced by Yao et al. in 2022, in which a language model interleaves reasoning traces with actions such as search or tool calls. It reads each observation, updates its plan, and repeats until it can give a final answer. - [Reasoning tokens](https://broadrouter.com/glossary/reasoning-tokens/): Reasoning tokens are tokens a reasoning model generates to work through a problem before it writes its visible answer. They are usually billed as output tokens, count toward the output limit and the context window, and may be hidden or summarized by the API. - [Request timeout](https://broadrouter.com/glossary/request-timeout/): A request timeout is the longest time an HTTP client waits for a server before it abandons the request. Clients usually split it into a connect timeout for opening the connection and a read timeout for waiting on data. LLM calls need generous read timeouts because long answers take time. - [Reranking](https://broadrouter.com/glossary/reranking/): Reranking is a second retrieval stage in which a more accurate but slower model, usually a cross-encoder, scores each query and candidate document pair from a first-pass search and reorders them, so only the most relevant few are passed to the language model. - [Right-to-left (RTL) text](https://broadrouter.com/glossary/rtl-text/): Right-to-left (RTL) text is writing, such as Arabic or Hebrew, displayed from right to left. Computers store it in logical reading order, and the Unicode Bidirectional Algorithm decides display order when RTL text mixes with left-to-right text, digits, model names or code. - [Sample rate](https://broadrouter.com/glossary/sample-rate/): Sample rate is the number of audio samples captured per second, measured in hertz (Hz). Speech recognition models commonly expect 16,000 Hz (16 kHz) mono audio. Together with bit depth and channel count, the sample rate sets how large an uncompressed audio file is per minute. - [Seed parameter](https://broadrouter.com/glossary/seed-parameter/): The seed parameter is an integer you send with a request so the API tries to sample the same tokens when the same request is repeated. OpenAI describes it as best effort: determinism is not guaranteed, and the system_fingerprint field shows when the backend configuration has changed. - [Server-sent events (SSE)](https://broadrouter.com/glossary/sse/): Server-sent events (SSE) is a web standard for a server to push a stream of text events to a client over one long-lived HTTP response with the content type text/event-stream. Each event is one or more field lines, such as data:, ended by a blank line. - [Serverless inference](https://broadrouter.com/glossary/serverless-inference/): Serverless inference is a way to run machine learning models in which the platform starts, scales and stops the compute behind an endpoint automatically, often down to zero when idle, and bills for usage rather than for reserved machines. The trade-off is possible cold-start delay. - [SLA (service level agreement)](https://broadrouter.com/glossary/sla/): An SLA (service level agreement) is a contractual commitment from a provider to a measurable level of service, usually a monthly uptime percentage, with a defined remedy, usually service credits, if it is missed. An SLA defines what counts as downtime, what is excluded and how to claim. - [Speaker diarization](https://broadrouter.com/glossary/diarization/): Speaker diarization is the task of dividing an audio recording into segments by speaker, answering "who spoke when". It labels each stretch of speech with an anonymous speaker ID such as SPEAKER_00, and is usually combined with a transcript to show which person said each line. - [Speculative decoding](https://broadrouter.com/glossary/speculative-decoding/): Speculative decoding is an inference technique in which a small, fast draft model proposes several next tokens and the large target model verifies them in a single forward pass, keeping the accepted ones. It speeds up generation without changing the distribution of the output. - [Speech-to-text](https://broadrouter.com/glossary/speech-to-text/): Speech-to-text, also called automatic speech recognition (ASR), is technology that converts spoken audio into written text. A speech-to-text API accepts an audio file or stream and returns a transcript, usually priced per minute of audio. Accuracy is measured with word error rate and varies with accent, dialect and recording quality. - [Spend limit](https://broadrouter.com/glossary/spend-limit/): A spend limit is a cap on how much an API account, project or key may spend in a period, usually a month. When usage reaches it, the provider rejects further requests, often with an HTTP 429 or 400 error, until the period resets or the limit is raised. - [SRT subtitles](https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) is a plain-text subtitle format. Each caption is a sequence number, a start and end time written as hh:mm:ss,mmm --> hh:mm:ss,mmm, one or more lines of text, and a blank line. WebVTT is the similar web format used by HTML video. - [Stop sequences](https://broadrouter.com/glossary/stop-sequences/): Stop sequences are strings you send with a request that make the model stop generating as soon as it produces one of them. OpenAI calls the field stop and accepts up to 4 sequences; Anthropic calls it stop_sequences. The matched text is normally left out of the returned output. - [Streaming](https://broadrouter.com/glossary/streaming/): Streaming is a way of returning a language model response in small chunks while it is still being generated, instead of as one reply at the end. OpenAI-compatible APIs stream over server-sent events when the request sets stream to true. - [Structured output](https://broadrouter.com/glossary/structured-output/): Structured output is an LLM API feature that constrains a model's reply to a format the developer specifies, usually a JSON Schema, so code can parse it reliably. It is stricter than JSON mode, which guarantees valid JSON but not a particular shape. - [Subagent](https://broadrouter.com/glossary/subagent/): A subagent is a separate agent that a main agent starts to handle one focused task. It runs with its own context window, system prompt and tool access, does the work, and returns only a result or summary to the main agent, which keeps the main conversation small. - [Subprocessor](https://broadrouter.com/glossary/subprocessor/): A subprocessor is a third party that a processor engages to carry out processing of personal data on behalf of the controller. For an LLM API, subprocessors can include GPU hosts, partner inference providers and backup capacity. Under GDPR Article 28 they need the controller's authorisation and the same data protection obligations. - [System prompt](https://broadrouter.com/glossary/system-prompt/): A system prompt is the instruction text an application sends to a language model ahead of the conversation, usually as a message with the system or developer role. It sets the model's role, rules, tone and output format, and the model is meant to follow it over conflicting user messages. - [Taa marbuta](https://broadrouter.com/glossary/taa-marbuta/): Taa marbuta (ة, U+0629) is an Arabic letter that appears only at the end of a word, often marking feminine nouns and adjectives. It looks like heh (ه, U+0647) with two dots, so typed text mixes the two, and many search normalizers map ة to ه. - [Temperature](https://broadrouter.com/glossary/temperature/): Temperature is a sampling parameter that controls how random a language model's output is. The model divides its token scores by the temperature before turning them into probabilities: low values make the likeliest token dominate, high values flatten the distribution so less likely tokens are chosen more often. - [Text-to-speech (TTS)](https://broadrouter.com/glossary/text-to-speech/): Text-to-speech (TTS), or speech synthesis, converts written text into spoken audio. A TTS API takes text, a voice and an output format such as MP3 or WAV, and returns generated speech. It is the reverse of speech-to-text, which turns audio into text. - [Throughput](https://broadrouter.com/glossary/throughput/): Throughput, in LLM serving, is the total amount of work a system completes per unit of time across all concurrent requests, usually measured as output tokens per second or requests per second. It differs from per-stream speed, which is how fast a single request receives its tokens. - [Time to first token (TTFT)](https://broadrouter.com/glossary/time-to-first-token/): Time to first token (TTFT) is the time from sending a request to a language model until the first output token arrives. It includes network time, queueing on the server and prefill, the step where the model reads the whole prompt, so it grows with prompt length and load. - [Timestamps (speech-to-text)](https://broadrouter.com/glossary/timestamps-asr/): Timestamps in speech-to-text are start and end times, in seconds, attached to parts of a transcript. Segment timestamps cover a phrase or sentence; word timestamps cover each word. They are needed for subtitles, search inside recordings and aligning text with speakers. - [Token](https://broadrouter.com/glossary/token/): A token is the unit of text a language model reads and writes: a word, part of a word, a punctuation mark or a byte sequence produced by the model's tokenizer. LLM APIs count, limit and usually bill usage in tokens. - [Tokenizer](https://broadrouter.com/glossary/tokenizer/): A tokenizer is the component that converts text into the sequence of integer token IDs a language model reads, and converts generated IDs back into text. Each model ships its own tokenizer, so the same text can produce a different number of tokens on different models. - [Tokens per second](https://broadrouter.com/glossary/tokens-per-second/): Tokens per second (TPS) is the rate at which a language model generates output tokens, usually measured for one request after its first token arrives. It sets how fast a reply streams once it starts. Summed across all concurrent requests it becomes system throughput; the wait before the first token is measured separately. - [Tool calling](https://broadrouter.com/glossary/tool-calling/): Tool calling, also called function calling, is an LLM API feature in which the model returns a structured request to run a function the developer has defined, with JSON arguments. The application runs the function and sends the result back; the model never runs code itself. - [tool_choice](https://broadrouter.com/glossary/tool-choice/): tool_choice is a request parameter that controls tool use. In OpenAI's Chat Completions it can be auto (the model decides), required (it must call one or more tools), none (no tools), or a named function to force. Anthropic's equivalent options are auto, any, tool and none. - [Top-p (nucleus sampling)](https://broadrouter.com/glossary/top-p/): Top-p, also called nucleus sampling, is a decoding method in which a language model samples the next token only from the smallest set of most likely tokens whose combined probability reaches a threshold p, such as 0.9, and discards the unlikely tail. - [Unicode normalization](https://broadrouter.com/glossary/unicode-normalization/): Unicode normalization converts text to a standard form so that strings that look the same use the same code points. The four forms are NFD, NFC, NFKD and NFKC; the K forms also fold compatibility characters, such as Arabic presentation forms and ligatures, into ordinary letters. - [Vector database](https://broadrouter.com/glossary/vector-database/): A vector database is a data store that holds embedding vectors, often with their source text and metadata, and answers nearest-neighbour queries: given a query vector, it returns the stored vectors most similar to it, usually by cosine distance, Euclidean distance or inner product. - [Vision-language model](https://broadrouter.com/glossary/vision-language-model/): A vision-language model (VLM) is a model that takes images and text together as input and produces text, so it can describe pictures, answer questions about them and read documents, charts or screenshots. VLMs are also called image-text-to-text or multimodal models. - [vLLM](https://broadrouter.com/glossary/vllm/): vLLM is an open-source library for LLM inference and serving, first developed at UC Berkeley's Sky Computing Lab. It manages KV cache memory with PagedAttention, batches incoming requests continuously, supports many quantization formats and exposes an OpenAI-compatible HTTP server. - [Vocabulary size](https://broadrouter.com/glossary/vocabulary-size/): Vocabulary size is the number of distinct tokens in a language model's tokenizer, each with its own ID and embedding. A larger vocabulary can encode common text in fewer tokens, but it makes the embedding and output layers bigger, and it only helps for text the vocabulary was trained on. - [Voice activity detection (VAD)](https://broadrouter.com/glossary/vad/): Voice activity detection (VAD) is a signal-processing step that classifies short frames of audio as speech or non-speech. Speech systems use it to skip silence, split long recordings at pauses, detect when a speaker has finished, and avoid transcribing background noise. - [WAV](https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format, or WAVE) is an audio container built on Microsoft and IBM's RIFF format. It usually holds uncompressed linear PCM samples, so its size is predictable: sample rate times bytes per sample times channels, per second. - [Word error rate](https://broadrouter.com/glossary/word-error-rate/): Word error rate (WER) is the standard accuracy measure for speech recognition: the number of word substitutions, deletions and insertions needed to turn a transcript into the correct reference, divided by the number of words in the reference. Lower is better. - [Zero data retention](https://broadrouter.com/glossary/zero-data-retention/): Zero data retention (ZDR) is a data-handling policy in which an AI provider processes prompts and outputs only to serve a request and does not store that content afterward, in logs, caches or training data. Content-free records such as token counts and timings are usually still kept. ## Blog - [Pricing articles](https://broadrouter.com/blog/category/pricing/) - [Engineering articles](https://broadrouter.com/blog/category/engineering/) - [Arabic AI articles](https://broadrouter.com/blog/category/arabic-ai/) - [Comparisons articles](https://broadrouter.com/blog/category/comparisons/) - [Agents articles](https://broadrouter.com/blog/category/agents/) - [Privacy articles](https://broadrouter.com/blog/category/privacy/) - [LLM API price changes in October 2026: 3 rises, no cuts](https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 pricing-page changes at 5 of 7 LLM API providers in October 2026: 3 price rises, no cuts, 9 new models priced and 11 removed. Archive links for each. - [A private endpoint starts with a clear workload](https://broadrouter.com/blog/private-endpoint-planning/): What to define before asking for a private LLM endpoint: workload, tokens per second, concurrency, context, region, licence and budget, then how to test it. - [Add diacritics to Arabic text with an LLM: prompt, checks, DER](https://broadrouter.com/blog/arabic-diacritization-llm/): Restore Arabic diacritics (tashkeel) with an LLM API: a prompt that keeps every letter, a check that rejects edits, DER and WER scoring, and when to use CATT. - [Arabic grammar and spelling correction with an LLM: a how-to](https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Correct Arabic spelling and grammar with an LLM API: a minimal-edit prompt, a guard against rewrites, word-level edits by error class and M2-style scoring. - [Arabic meeting minutes from recordings: an API pipeline](https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Turn Arabic meeting recordings into minutes with decisions and action items: transcribe, clean, summarise with an LLM into JSON, and check names and dates. - [Arabic named entity extraction with an LLM: JSON, dates and checks](https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Extract Arabic names, organisations, places and Hijri dates to JSON with an LLM: a prompt, span checks that ignore diacritics, date conversion and a scoring script. - [Arabic OCR API comparison: Google, Azure, Textract, Mistral, LLMs](https://broadrouter.com/blog/arabic-ocr-api-comparison/): Which OCR APIs officially support Arabic, printed and handwritten, what each returns and what it costs per 1,000 pages, compared with vision LLM APIs. - [Arabic subtitles (SRT) from a speech-to-text API: what you need](https://broadrouter.com/blog/arabic-subtitles-srt-api/): What an SRT file needs, which speech-to-text APIs return timestamps or SRT for Arabic, and a tested pipeline that builds subtitles from a text-only API. - [Arabic text classification with an LLM: sentiment, intent, triage](https://broadrouter.com/blog/arabic-text-classification-llm/): Classify Arabic reviews, messages and support tickets with an LLM API: label design, a dialect-aware prompt, JSON output, macro F1 testing and cost at volume. - [Arabic text normalization for search and RAG: what to change, what not](https://broadrouter.com/blog/arabic-text-normalization-search/): Tashkeel, alef forms, ta marbuta, hamza seats, Persian letters, digits and Unicode: which Arabic normalization rules help search, which hurt, and tested Python code. - [Arabic text summarization with an LLM: prompts, length and scoring](https://broadrouter.com/blog/arabic-text-summarization-llm/): Summarise Arabic articles, reports and dialect messages with an LLM API: prompt patterns, word-count control, number checks, and why default ROUGE gives Arabic 0. - [Artificial Analysis alternatives: where to compare LLMs in 2026](https://broadrouter.com/blog/artificial-analysis-alternatives/): Artificial Analysis alternatives compared: Arena, Epoch AI, Vals AI, OpenRouter rankings, Hugging Face and price data, by what each one measures and when to use it. - [Before fine-tuning: define the problem to solve](https://broadrouter.com/blog/fine-tuning-readiness/): When prompting, RAG or evaluation fixes the problem and fine-tuning is not needed, what training data you need, and how to test a fine-tune before you ask for one. - [BroadNet's GPU hardware and serving approach, as BroadNet describes it](https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): What BroadNet says about the hardware behind BroadNet Inference: three GPU tiers from B300 to client appliances, throughput per GPU, and what its tests taught it. - [Cache LLM API responses to cut cost: an exact-match cache in Python](https://broadrouter.com/blog/cache-llm-responses-cut-cost/): A tested exact-match cache for LLM API responses in Python: what goes in the key, what never to cache, TTLs, privacy and how to measure the hit rate before building. - [Claude Code with open models: proxies, native endpoints and limits](https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code speaks the Anthropic Messages API. How to run it on open models through a LiteLLM proxy or a native endpoint, and what breaks along the way. - [Design an AI experience for multiple languages](https://broadrouter.com/blog/multilingual-product-design/): Design an AI product for Arabic, English and Spanish users: UI locale vs model language, RTL and mixed text, numbers, dates, detection and cost. - [Do LLM API prices only go down? 52 rises and how to budget for them](https://broadrouter.com/blog/llm-api-price-increases/): LLM API prices do go up. We found 52 list-price rises on the same model in archived pricing pages, 2024 to 2026, median 40%. All sources, plus a budget rule. - [Egypt's data protection law (Law 151 of 2020) and LLM APIs](https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Sending Egyptians' personal data to an LLM or speech-to-text API abroad: what Law 151 of 2020 and Decision 816 of 2025 say about transfers, licences and breaches. - [How often do LLM API prices change? 3 years of archived pricing pages](https://broadrouter.com/blog/how-often-llm-api-prices-change/): We read 308 archived pricing pages from 11 LLM API providers, 2023 to 2026. 80% of models never changed price; DeepInfra made 75% of all list-price changes. - [How to fix invalid JSON from an LLM: causes, repair, retry](https://broadrouter.com/blog/fix-invalid-json-from-llm/): Why LLMs return broken JSON (code fences, prose, max_tokens cut-offs, trailing commas) and a tested Python function that repairs what is safe and rejects the rest. - [How to reduce LLM API costs: a 10-step checklist, in order](https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): Cut an LLM API bill in the right order: measure, cache the prefix, trim prompts, cap output and reasoning, right-size the model, batch, and know when flat rate wins. - [How to test an LLM for sensitive-data safety: BroadNet's method](https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): BroadNet's published method for testing whether an LLM leaks credentials, PII, health or financial data before you deploy it: categories, leak vs miss, 3+ runs. - [Live, on request, research: read a model catalogue clearly](https://broadrouter.com/blog/understand-model-status/): How to tell which catalogue models you can call today, which need an enquiry and which were only tested, using BroadNet's catalogue and /v1/models as the example. - [LLM API failover: fall back to a second provider in your own code](https://broadrouter.com/blog/llm-api-failover-fallback/): How to fail over between OpenAI-compatible LLM providers: which errors to fail over on, a circuit breaker, model ID mapping, streaming, cost and privacy parity. - [LLM API price cuts 2023–2026: the biggest drops, with sources](https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): The largest LLM API price cuts from 2023 to 2026, each with before and after Internet Archive links: o3, Codestral, Gemini 1.5, plus the rises few people noticed. - [LLM API rate limits by provider: RPM, TPM and concurrency compared](https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): Entry-tier LLM API rate limits from 13 providers, read from their docs: OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks and others, plus what RPM and TPM mean. - [LLM API SLAs compared: what Azure, Bedrock, Google and OpenAI promise](https://broadrouter.com/blog/llm-api-sla-uptime-compared/): What each LLM API provider commits to in writing: uptime percentage, what counts as downtime, service credits and latency targets, read from the SLA pages. - [LLM batch processing without a Batch API: a resumable runner](https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Run 10,000 to 1,000,000 prompts through a chat-completions API with no Batch API: concurrency caps, rate pacing, retries, resumable checkpoints and the time math. - [Make human review part of your document AI workflow](https://broadrouter.com/blog/document-review-queues/): Design the human review step of a document-AI pipeline: validation rules, what to send to a person, random audit samples, an audit trail and the review rate. - [Migrating off the OpenAI API: a code walkthrough in Python, Node, cURL](https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): The code changes to move from OpenAI to an OpenAI-compatible API: client setup, Responses to Chat Completions, tools, streaming, embeddings and a smoke test. - [Models BroadNet evaluated and ruled out, and why](https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet publishes the models it tested and rejected, with reasons: chat models that failed its credential-safety gate, speech, OCR and embedding models that lost. - [Open-weight vs closed model API cost: what the price data shows](https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): Are open-weight models cheaper through an API than GPT, Claude or Gemini? Per-token prices from providers' own pages, host spreads and a worked cost per user. - [OpenAI vs Anthropic vs Google API price history, 2023–2026](https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): GPT, Claude and Gemini API prices side by side from archived pricing pages, 2023 to 2026: which lines got cheaper, which got dearer, each with an archive link. - [OpenAI-compatible API base URLs: 14 providers, checked](https://broadrouter.com/blog/openai-compatible-api-base-urls/): The base_url for 14 OpenAI-compatible LLM APIs, each read from the provider's own docs, with the endpoints they list and the quirks that break a drop-in switch. - [OpenClaw and Hermes Agent on a cheap model: setup and token math](https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): Connect OpenClaw or Hermes Agent to an OpenAI-compatible endpoint, see why scheduled checks burn tokens, and cut the bill with lighter context and cheaper models. - [Prompt injection in coding agents: defences that actually hold](https://broadrouter.com/blog/prompt-injection-coding-agents/): Where prompt injection reaches a coding agent, and the layered defences vendors and researchers document: sandboxing, egress limits, allow-lists and approval. - [RAG chunk size and overlap: what they do to retrieval and cost](https://broadrouter.com/blog/rag-chunk-size-overlap/): How chunk size and overlap change RAG recall, index size, embedding cost and the tokens every query sends to the LLM, with published results and worked numbers. - [Record audio in the browser and send it to a speech-to-text API](https://broadrouter.com/blog/record-browser-audio-transcription-api/): Record the microphone with MediaRecorder, turn WebM or Safari MP4 into 16 kHz WAV, and send it to a transcription API through a small server that keeps the key. - [Reduce prompt tokens: measured cuts for system prompts, tools, data](https://broadrouter.com/blog/reduce-prompt-tokens/): Measured token savings from trimming system prompts, few-shot examples, tool schemas, data formats, chat history and Arabic diacritics, counted with real tokenizers. - [Same model, different host: how much LLM API prices vary](https://broadrouter.com/blog/same-llm-different-host-prices/): For 119 models sold by several hosts, the dearest list price is a median 1.76 times the cheapest. Data from 41 providers' own pricing pages, November 2026. - [Saudi and UAE PDPL and LLM APIs: a cross-border checklist](https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Sending personal data to an LLM API from Saudi Arabia or the UAE: what the PDPL transfer rules say, what to ask a provider, and when in-country inference matters. - [Setting max_tokens for reasoning models: a budgeting method](https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Why reasoning models return empty replies at low max_tokens, and how to size the budget from measured percentiles, the context window and a retry that escalates. - [Text, speech or vision? Start with the task](https://broadrouter.com/blog/text-speech-vision/): Pick text chat, image input or audio transcription by task, with the models and limits BroadNet documents, and what it does not offer, such as TTS. - [The Safety Map: how 39 LLMs handle sensitive data](https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet surveyed 39 LLMs from 14 labs on credentials, PII, health and financial data: what transferred, what did not, and the tool-call and reasoning leaks. - [Transcribe long audio over 25 MB: split, send and stitch](https://broadrouter.com/blog/transcribe-long-audio-files-api/): How to transcribe recordings over a speech-to-text API file limit: shrink the file, split at silences with ffmpeg, send chunks in parallel and stitch the text. - [Translate app strings to Arabic with an LLM: placeholders and plurals](https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): Translate i18n JSON files to Arabic with an LLM API: freeze placeholders, generate all six Arabic plural forms, validate ICU syntax in CI and isolate inserted names. - [Translate Arabic audio to English text with an API: a tested pipeline](https://broadrouter.com/blog/translate-arabic-audio-to-english/): Turn Arabic recordings into English text with two API calls: transcribe with language=ar, then translate with an LLM, with number checks, chunking, cost and scoring. - [Uno in vLLM: what BroadNet's serving study measured](https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet built the Uno self-drafting adapter into vLLM and measured it on an H100: 2.58x plain throughput at one stream, 1.74x at 32. The method, DOI and limits. - [Uno on Gemma 4 26B: BroadNet's mixture-of-experts study](https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet trained the first Uno self-drafting adapter for a mixture-of-experts model, Gemma 4 26B A4B. What it measured, when self-drafting pays, and the DOI. - [What one AI coding agent run costs: steps, context growth, caching](https://broadrouter.com/blog/ai-coding-agent-run-cost/): Why a coding agent run costs more per step the longer it goes, how caching changes the total, and when a flat session beats per-token billing. Worked token math. - [10 ways to cut coding agent LLM costs, with estimated savings](https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Ten concrete ways to lower what Cline, Aider, opencode or your own coding agent spends on tokens, each with the math behind the saving and how to measure it. - [A weekly LLM plan for short projects: Orbit vs hour vs month](https://broadrouter.com/blog/weekly-llm-plan-short-projects/): When a one-week unlimited DeepSeek V4.1 Flash plan beats an hourly session, a monthly plan or pay-per-token, with the math for hackathons, sprints and migrations. - [Agent token budget: where an agent loop's tokens go](https://broadrouter.com/blog/agent-token-budget/): System prompt, tool schemas, file reads, history and reasoning: how an agent loop spends tokens, how to measure each part from usage data, and what to trim first. - [Arabic ASR word error rate (WER) explained](https://broadrouter.com/blog/arabic-asr-wer-explained/): What word error rate means, how to compute it, and why Arabic WER is so sensitive to diacritics, spelling normalisation and dialects. With Python and real numbers. - [Arabic dialect speech recognition: Gulf, Egyptian, Levantine, Maghrebi](https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Why speech-to-text struggles with Gulf, Egyptian, Levantine and Maghrebi Arabic, and how to build a dialect test set that shows how a provider will do on your audio. - [Arabic OCR vs vision LLMs: reading Arabic documents and invoices](https://broadrouter.com/blog/arabic-ocr-vision-llm/): When to use classic OCR and when a vision LLM for Arabic invoices, receipts and forms: accuracy traps, JSON output, cost per page and a validation pipeline. - [Arabic prompt engineering: register, RTL, numerals and mixed text](https://broadrouter.com/blog/arabic-prompt-engineering/): Write prompts that get good Arabic from an LLM: pick register and dialect, control numerals and dates, keep RTL formatting clean and handle Arabic–English mixing. - [Arabic RAG: normalisation, chunking and retrieval that works](https://broadrouter.com/blog/arabic-rag-search/): Retrieval-augmented generation over Arabic documents: normalise alef, ya and ta marbuta, chunk sensibly, pick multilingual embeddings, add keyword search. - [Arabic speech-to-text API: how to choose one in 2026](https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): How to pick an Arabic transcription API: word error rate, dialects, file limits, price per audio hour and privacy, plus a test plan for your own recordings. - [Arabic WhatsApp chatbot architecture: voice notes to replies](https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): A working architecture for an Arabic WhatsApp assistant: webhook, queue, voice-note transcription, LLM reply, policy rules and costs, with Python you can adapt. - [Arabic–English translation with LLMs: prompts, glossaries, evaluation](https://broadrouter.com/blog/arabic-english-translation-llm/): Translate Arabic to English and back with an LLM API: prompts that keep tone, glossaries for fixed terms, numbers and names, and how to evaluate quality. - [Arabizi and LLMs: handling Franco-Arabic with numbers like 3 and 7](https://broadrouter.com/blog/arabizi-llm-handling/): Arabizi writes Arabic in Latin letters with digits like 3 for ع and 7 for ح. How to detect it, convert it to Arabic script with an LLM, and reply the right way. - [Best LLM for Aider: how to choose a model and connect it](https://broadrouter.com/blog/best-llm-for-aider/): How to pick the best LLM for Aider: edit-format discipline, context size and cost per hour, a one-hour test on your own repo, and the exact OPENAI_API_BASE setup. - [Best open model for coding agents: pick by tools, context and cost](https://broadrouter.com/blog/best-open-model-for-coding-agents/): GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B or Gemma 4 for your coding agent? Pick by tool calling, context size, reasoning and cost per task, with the math. - [Cached input tokens: how to cut an LLM API bill](https://broadrouter.com/blog/cached-input-tokens-explained/): Cached input costs a fraction of normal input. How prompt caching is billed, how to read cached_tokens in usage, and how to structure prompts to hit it. - [Cheapest vision LLM API: cost per image and per page, worked out](https://broadrouter.com/blog/cheapest-vision-llm-api/): What a vision LLM call costs per image and per document page: token math for Gemma 4, Qwen3.8 27B and DeepSeek V4.1 Flash, plus where DeepInfra is cheaper. - [Choosing an LLM API for Arabic: what to test](https://broadrouter.com/blog/choosing-an-llm-for-arabic/): How to pick an LLM for Arabic apps: dialect understanding, writing quality, translation, Arabic documents and images, cost, plus a test set you can build in a day. - [Cline vs Roo Code vs Aider for teams on open models (2026)](https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (now archived, continued as Zoo Code) and Aider compared on what matters with open models: tool calls, context use, MCP, git and approvals. - [Context management for long-running agents with a 1M window](https://broadrouter.com/blog/long-running-agent-context-management/): How to keep long agent sessions accurate and affordable with a 1M-token window. Append-only history, cache-friendly prompts, when to compact, and what to drop first. - [DeepSeek API alternatives: 5 ways to call DeepSeek models](https://broadrouter.com/blog/deepseek-api-alternatives/): Calling DeepSeek models through the official API, an aggregator, a cloud catalogue, your own GPUs or BroadNet Inference: trade-offs on price, privacy and limits. - [DeepSeek V4.1 Flash API pricing: per token, cached and plans](https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): DeepSeek V4.1 Flash API pricing per 1M input, cached and output tokens, flat monthly, weekly and hourly plans, and how DeepSeek, DeepInfra and Fireworks compare. - [Extract invoice, receipt and form data to JSON with a vision LLM](https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): A production pipeline for invoices, receipts and forms: image prep, a JSON schema prompt, Pydantic validation, arithmetic checks, retries and a human review queue. - [Flat-rate vs pay-per-token LLM API: the break-even math](https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): When does a flat-rate LLM plan beat paying per token? Four worked monthly bills for DeepSeek V4.1 Flash, the break-even volume, and how to measure your own. - [Free LLM API credit: how far a starting credit really goes](https://broadrouter.com/blog/free-llm-api-credit/): What a free LLM API credit really buys: tokens, audio minutes and hourly keys on the BroadNet starting credit, compared with Gemini and OpenRouter free tiers. - [GDPR and LLM APIs: 15 questions to ask your provider](https://broadrouter.com/blog/gdpr-llm-api-checklist/): A GDPR checklist for LLM APIs: DPA terms, subprocessors, retention, training, international transfers and breach notice, with the GDPR articles behind each question. - [GLM 5.3 vs DeepSeek V4.1 Flash: which one to use, and when](https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 vs DeepSeek V4.1 Flash on price, context, tools, structured output, images and reasoning, with worked workload costs and a simple rule for when to use each. - [Gulf, Egyptian and Levantine Arabic: how LLMs handle dialects](https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): How LLMs cope with Gulf, Egyptian and Levantine Arabic: where they fail, how to prompt for dialect, how to test it, and what BroadNet reports for Qwen3.8 27B. - [Hourly LLM API key: a flat price per hour, explained](https://broadrouter.com/blog/hourly-llm-api-key-explained/): BroadNet's hourly key buys one hour of Gemma 4 VL 26B on one stream for a fixed price. What it includes, its limits, who it suits and when it pays off. - [How inference providers price open models: 6 compared](https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): How OpenRouter, Together AI, Fireworks, Groq, DeepInfra and BroadNet price open models: per token, cached, batch, GPU hours and flat plans, with cited prices. - [How to choose an LLM API provider: a 20-point checklist](https://broadrouter.com/blog/choose-llm-api-provider-checklist/): Twenty questions to ask before you pick an LLM API: pricing, limits, OpenAI compatibility, privacy, models, support and uptime, plus a scoring sheet and test plan. - [How to count tokens before you send an LLM request](https://broadrouter.com/blog/count-tokens-before-sending/): Estimate prompt tokens by language and model before calling an LLM API: why tokenizers differ, measured examples in five languages, and safe max_tokens margins. - [How to estimate LLM API cost before you launch a feature](https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): A step-by-step cost model for a new AI feature: users times actions times calls times tokens, with cache, reasoning, retries, rate limits and a worked monthly table. - [How to evaluate an LLM before switching providers](https://broadrouter.com/blog/evaluate-llm-before-switching/): Build a small eval set from real traffic, score it with code checks and a debiased LLM judge, and compare quality, cost and latency before you migrate. - [How to handle LLM API 429 errors: backoff, jitter and limiters](https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 vs 5xx, exponential backoff with full jitter, concurrency limiters and queues for LLM APIs, with Python and Node code you can copy into production. - [How to run coding agents unattended overnight, safely](https://broadrouter.com/blog/overnight-coding-agent-runs/): A practical setup for unattended coding agent runs: sandboxing, git checkpoints, task files, step limits, morning review, and flat plan sessions overnight. - [How to run parallel coding agents on a flat-rate LLM API](https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Sub-agents, orchestrators and batch workers on unlimited DeepSeek V4.1 Flash. How many plan sessions you need, how to split them across keys, and how to cap overlap. - [How to structure prompts so the prefix caches](https://broadrouter.com/blog/prompt-structure-for-caching/): How to lay out LLM prompts so the cached prefix stays identical: system prompt, tool schemas, documents, append-only history, plus a test that proves it. - [How to transcribe call center recordings: an Arabic-ready pipeline](https://broadrouter.com/blog/transcribe-call-center-recordings/): Split stereo calls by channel, cut speech turns, transcribe Arabic and English, redact PII and summarise each call with an LLM. A working Python pipeline with costs. - [JSON mode vs json_schema: structured output that holds up](https://broadrouter.com/blog/structured-output-json-schema/): JSON mode gives valid JSON; json_schema gives your shape. How to request each on OpenAI-compatible models, validate with Pydantic, and repair bad output in a loop. - [LLM API data retention compared: OpenAI, Anthropic, Gemini, Mistral](https://broadrouter.com/blog/llm-api-data-retention-compared/): What OpenAI, Anthropic, Gemini API, Mistral and BroadNet officially say about storing API prompts and training on them, quoted from their pages in November 2026. - [LLM API key security: storage, rotation and IP allow-lists](https://broadrouter.com/blog/llm-api-key-security/): Keep LLM API keys out of code and logs, give each workload its own key, rotate on a schedule, lock keys to server IPs and revoke a leak in seconds. - [LLM API pricing explained: input, cached, output and reasoning](https://broadrouter.com/blog/llm-api-pricing-explained/): How per-token LLM API pricing works: input, cached input, output, reasoning tokens, images and per-minute audio, with worked examples from a published rate card. - [LLM API spend control: budgets, key limits and loop protection](https://broadrouter.com/blog/llm-api-spend-control/): Stop an LLM API bill from running away: prepaid balances, one key per app, IP allow-lists, usage logging, in-code budgets, loop caps and flat-price keys. - [LLM API usage logging: cost per user and feature, PII redacted](https://broadrouter.com/blog/llm-api-usage-logging/): Log every LLM API call with tokens, cost, latency and request ID, attribute spend to users and features, and redact PII before it is written. Python code. - [LLM cost per user: unit economics for a SaaS AI feature](https://broadrouter.com/blog/llm-cost-per-user-saas/): Work out what an AI feature costs per active user per month, why heavy users matter more than averages, and how to price SaaS tiers so the model bill stays covered. - [LLM latency: TTFT vs tokens per second vs end-to-end time](https://broadrouter.com/blog/llm-latency-ttft-throughput/): What time to first token, tokens per second and end-to-end latency measure, how to time each from the client with code, and which fixes move which number. - [LLM streaming explained: SSE chunks, deltas, usage and cancel](https://broadrouter.com/blog/llm-streaming-sse-explained/): How chat-completion streaming works on the wire, how to rebuild the answer from deltas, read usage from the last chunk and cancel cleanly in Python, Node and curl. - [Long context vs RAG: when 1M tokens replaces a vector DB](https://broadrouter.com/blog/long-context-vs-rag/): When a 1M-token context window can replace a vector database and when RAG still wins, compared on cost per query, latency, recall and upkeep. - [Long-running agents: timeouts, retries and checkpoints that hold](https://broadrouter.com/blog/long-running-agents-timeouts-retries/): Keep an LLM agent running for hours: timeouts per call, streaming to spot stalls, safe retries, idempotent tool side effects and a checkpoint after every step. - [Multilingual speech-to-text API: 14 languages, one endpoint](https://broadrouter.com/blog/multilingual-speech-to-text-api/): Transcribe English, Spanish, French, Chinese, Arabic and 9 more languages through one OpenAI-compatible endpoint priced per minute, and when to use the Arabic model. - [n8n AI Agent on a cheap LLM: setup, model choice and cost per run](https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): Build n8n AI Agent workflows on a low-cost OpenAI-compatible model: which model to pick for tools, the settings that cut tokens, and the cost math per 10,000 runs. - [Open-weight vs closed AI models for business: cost, control, risk](https://broadrouter.com/blog/open-weight-vs-closed-models/): Open-weight or closed LLMs for your business? Compare cost, control, privacy, the quality gap and lock-in, with the three ways to run open models and when each wins. - [OpenRouter vs a flat-rate plan for AI agents: which costs less?](https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter bills per token plus a fee on credits; a flat plan bills per session. The cost math for always-on agents, OpenRouter's fees and limits, and which fits. - [Reasoning tokens: why you pay for more output than you see](https://broadrouter.com/blog/reasoning-tokens-cost/): Reasoning models bill hidden thinking as output tokens. How to measure it, what it costs per model, and how to cut it with reasoning_effort, max_tokens and routing. - [Self-hosting an LLM vs an API: the real cost and break-even](https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): GPU rental prices from official pages, the formula for cost per 1M tokens on your own GPU, the utilization you need to beat an API, and the ops costs people forget. - [Sending sensitive data to an LLM API: redact, minimise, log less](https://broadrouter.com/blog/llm-api-sensitive-data/): How to send customer, health or financial data to an LLM API safely: classify fields, redact with reversible placeholders, use zero retention, keep logs clean. - [Speech-to-text API pricing compared: price per minute in 2026](https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): Per-minute and per-1,000-hour prices for OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq and BroadNet Transcribe from official pages, plus add-ons. - [Switching to an OpenAI-compatible API: a migration checklist](https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): Changing base_url is the easy part. Model IDs, reasoning tokens, streaming usage, errors, rate limits and keys are what break. A checklist with code. - [Tokens per word in Arabic vs English: why Arabic costs more](https://broadrouter.com/blog/tokens-per-word-arabic-english/): Why Arabic often uses more LLM tokens than English for the same meaning, why tokens per word misleads, and a script to measure your own ratio from the usage field. - [Tool calling with open models: schemas, parallel calls, retries](https://broadrouter.com/blog/tool-calling-open-models/): Function calling on OpenAI-compatible open models: schema tips, parallel tool calls, the five ways tool calls fail, and a Python loop that validates and retries. - [Use MCP servers with OpenAI-compatible models (any provider)](https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP works with any LLM: the client turns MCP tools into function calls. How to use MCP with open models in Cline, Continue, opencode, n8n or your own Python code. - [Use open models in coding agents: Aider, Cline, opencode, Codex](https://broadrouter.com/blog/use-open-models-in-coding-agents/): Point Aider, Cline, Roo Code, opencode, Continue and Codex CLI at any OpenAI-compatible endpoint. The exact field names for each, plus the settings that break. - [Voice notes to text for business: WhatsApp and Telegram to tickets](https://broadrouter.com/blog/voice-notes-to-text-for-business/): Turn customer voice notes from WhatsApp Business and Telegram bots into transcripts and support tickets: download, transcribe, extract fields with an LLM, route. - [What "unlimited" really means in an LLM API plan](https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Token meters, rolling 5-hour windows, weekly caps, rate limits and concurrency. Seven questions that show whether an "unlimited" AI plan is unlimited for your agent. - [What multi-agent setups really cost: per token vs flat sessions](https://broadrouter.com/blog/multi-agent-cost/): Multi-agent systems burn several times the tokens of one agent. The session math for flat plans vs per-token billing, the hidden multipliers, and when each wins. - [When a 1M-token context window is worth it (and when not)](https://broadrouter.com/blog/million-token-context-window-guide/): What fits in 1,048,576 tokens, what it costs on DeepSeek V4.1 Flash, how reasoning and caching change the math, and when RAG is still the better choice. - [Whisper API alternative: when to switch, and how to test one](https://broadrouter.com/blog/whisper-api-alternative/): When a hosted alternative to the OpenAI Whisper API makes sense: word error rate on your audio, languages, price per hour, limits and privacy, with test code. - [Zero data retention LLM API: a 10-question checklist](https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): What "zero data retention" should mean for an LLM API: prompts, outputs, logs, training, subprocessors and audits. Ten questions to ask before sending data. ## Localized pages: 17 languages, 238 pages The launch pages also exist in Arabic, Spanish, German, Portuguese, French, Italian, Chinese (Simplified), Turkish, Urdu, Hindi, Indonesian, Filipino, Malay, Japanese, Korean, Russian and Vietnamese, with the same facts and prices as the English page each one translates. Some blog posts, guides and glossary entries are translated too (see "Translated articles, guides and glossary"); everything else (tools, data, compare, alternatives, LLM API pricing) is English-only. Model IDs, code and URLs stay in Latin script. ### Arabic (ar, العربية): 14 pages, hreflang ar, right-to-left - [BroadRouter: API لـ DeepSeek وGLM وQwen وGemma والصوت العربي](https://broadrouter.com/ar/) (Arabic; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/ar/index.md): واجهة API واحدة متوافقة مع OpenAI لنماذج DeepSeek وGLM وQwen وGemma وتحويل الصوت العربي إلى نص. ادفع حسب الاستخدام، أو اشترِ جلسات DeepSeek بسعر ثابت لوكلائك. - [واجهة API متوافقة مع OpenAI: بدء سريع بـ Python وcURL](https://broadrouter.com/ar/developers/) (Arabic; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/ar/developers/index.md): ابدأ مع واجهة API متوافقة مع OpenAI: أرسل أول طلب محادثة وأول طلب تفريغ صوتي إلى BroadNet Inference باستخدام Python أو cURL، ثم راجع قائمة التحقق للانتقال. - [تواصل مع BroadNet بشأن BroadNet Inference](https://broadrouter.com/ar/contact/) (Arabic; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/ar/contact/index.md): كيف تتواصل مع BroadNet بشأن BroadNet Inference: البريد العام والدعم الفني والمبيعات، وطريقة طلب نقاط نهاية خاصة أو معالجة إقليمية أو نماذج غير مدرجة. - [API لتحويل الصوت العربي إلى نص: النماذج وسعر الدقيقة](https://broadrouter.com/ar/arabic-transcription/) (Arabic; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/ar/arabic-transcription/index.md): حوّل الصوت العربي إلى نص (تفريغ صوتي) عبر BroadNet Inference: النموذج العربي مقابل متعدد اللغات، والصيغ وحد الحجم، وحاسبة التكلفة، وشيفرة جاهزة للنسخ. - [خصوصية BroadNet Inference: ما البيانات التي يُحتفظ بها؟](https://broadrouter.com/ar/privacy/) (Arabic; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/ar/privacy/index.md): ملخص قصير: لا تحتفظ BroadNet Inference بأوامرك وملفاتك الصوتية وإجاباتك ولا تستخدمها لتدريب النماذج. اعرف البيانات الوصفية التي تحتفظ بها، مع روابط سياساتها. - [نماذج الذكاء الاصطناعي عبر API: DeepSeek وGLM وQwen وGemma](https://broadrouter.com/ar/models/) (Arabic; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/ar/models/index.md): كل النماذج المتاحة على BroadNet Inference: DeepSeek V4.1 Flash وGLM 5.3 وQwen3.8 27B وGemma 4 VL ونموذجان لتحويل الصوت إلى نص، مع المعرّفات والأسعار والسياق. - [أسعار API نماذج الذكاء الاصطناعي: لكل رمز أو خطة غير محدودة](https://broadrouter.com/ar/pricing/) (Arabic; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/ar/pricing/index.md): أسعار 4 نماذج محادثة لكل مليون رمز (توكن)، وتحويل الصوت إلى نص لكل دقيقة، وخطط DeepSeek بسعر ثابت من $60 شهريًا، مع حاسبات للتكلفة. - [خطط DeepSeek API غير محدودة: البدر والمدار والشهاب](https://broadrouter.com/ar/plans/) (Arabic; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/ar/plans/index.md): جلسات DeepSeek V4.1 Flash برموز غير محدودة وسعر ثابت شهريًا أو أسبوعيًا أو بالساعة، بطلب واحد في كل مرة لكل جلسة. احسب جلساتك وقارنها بالدفع حسب الاستخدام. - [DeepSeek V4.1 Flash API: خطة غير محدودة أو الدفع لكل رمز](https://broadrouter.com/ar/models/deepseek-v4.1-flash/) (Arabic; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/ar/models/deepseek-v4.1-flash/index.md): استخدم DeepSeek V4.1 Flash عبر واجهة API متوافقة مع OpenAI، بسياق مليون رمز ودون الاحتفاظ بمحتوى الاستدلال. ادفع لكل رمز أو اشترِ خطة غير محدودة بسعر ثابت. - [واجهة GLM 5.3 API: السعر والسياق والإعداد المتوافق مع OpenAI](https://broadrouter.com/ar/models/glm-5.3/) (Arabic; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/ar/models/glm-5.3/index.md): استخدم GLM 5.3 عبر نقطة نهاية متوافقة مع OpenAI: أدوات ومخرجات منظَّمة واستدلال منطقي، وسياق مليون رمز، ودون الاحتفاظ بمحتوى الاستدلال، وبأسعار منشورة لكل رمز. - [واجهة Qwen3.8 27B API: نموذج نصوص وصور مُختبَر بالعربية](https://broadrouter.com/ar/models/qwen3.8-27b/) (Arabic; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/ar/models/qwen3.8-27b/index.md): Qwen3.8 27B عبر واجهة API متوافقة مع OpenAI: نصوص وصور، وأدوات، ومخرجات منظَّمة، واستدلال منطقي، وسياق 262K، ومُختبَر بالعربية وباللهجتين الخليجية والمصرية. - [واجهة Gemma 4 26B للرؤية: سعر منخفض لكل رمز أو مفتاح بالساعة](https://broadrouter.com/ar/models/gemma-4-vl-26b/) (Arabic; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/ar/models/gemma-4-vl-26b/index.md): نسخة BroadNet من Gemma 4 26B من Google: محادثة بالنص والصور عبر واجهة API متوافقة مع OpenAI، تعمل بالرصيد المبدئي أو بمفتاح بالساعة بسعر ثابت. - [API لتحويل الصوت العربي إلى نص بسعر لكل دقيقة](https://broadrouter.com/ar/models/cohere-transcribe-arabic/) (Arabic; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/ar/models/cohere-transcribe-arabic/index.md): حوّل الصوت العربي والرسائل الصوتية والمكالمات إلى نص عبر نقطة نهاية متوافقة مع OpenAI، بسعر لكل دقيقة صوت ودون الاحتفاظ بمحتوى الاستدلال. - [API لتحويل الصوت إلى نص في 14 لغة بسعر لكل دقيقة](https://broadrouter.com/ar/models/cohere-transcribe/) (Arabic; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/ar/models/cohere-transcribe/index.md): فرّغ الصوت في 14 لغة، منها الإنجليزية والإسبانية والفرنسية والصينية والعربية، عبر نقطة نهاية متوافقة مع OpenAI وبسعر لكل دقيقة صوت. ### Spanish (es, Español): 14 pages, hreflang es, neutral Latin-American Spanish - [BroadRouter: API de DeepSeek, GLM, Qwen, Gemma y voz árabe](https://broadrouter.com/es/) (Spanish; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/es/index.md): API compatible con OpenAI para DeepSeek, GLM, Qwen, Gemma y voz a texto en árabe. Pague por uso o compre sesiones de DeepSeek a precio fijo para sus agentes. - [API compatible con OpenAI: guía rápida con Python y cURL](https://broadrouter.com/es/developers/) (Spanish; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/es/developers/index.md): Use la API compatible con OpenAI de BroadNet Inference: envíe su primera solicitud de chat y de transcripción con Python o cURL y revise la lista de migración. - [Contacte a BroadNet sobre BroadNet Inference](https://broadrouter.com/es/contact/) (Spanish; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/es/contact/index.md): Cómo contactar a BroadNet: correo general, soporte técnico, ventas y consultas sobre endpoints privados, procesamiento regional o modelos no listados. - [API de voz a texto en árabe: modelos, formatos y costo](https://broadrouter.com/es/arabic-transcription/) (Spanish; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/es/arabic-transcription/index.md): Transcriba audio en árabe con una API de voz a texto: modelo árabe o multilingüe, formatos y tamaño máximo, estimador de costo y código para copiar. - [Privacidad de BroadNet Inference: qué datos se conservan](https://broadrouter.com/es/privacy/) (Spanish; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/es/privacy/index.md): Resumen: BroadNet Inference no conserva prompts, audio ni respuestas, ni los usa para entrenar modelos. Qué metadatos guarda y enlaces a sus políticas. - [Modelos de IA por API: DeepSeek, GLM, Qwen y Gemma](https://broadrouter.com/es/models/) (Spanish; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/es/models/index.md): Los modelos activos de BroadNet Inference: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL y dos de voz a texto, con IDs, precios y tamaños de contexto. - [Precios de API de LLM: planes ilimitados y tarifas por token](https://broadrouter.com/es/pricing/) (Spanish; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/es/pricing/index.md): Precios por millón de tokens de 4 modelos de chat, voz a texto por minuto y planes de DeepSeek con tarifa plana desde USD 60 al mes, con calculadoras de costo. - [Planes ilimitados de DeepSeek: Luna Llena, Órbita y Meteoro](https://broadrouter.com/es/plans/) (Spanish; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/es/plans/index.md): Sesiones de DeepSeek V4.1 Flash con tarifa plana por mes, semana u hora, una solicitud a la vez por sesión. Calcule sus sesiones y compare con el pago por uso. - [API de DeepSeek V4.1 Flash: plan ilimitado o pago por token](https://broadrouter.com/es/models/deepseek-v4.1-flash/) (Spanish; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/es/models/deepseek-v4.1-flash/index.md): Use DeepSeek V4.1 Flash con una API compatible con OpenAI y 1M de tokens de contexto, sin conservar el contenido. Pague por token o compre un plan ilimitado. - [API de GLM 5.3: precio, contexto y uso con el SDK de OpenAI](https://broadrouter.com/es/models/glm-5.3/) (Spanish; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/es/models/glm-5.3/index.md): GLM 5.3 por API compatible con OpenAI: herramientas, salida estructurada, razonamiento, 1M de tokens de contexto, precios publicados y sin retener contenido. - [API de Qwen3.8 27B: texto y visión, probado en árabe](https://broadrouter.com/es/models/qwen3.8-27b/) (Spanish; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/es/models/qwen3.8-27b/index.md): Qwen3.8 27B por API compatible con OpenAI: texto e imagen, herramientas, razonamiento y 262K de contexto. Probado en árabe, incluido el del Golfo y el egipcio. - [API de Gemma 4 26B con visión: precio por token o por hora](https://broadrouter.com/es/models/gemma-4-vl-26b/) (Spanish; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/es/models/gemma-4-vl-26b/index.md): Gemma 4 26B de Google, versión de BroadNet: chat con texto e imagen por API compatible con OpenAI, con el crédito inicial o una clave por hora a precio fijo. - [API de voz a texto en árabe con precio por minuto de audio](https://broadrouter.com/es/models/cohere-transcribe-arabic/) (Spanish; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/es/models/cohere-transcribe-arabic/index.md): Transcriba audio, notas de voz y llamadas en árabe con un endpoint compatible con OpenAI, con precio por minuto de audio y sin conservar el contenido. - [API de voz a texto en 14 idiomas, con precio por minuto](https://broadrouter.com/es/models/cohere-transcribe/) (Spanish; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/es/models/cohere-transcribe/index.md): Transcriba audio en 14 idiomas, entre ellos inglés, español, francés, chino y árabe, con un endpoint compatible con OpenAI y precio por minuto de audio. ### German (de, Deutsch): 14 pages, hreflang de - [BroadRouter: DeepSeek, GLM, Qwen, Gemma + Arabisch-STT-API](https://broadrouter.com/de/) (German; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/de/index.md): Eine OpenAI-kompatible API für DeepSeek, GLM, Qwen, Gemma und arabische Speech-to-Text. Zahlen Sie pro Nutzung oder buchen Sie DeepSeek-Sessions zum Festpreis. - [BroadNet Inference Schnellstart: Python und cURL](https://broadrouter.com/de/developers/) (German; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/de/developers/index.md): Senden Sie Ihre erste Chat- und Transkriptionsanfrage an BroadNet Inference mit Python oder cURL, und gehen Sie die Migrations-Checkliste für Ihr Modell durch. - [BroadNet zu BroadNet Inference kontaktieren](https://broadrouter.com/de/contact/) (German; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/de/contact/index.md): So erreichen Sie BroadNet zu BroadNet Inference: E-Mail, Support, Vertrieb sowie private Endpoints, regionale Verarbeitung oder weitere Modelle. - [Speech-to-Text API für Arabisch: Modelle, Formate, Preise](https://broadrouter.com/de/arabic-transcription/) (German; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/de/arabic-transcription/index.md): Arabisches Audio mit BroadNet Inference transkribieren: Arabisch- und Multilingual-Modell im Vergleich, Formate, Größenlimit, Kostenrechner und Code. - [Datenschutz: BroadNet Inference und diese Website](https://broadrouter.com/de/privacy/) (German; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/de/privacy/index.md): Kurz erklärt: Wie BroadNet Inference Prompts, Audio und Antworten behandelt, welche Metadaten bleiben, plus Links zu den offiziellen Richtlinien. - [DeepSeek, GLM, Qwen, Gemma: Modelle auf BroadNet Inference](https://broadrouter.com/de/models/) (German; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/de/models/index.md): Alle Modelle auf BroadNet Inference: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL und zwei Speech-to-Text-Modelle, mit IDs, Preisen, Kontext. - [LLM-API-Preise: Flatrate-Pläne und Preise pro Token](https://broadrouter.com/de/pricing/) (German; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/de/pricing/index.md): Token-Preise für 4 Chat-Modelle, Minutenpreise für Audio und DeepSeek-Session-Pläne ab USD 60 pro Monat, mit Kostenrechnern für Ihren Workload. - [DeepSeek-Flatrate: Session-Pläne Full Moon, Orbit, Meteor](https://broadrouter.com/de/plans/) (German; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/de/plans/index.md): DeepSeek V4.1 Flash-Sessions zum Festpreis pro Monat, Woche oder Stunde, eine Anfrage gleichzeitig pro Session. Sessions planen, mit Pay-as-you-go vergleichen. - [DeepSeek V4.1 Flash API: Flatrate-Plan oder Preis pro Token](https://broadrouter.com/de/models/deepseek-v4.1-flash/) (German; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/de/models/deepseek-v4.1-flash/index.md): DeepSeek V4.1 Flash über eine OpenAI-kompatible API, mit 1 Mio. Tokens Kontext und ohne gespeicherte Inhalte. Pro Token zahlen oder einen Festpreis-Plan buchen. - [GLM 5.3 API: Preis, Kontext und OpenAI-kompatibles Setup](https://broadrouter.com/de/models/glm-5.3/) (German; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/de/models/glm-5.3/index.md): GLM 5.3 per OpenAI-kompatiblem Endpoint: Tools, strukturierte Ausgabe, Reasoning, 1 Mio. Tokens Kontext, keine Speicherung, veröffentlichte Preise. - [Qwen3.8 27B API: Text + Vision, auf Arabisch getestet](https://broadrouter.com/de/models/qwen3.8-27b/) (German; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/de/models/qwen3.8-27b/index.md): Qwen3.8 27B per OpenAI-kompatibler API: Text und Bild, Tools, strukturierte Ausgabe, Reasoning, 262K Kontext. Auf Arabisch inkl. Golf und Ägypten getestet. - [Gemma 4 26B Vision API: günstig pro Token oder pro Stunde](https://broadrouter.com/de/models/gemma-4-vl-26b/) (German; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/de/models/gemma-4-vl-26b/index.md): Gemma 4 26B von Google, BroadNet-Version: Text- und Bild-Chat per OpenAI-kompatibler API, mit Startguthaben oder Festpreis-Stundenschlüssel. - [Speech-to-Text API für Arabisch, Preis pro Audiominute](https://broadrouter.com/de/models/cohere-transcribe-arabic/) (German; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/de/models/cohere-transcribe-arabic/index.md): Arabisches Audio, Sprachnachrichten und Anrufe per OpenAI-kompatiblem Endpoint transkribieren: Preis pro Audiominute, keine gespeicherten Inhalte. - [Speech-to-Text API in 14 Sprachen, Preis pro Minute](https://broadrouter.com/de/models/cohere-transcribe/) (German; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/de/models/cohere-transcribe/index.md): Audio in 14 Sprachen transkribieren, u. a. Englisch, Spanisch, Französisch, Chinesisch, Arabisch, per OpenAI-kompatiblem Endpoint, Preis pro Minute. ### Portuguese (pt, Português): 14 pages, hreflang pt-BR, Brazilian Portuguese - [BroadRouter: API DeepSeek, GLM, Qwen, Gemma e voz em árabe](https://broadrouter.com/pt/) (Portuguese; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/pt/index.md): Uma API compatível com a OpenAI para DeepSeek, GLM, Qwen, Gemma e transcrição em árabe. Pague por uso ou compre sessões DeepSeek a preço fixo para seus agentes. - [Guia rápido da BroadNet Inference: API em Python e cURL](https://broadrouter.com/pt/developers/) (Portuguese; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/pt/developers/index.md): Envie sua primeira requisição de chat e de transcrição para a BroadNet Inference com Python ou cURL e siga o checklist de migração do modelo escolhido. - [Fale com a BroadNet sobre a BroadNet Inference](https://broadrouter.com/pt/contact/) (Portuguese; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/pt/contact/index.md): Como falar com a BroadNet sobre a BroadNet Inference: e-mail, suporte, vendas e o caminho para endpoint privado, processamento regional ou outros modelos. - [API de transcrição em árabe: modelos, formatos e preço](https://broadrouter.com/pt/arabic-transcription/) (Portuguese; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/pt/arabic-transcription/index.md): Transcreva áudio em árabe com a BroadNet Inference: modelos árabe e multilíngue comparados, formatos, limite de tamanho, estimativa de custo e código pronto. - [Resumo de privacidade: BroadNet Inference e este site](https://broadrouter.com/pt/privacy/) (Portuguese; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/pt/privacy/index.md): Um resumo curto de como a BroadNet Inference trata seus prompts, áudios e respostas, quais metadados ela guarda e links para as políticas oficiais da BroadNet. - [Modelos de LLM na BroadNet Inference: DeepSeek, GLM, Qwen](https://broadrouter.com/pt/models/) (Portuguese; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/pt/models/index.md): Todos os modelos ativos na BroadNet Inference: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL e dois de transcrição, com IDs, preços e contexto. - [Preço de API de LLM: planos ilimitados e preço por token](https://broadrouter.com/pt/pricing/) (Portuguese; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/pt/pricing/index.md): Preço por token de 4 modelos de chat, preço por minuto de áudio e planos de sessão DeepSeek a partir de US$ 60 por mês, com calculadoras de custo. - [Planos de sessão DeepSeek: Full Moon, Orbit e Meteor](https://broadrouter.com/pt/plans/) (Portuguese; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/pt/plans/index.md): Sessões de DeepSeek V4.1 Flash a preço fixo por mês, semana ou hora, uma requisição por vez por sessão. Planeje suas sessões e compare com o pagamento por uso. - [API do DeepSeek V4.1 Flash: plano ilimitado ou por token](https://broadrouter.com/pt/models/deepseek-v4.1-flash/) (Portuguese; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/pt/models/deepseek-v4.1-flash/index.md): Use o DeepSeek V4.1 Flash por API compatível com OpenAI, com contexto de 1M de tokens e sem retenção de conteúdo. Pague por token ou compre um plano ilimitado. - [API do GLM 5.3: preço, contexto e uso compatível com OpenAI](https://broadrouter.com/pt/models/glm-5.3/) (Portuguese; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/pt/models/glm-5.3/index.md): GLM 5.3 por endpoint compatível com a OpenAI: ferramentas, saída estruturada, raciocínio, contexto de 1M de tokens e preço por token. Sem retenção de conteúdo. - [API do Qwen3.8 27B: modelo de texto e visão testado em árabe](https://broadrouter.com/pt/models/qwen3.8-27b/) (Portuguese; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/pt/models/qwen3.8-27b/index.md): Qwen3.8 27B por API compatível com a OpenAI: texto e visão, ferramentas, saída estruturada, raciocínio, contexto de 262K. Testado em árabe do Golfo e do Egito. - [API de visão Gemma 4 26B: barata por token ou chave por hora](https://broadrouter.com/pt/models/gemma-4-vl-26b/) (Portuguese; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/pt/models/gemma-4-vl-26b/index.md): O Gemma 4 26B do Google na versão da BroadNet: chat com texto e imagem por API compatível com a OpenAI, com o crédito inicial ou chave por hora a preço fixo. - [API de transcrição de áudio em árabe com preço por minuto](https://broadrouter.com/pt/models/cohere-transcribe-arabic/) (Portuguese; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/pt/models/cohere-transcribe-arabic/index.md): Transcreva áudio, mensagens de voz e ligações em árabe por um endpoint de transcrição compatível com a OpenAI, com preço por minuto e sem retenção de conteúdo. - [API de transcrição de áudio em 14 idiomas, preço por minuto](https://broadrouter.com/pt/models/cohere-transcribe/) (Portuguese; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/pt/models/cohere-transcribe/index.md): Transcreva áudio em 14 idiomas, entre eles inglês, espanhol, francês, chinês e árabe, por um endpoint compatível com a OpenAI e com preço por minuto de áudio. ### French (fr, Français): 14 pages, hreflang fr - [BroadRouter : API DeepSeek, GLM, Qwen, Gemma et voix arabe](https://broadrouter.com/fr/) (French; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/fr/index.md): API compatible OpenAI pour DeepSeek, GLM, Qwen, Gemma et le speech-to-text arabe. Payez à l’usage ou prenez des sessions DeepSeek à prix fixe pour vos agents. - [Démarrage rapide BroadNet Inference : Python et cURL](https://broadrouter.com/fr/developers/) (French; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/fr/developers/index.md): Envoyez votre première requête de chat et de transcription à BroadNet Inference avec Python ou cURL, puis suivez la checklist de migration pour votre modèle. - [Contacter BroadNet au sujet de BroadNet Inference](https://broadrouter.com/fr/contact/) (French; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/fr/contact/index.md): Comment joindre BroadNet pour BroadNet Inference : e-mail, support, ventes, et la démarche pour un endpoint privé, une région ou un modèle non listé. - [API speech-to-text arabe : modèles, formats et prix/minute](https://broadrouter.com/fr/arabic-transcription/) (French; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/fr/arabic-transcription/index.md): Transcrivez de l’audio arabe avec BroadNet Inference : modèles arabe et multilingue comparés, formats, taille maximale, estimateur de coût et code à copier. - [Confidentialité : BroadNet Inference et ce site, en bref](https://broadrouter.com/fr/privacy/) (French; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/fr/privacy/index.md): Résumé : comment BroadNet Inference traite vos prompts, votre audio et vos réponses, quelles métadonnées il conserve, et liens vers ses politiques. - [Modèles sur BroadNet Inference : DeepSeek, GLM, Qwen, Gemma](https://broadrouter.com/fr/models/) (French; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/fr/models/index.md): Les modèles de BroadNet Inference : DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL et deux modèles speech-to-text, avec IDs, prix et taille de contexte. - [Prix API LLM : forfaits illimités et tarifs au token](https://broadrouter.com/fr/pricing/) (French; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/fr/pricing/index.md): Prix au token de 4 modèles de chat, prix audio à la minute et forfaits par session DeepSeek à partir de USD 60 par mois, avec des calculateurs de coût. - [Forfaits DeepSeek illimités : Full Moon, Orbit et Meteor](https://broadrouter.com/fr/plans/) (French; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/fr/plans/index.md): Sessions DeepSeek V4.1 Flash à prix fixe (mois, semaine ou heure), une requête à la fois par session. Planifiez vos sessions et comparez au paiement à l’usage. - [API DeepSeek V4.1 Flash : forfait illimité ou prix au token](https://broadrouter.com/fr/models/deepseek-v4.1-flash/) (French; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/fr/models/deepseek-v4.1-flash/index.md): DeepSeek V4.1 Flash via API compatible OpenAI, contexte d’un million de tokens, contenu non conservé. Payez au token ou prenez un forfait illimité à prix fixe. - [API GLM 5.3 : prix, contexte et config compatible OpenAI](https://broadrouter.com/fr/models/glm-5.3/) (French; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/fr/models/glm-5.3/index.md): GLM 5.3 via un endpoint compatible OpenAI : outils, sortie structurée, raisonnement, contexte d’un million de tokens, contenu non conservé, prix publiés. - [API Qwen3.8 27B : modèle texte et vision testé en arabe](https://broadrouter.com/fr/models/qwen3.8-27b/) (French; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/fr/models/qwen3.8-27b/index.md): Qwen3.8 27B via API compatible OpenAI : texte et image, outils, sortie structurée, raisonnement, contexte 262K. Testé en arabe, dialectes du Golfe et égyptien. - [API vision Gemma 4 26B : bas prix au token ou clé horaire](https://broadrouter.com/fr/models/gemma-4-vl-26b/) (French; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/fr/models/gemma-4-vl-26b/index.md): Gemma 4 26B de Google, version BroadNet : chat texte et image via une API compatible OpenAI, avec le crédit de départ ou une clé horaire à prix fixe. - [API speech-to-text arabe, prix à la minute d’audio](https://broadrouter.com/fr/models/cohere-transcribe-arabic/) (French; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/fr/models/cohere-transcribe-arabic/index.md): Transcrivez audio, notes vocales et appels en arabe via un endpoint compatible OpenAI, au prix à la minute d’audio, sans conservation du contenu. - [API speech-to-text en 14 langues, prix à la minute](https://broadrouter.com/fr/models/cohere-transcribe/) (French; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/fr/models/cohere-transcribe/index.md): Transcrivez de l’audio en 14 langues, dont l’anglais, l’espagnol, le français, le chinois et l’arabe, via un endpoint compatible OpenAI, au prix à la minute. ### Italian (it, Italiano): 14 pages, hreflang it - [BroadRouter: API DeepSeek, GLM, Qwen, Gemma e voce araba](https://broadrouter.com/it/) (Italian; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/it/index.md): Un’API compatibile con OpenAI per DeepSeek, GLM, Qwen, Gemma e speech-to-text in arabo. Paga a consumo o compra sessioni DeepSeek a prezzo fisso per gli agenti. - [Guida rapida a BroadNet Inference: Python e cURL](https://broadrouter.com/it/developers/) (Italian; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/it/developers/index.md): Invia la tua prima richiesta di chat e di trascrizione a BroadNet Inference con Python o cURL, poi segui la checklist di migrazione per il modello scelto. - [Contatta BroadNet per BroadNet Inference](https://broadrouter.com/it/contact/) (Italian; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/it/contact/index.md): Come contattare BroadNet per BroadNet Inference: email, supporto tecnico, vendite e richieste di endpoint privati, elaborazione regionale o altri modelli. - [API speech-to-text in arabo: modelli, formati e prezzi](https://broadrouter.com/it/arabic-transcription/) (Italian; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/it/arabic-transcription/index.md): Trascrivi audio in arabo con BroadNet Inference: modello arabo e multilingue a confronto, formati e dimensione massima, stima dei costi e codice da copiare. - [Privacy in breve: BroadNet Inference e questo sito](https://broadrouter.com/it/privacy/) (Italian; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/it/privacy/index.md): Un riassunto di come BroadNet Inference tratta i tuoi prompt, il tuo audio e le risposte, quali metadati conserva, con i link alle policy ufficiali di BroadNet. - [Modelli su BroadNet Inference: DeepSeek, GLM, Qwen, Gemma](https://broadrouter.com/it/models/) (Italian; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/it/models/index.md): I modelli attivi su BroadNet Inference: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL e due speech-to-text, con ID, prezzi e dimensione del contesto. - [Prezzi API LLM: piani illimitati e tariffe per token](https://broadrouter.com/it/pricing/) (Italian; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/it/pricing/index.md): Prezzi per token di 4 modelli di chat, prezzi audio al minuto e piani a sessione DeepSeek da USD 60 al mese, con calcolatori dei costi. - [Piani DeepSeek illimitati: Full Moon, Orbit e Meteor](https://broadrouter.com/it/plans/) (Italian; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/it/plans/index.md): Sessioni DeepSeek V4.1 Flash a prezzo fisso al mese, alla settimana o all’ora, una richiesta alla volta per sessione. Calcola le sessioni e confronta i costi. - [API DeepSeek V4.1 Flash: piano illimitato o prezzo per token](https://broadrouter.com/it/models/deepseek-v4.1-flash/) (Italian; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/it/models/deepseek-v4.1-flash/index.md): DeepSeek V4.1 Flash via API compatibile con OpenAI, contesto da un milione di token, contenuti non conservati. Paga per token o scegli un piano a prezzo fisso. - [API GLM 5.3: prezzi, contesto e setup compatibile OpenAI](https://broadrouter.com/it/models/glm-5.3/) (Italian; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/it/models/glm-5.3/index.md): GLM 5.3 tramite un endpoint compatibile con OpenAI: tool calling, output strutturato, ragionamento, contesto da un milione di token, prezzi pubblicati. - [API Qwen3.8 27B: modello testo e visione testato in arabo](https://broadrouter.com/it/models/qwen3.8-27b/) (Italian; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/it/models/qwen3.8-27b/index.md): Qwen3.8 27B via API compatibile con OpenAI: testo e immagini, tool calling, output strutturato, contesto 262K. Testato in arabo, dialetti del Golfo ed egiziano. - [API vision Gemma 4 26B: prezzo per token o chiave oraria](https://broadrouter.com/it/models/gemma-4-vl-26b/) (Italian; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/it/models/gemma-4-vl-26b/index.md): Gemma 4 26B di Google, versione BroadNet: chat testo e immagini via API compatibile con OpenAI, con il credito iniziale o una chiave oraria a prezzo fisso. - [API speech-to-text arabo, prezzo al minuto di audio](https://broadrouter.com/it/models/cohere-transcribe-arabic/) (Italian; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/it/models/cohere-transcribe-arabic/index.md): Trascrivi audio, messaggi vocali e chiamate in arabo tramite un endpoint compatibile con OpenAI, con prezzo al minuto di audio e senza conservare i contenuti. - [API speech-to-text in 14 lingue, prezzo al minuto](https://broadrouter.com/it/models/cohere-transcribe/) (Italian; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/it/models/cohere-transcribe/index.md): Trascrivi audio in 14 lingue, tra cui inglese, spagnolo, italiano, cinese e arabo, tramite un endpoint compatibile con OpenAI, con prezzo al minuto di audio. ### Chinese (Simplified) (zh, 简体中文): 14 pages, hreflang zh-Hans, Simplified Chinese script - [BroadRouter:DeepSeek、GLM、Qwen、Gemma 与阿拉伯语语音 API](https://broadrouter.com/zh/) (Chinese (Simplified); translation of https://broadrouter.com/; Markdown: https://broadrouter.com/zh/index.md): 一个 OpenAI 兼容接口,接入 DeepSeek、GLM、Qwen、Gemma 和阿拉伯语语音转文字。按量付费,或为你的智能体购买固定价格的 DeepSeek 会话。 - [BroadNet Inference 开发者快速入门:Python 与 cURL](https://broadrouter.com/zh/developers/) (Chinese (Simplified); translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/zh/developers/index.md): 用 Python 或 cURL 向 BroadNet Inference 发送第一个对话请求和转写请求,再按所选模型逐项完成迁移清单。 - [就 BroadNet Inference 联系 BroadNet](https://broadrouter.com/zh/contact/) (Chinese (Simplified); translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/zh/contact/index.md): 如何就 BroadNet Inference 联系 BroadNet:通用邮箱、技术支持、销售,以及私有端点、区域处理或未列出模型的咨询途径。 - [阿拉伯语语音转文字 API:模型、格式和每分钟价格](https://broadrouter.com/zh/arabic-transcription/) (Chinese (Simplified); translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/zh/arabic-transcription/index.md): 通过 BroadNet Inference 转写阿拉伯语音频:阿拉伯语模型与多语言模型对比、支持格式和大小上限、费用估算器,以及可直接复制的代码。 - [隐私摘要:BroadNet Inference 与本网站](https://broadrouter.com/zh/privacy/) (Chinese (Simplified); translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/zh/privacy/index.md): 简要说明 BroadNet Inference 如何处理你的提示词、音频和回答,保留哪些元数据,并附 BroadNet 官方政策链接。 - [BroadNet Inference 模型:DeepSeek、GLM、Qwen、Gemma API](https://broadrouter.com/zh/models/) (Chinese (Simplified); translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/zh/models/index.md): BroadNet Inference 全部在线模型:DeepSeek V4.1 Flash、GLM 5.3、Qwen3.8 27B、Gemma 4 VL 和两个语音转文字模型,附模型 ID、价格和上下文长度。 - [大模型 API 价格:不限量套餐与按 token 计价](https://broadrouter.com/zh/pricing/) (Chinese (Simplified); translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/zh/pricing/index.md): 4 个对话模型的按 token 价格、按分钟计价的音频价格,以及 US$60/月起的 DeepSeek 会话套餐,附费用估算器。 - [DeepSeek 会话套餐:Full Moon、Orbit 和 Meteor](https://broadrouter.com/zh/plans/) (Chinese (Simplified); translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/zh/plans/index.md): 按月、周或小时购买固定价格的 DeepSeek V4.1 Flash 会话,每个会话同一时间一个请求。规划会话数量,并与按量付费比较。 - [DeepSeek V4.1 Flash API:不限量套餐或按 token 付费](https://broadrouter.com/zh/models/deepseek-v4.1-flash/) (Chinese (Simplified); translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/zh/models/deepseek-v4.1-flash/index.md): 通过 OpenAI 兼容接口调用 DeepSeek V4.1 Flash,上下文 1M token,推理内容不保留。可按 token 付费,也可购买固定价格的不限量套餐。 - [GLM 5.3 API:价格、上下文与 OpenAI 兼容配置](https://broadrouter.com/zh/models/glm-5.3/) (Chinese (Simplified); translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/zh/models/glm-5.3/index.md): 通过 OpenAI 兼容端点使用 GLM 5.3:工具调用、结构化输出、思考、1M token 上下文,按 token 公布价格。推理内容不保留。 - [Qwen3.8 27B API:经阿拉伯语测试的视觉与文本模型](https://broadrouter.com/zh/models/qwen3.8-27b/) (Chinese (Simplified); translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/zh/models/qwen3.8-27b/index.md): 通过 OpenAI 兼容 API 调用 Qwen3.8 27B:文本和视觉、工具调用、结构化输出和思考,上下文 262K,经阿拉伯语及海湾、埃及方言测试。 - [Gemma 4 26B 视觉 API:按 token 低价或按小时密钥](https://broadrouter.com/zh/models/gemma-4-vl-26b/) (Chinese (Simplified); translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/zh/models/gemma-4-vl-26b/index.md): BroadNet 版 Google Gemma 4 26B:通过 OpenAI 兼容 API 进行文本和图像对话,可使用初始额度,或购买固定价格的按小时密钥。 - [阿拉伯语语音转文字 API,按音频分钟计价](https://broadrouter.com/zh/models/cohere-transcribe-arabic/) (Chinese (Simplified); translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/zh/models/cohere-transcribe-arabic/index.md): 通过 OpenAI 兼容的转写端点转写阿拉伯语音频、语音消息和通话,按音频分钟计价。推理内容不保留。 - [支持 14 种语言的语音转文字 API,按分钟计价](https://broadrouter.com/zh/models/cohere-transcribe/) (Chinese (Simplified); translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/zh/models/cohere-transcribe/index.md): 转写 14 种语言的音频,包括英语、西班牙语、法语、中文和阿拉伯语,使用 OpenAI 兼容端点,按音频分钟计价。 ### Turkish (tr, Türkçe): 14 pages, hreflang tr - [BroadRouter: DeepSeek, GLM, Qwen, Gemma ve Arapça ses API](https://broadrouter.com/tr/) (Turkish; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/tr/index.md): DeepSeek, GLM, Qwen, Gemma ve Arapça ses-metin için tek OpenAI uyumlu API. Kullandıkça ödeyin ya da ajanlarınız için sabit fiyatlı DeepSeek oturumu alın. - [BroadNet Inference hızlı başlangıç: Python ve cURL ile API](https://broadrouter.com/tr/developers/) (Turkish; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/tr/developers/index.md): İlk sohbet ve transkripsiyon isteğinizi Python veya cURL ile BroadNet Inference'a gönderin, ardından seçtiğiniz model için geçiş kontrol listesini izleyin. - [BroadNet Inference için BroadNet'e ulaşın](https://broadrouter.com/tr/contact/) (Turkish; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/tr/contact/index.md): BroadNet Inference için BroadNet'e nasıl ulaşılır: genel e-posta, teknik destek, satış ve özel uç nokta, bölgesel işleme veya listede olmayan modeller. - [Arapça ses metne çevirme API: modeller, formatlar, fiyat](https://broadrouter.com/tr/arabic-transcription/) (Turkish; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/tr/arabic-transcription/index.md): BroadNet Inference ile Arapça sesi metne dönüştürün: Arapça ve çok dilli modellerin karşılaştırması, formatlar, boyut sınırı, maliyet hesaplayıcı ve hazır kod. - [Gizlilik özeti: BroadNet Inference ve bu web sitesi](https://broadrouter.com/tr/privacy/) (Turkish; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/tr/privacy/index.md): BroadNet Inference'ın istemlerinizi, seslerinizi ve yanıtlarınızı nasıl işlediği, hangi meta verileri sakladığı: kısa özet ve resmi BroadNet politikaları. - [BroadNet Inference LLM modelleri: DeepSeek, GLM, Qwen, Gemma](https://broadrouter.com/tr/models/) (Turkish; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/tr/models/index.md): BroadNet Inference'taki tüm aktif modeller: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL ve iki konuşmadan metne modeli; kimlikler, fiyatlar ve bağlam. - [LLM API fiyatları: sınırsız planlar ve token başına ücret](https://broadrouter.com/tr/pricing/) (Turkish; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/tr/pricing/index.md): 4 sohbet modeli için token başına fiyat, dakika başına ses fiyatı ve aylık US$60 fiyatından başlayan DeepSeek oturum planları; maliyet hesaplayıcılarla. - [DeepSeek oturum planları: Full Moon, Orbit ve Meteor](https://broadrouter.com/tr/plans/) (Turkish; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/tr/plans/index.md): Aylık, haftalık veya saatlik sabit fiyatlı DeepSeek V4.1 Flash oturumları, oturum başına tek istek. Oturumlarınızı planlayın, kullandıkça öde ile kıyaslayın. - [DeepSeek V4.1 Flash API: sınırsız plan veya token başına](https://broadrouter.com/tr/models/deepseek-v4.1-flash/) (Turkish; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/tr/models/deepseek-v4.1-flash/index.md): DeepSeek V4.1 Flash'ı 1M token bağlamla OpenAI uyumlu API üzerinden kullanın; çıkarım içeriği saklanmaz. Token başına ödeyin veya sınırsız sabit plan alın. - [GLM 5.3 API: fiyat, bağlam ve OpenAI uyumlu kurulum](https://broadrouter.com/tr/models/glm-5.3/) (Turkish; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/tr/models/glm-5.3/index.md): GLM 5.3'ü OpenAI uyumlu bir uç nokta üzerinden kullanın: araçlar, yapılandırılmış çıktı, akıl yürütme, 1M token bağlam ve token başına fiyat. İçerik saklanmaz. - [Qwen3.8 27B API: Arapçada test edilmiş metin ve görüntü](https://broadrouter.com/tr/models/qwen3.8-27b/) (Turkish; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/tr/models/qwen3.8-27b/index.md): Qwen3.8 27B, OpenAI uyumlu API: metin ve görüntü, araçlar, yapılandırılmış çıktı, akıl yürütme, 262K bağlam. Arapça, Körfez ve Mısır lehçelerinde test edildi. - [Gemma 4 26B görüntü API: ucuz token veya saatlik anahtar](https://broadrouter.com/tr/models/gemma-4-vl-26b/) (Turkish; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/tr/models/gemma-4-vl-26b/index.md): Google'ın Gemma 4 26B modelinin BroadNet sürümü: OpenAI uyumlu API ile metin ve görsel sohbet; başlangıç kredisiyle veya sabit fiyatlı saatlik anahtarla. - [Arapça ses metne çevirme API: dakika başına fiyat](https://broadrouter.com/tr/models/cohere-transcribe-arabic/) (Turkish; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/tr/models/cohere-transcribe-arabic/index.md): Arapça sesleri, sesli mesajları ve aramaları OpenAI uyumlu bir transkripsiyon uç noktasıyla metne dönüştürün; ses dakikası başına fiyat, içerik saklanmaz. - [14 dilde konuşmadan metne API, dakika başına fiyat](https://broadrouter.com/tr/models/cohere-transcribe/) (Turkish; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/tr/models/cohere-transcribe/index.md): İngilizce, İspanyolca, Fransızca, Çince ve Arapça dahil 14 dilde sesi, dakika başına fiyatlı OpenAI uyumlu bir uç noktayla metne dönüştürün. ### Urdu (ur, اردو): 14 pages, hreflang ur, right-to-left, Arabic script - [BroadRouter: DeepSeek، GLM، Qwen، Gemma اور عربی اسپیچ API](https://broadrouter.com/ur/) (Urdu; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/ur/index.md): DeepSeek، GLM، Qwen، Gemma اور عربی اسپیچ ٹو ٹیکسٹ کے لیے OpenAI سے مطابقت رکھنے والا ایک API۔ استعمال کے مطابق ادائیگی، یا ایجنٹس کے لیے مقررہ قیمت کے سیشن۔ - [BroadNet Inference ڈویلپر کوئک اسٹارٹ: Python اور cURL](https://broadrouter.com/ur/developers/) (Urdu; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/ur/developers/index.md): Python یا cURL سے BroadNet Inference کو اپنی پہلی چیٹ اور ٹرانسکرپشن درخواست بھیجیں، پھر منتخب ماڈل کے لیے منتقلی کی چیک لسٹ مکمل کریں۔ - [BroadNet Inference کے بارے میں BroadNet سے رابطہ کریں](https://broadrouter.com/ur/contact/) (Urdu; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/ur/contact/index.md): BroadNet Inference کے بارے میں BroadNet سے کیسے رابطہ کریں: عمومی ای میل، تکنیکی سپورٹ، سیلز، اور پرائیویٹ اینڈپوائنٹ، علاقائی پروسیسنگ یا نئے ماڈلز کی درخواست۔ - [عربی ٹرانسکرپشن API: ماڈلز، فارمیٹس اور فی منٹ لاگت](https://broadrouter.com/ur/arabic-transcription/) (Urdu; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/ur/arabic-transcription/index.md): BroadNet Inference کے ذریعے عربی آڈیو کو متن میں بدلیں: عربی اور کثیر لسانی ماڈلز کا موازنہ، فارمیٹس اور سائز کی حد، لاگت کا تخمینہ اور کاپی کے قابل کوڈ۔ - [پرائیویسی کا خلاصہ: BroadNet Inference اور یہ ویب سائٹ](https://broadrouter.com/ur/privacy/) (Urdu; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/ur/privacy/index.md): BroadNet Inference آپ کے پرامپٹس، آڈیو اور جوابات کو کیسے سنبھالتا ہے اور کون سا میٹا ڈیٹا رکھتا ہے، اس کا خلاصہ، ساتھ BroadNet کی سرکاری پالیسیوں کے لنکس۔ - [BroadNet Inference کے ماڈلز: DeepSeek، GLM، Qwen، Gemma API](https://broadrouter.com/ur/models/) (Urdu; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/ur/models/index.md): BroadNet Inference کے تمام لائیو ماڈلز: DeepSeek V4.1 Flash، GLM 5.3، Qwen3.8 27B، Gemma 4 VL اور دو اسپیچ ٹو ٹیکسٹ ماڈلز، IDs، قیمتوں اور کانٹیکسٹ کے ساتھ۔ - [LLM API قیمتیں: لامحدود پلانز اور فی ٹوکن ریٹس](https://broadrouter.com/ur/pricing/) (Urdu; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/ur/pricing/index.md): 4 چیٹ ماڈلز کی فی ٹوکن قیمتیں، فی منٹ آڈیو قیمتیں اور US$60 ماہانہ سے DeepSeek سیشن پلانز، لاگت کے تخمینے کے ٹولز کے ساتھ۔ - [DeepSeek سیشن پلانز: Full Moon، Orbit اور Meteor](https://broadrouter.com/ur/plans/) (Urdu; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/ur/plans/index.md): مہینے، ہفتے یا گھنٹے کے لیے مقررہ قیمت کے DeepSeek V4.1 Flash سیشن، فی سیشن ایک وقت میں ایک درخواست۔ سیشنز طے کریں اور استعمال کے مطابق ادائیگی سے موازنہ کریں۔ - [DeepSeek V4.1 Flash API: لامحدود پلان یا فی ٹوکن](https://broadrouter.com/ur/models/deepseek-v4.1-flash/) (Urdu; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/ur/models/deepseek-v4.1-flash/index.md): OpenAI سے مطابقت رکھنے والے API پر DeepSeek V4.1 Flash، دس لاکھ ٹوکن کا کانٹیکسٹ۔ انفرنس کا مواد محفوظ نہیں ہوتا۔ فی ٹوکن ادائیگی کریں یا لامحدود پلان لیں۔ - [GLM 5.3 API: قیمت، کانٹیکسٹ اور OpenAI سے مطابقت](https://broadrouter.com/ur/models/glm-5.3/) (Urdu; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/ur/models/glm-5.3/index.md): OpenAI سے مطابقت رکھنے والے اینڈپوائنٹ پر GLM 5.3: ٹولز، اسٹرکچرڈ آؤٹ پٹ، ریزننگ، دس لاکھ ٹوکن کا کانٹیکسٹ اور فی ٹوکن قیمتیں۔ مواد محفوظ نہیں ہوتا۔ - [Qwen3.8 27B API: عربی میں ٹیسٹ شدہ وژن اور ٹیکسٹ ماڈل](https://broadrouter.com/ur/models/qwen3.8-27b/) (Urdu; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/ur/models/qwen3.8-27b/index.md): OpenAI سے مطابقت رکھنے والے API سے Qwen3.8 27B: ٹیکسٹ اور وژن، ٹولز، اسٹرکچرڈ آؤٹ پٹ، ریزننگ، 262K کانٹیکسٹ؛ عربی اور خلیجی و مصری لہجوں میں ٹیسٹ شدہ۔ - [Gemma 4 26B وژن API: سستا فی ٹوکن یا گھنٹے والی کلید](https://broadrouter.com/ur/models/gemma-4-vl-26b/) (Urdu; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/ur/models/gemma-4-vl-26b/index.md): Google کے Gemma 4 26B کا BroadNet بلڈ: OpenAI سے مطابقت رکھنے والے API سے ٹیکسٹ اور تصویری چیٹ، ابتدائی کریڈٹ پر یا مقررہ قیمت کی گھنٹے والی کلید کے طور پر۔ - [عربی اسپیچ ٹو ٹیکسٹ API، فی آڈیو منٹ قیمت](https://broadrouter.com/ur/models/cohere-transcribe-arabic/) (Urdu; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/ur/models/cohere-transcribe-arabic/index.md): OpenAI سے مطابقت رکھنے والے ٹرانسکرپشن اینڈپوائنٹ سے عربی آڈیو، وائس نوٹس اور کالز کو متن میں بدلیں، فی آڈیو منٹ قیمت پر۔ انفرنس کا مواد محفوظ نہیں رکھا جاتا۔ - [14 زبانوں میں اسپیچ ٹو ٹیکسٹ API، فی منٹ قیمت](https://broadrouter.com/ur/models/cohere-transcribe/) (Urdu; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/ur/models/cohere-transcribe/index.md): انگریزی، ہسپانوی، فرانسیسی، چینی اور عربی سمیت 14 زبانوں میں آڈیو کو متن میں بدلیں، OpenAI سے مطابقت رکھنے والے اینڈپوائنٹ سے، فی آڈیو منٹ قیمت پر۔ ### Hindi (hi, हिन्दी): 14 pages, hreflang hi, Devanagari script - [BroadRouter: DeepSeek, GLM, Qwen, Gemma और अरबी स्पीच API](https://broadrouter.com/hi/) (Hindi; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/hi/index.md): DeepSeek, GLM, Qwen, Gemma और अरबी स्पीच-टू-टेक्स्ट के लिए एक OpenAI-कम्पैटिबल API। उपयोग के अनुसार भुगतान करें या एजेंट के लिए फ़िक्स्ड-कीमत DeepSeek सेशन लें। - [BroadNet Inference डेवलपर क्विकस्टार्ट: Python और cURL](https://broadrouter.com/hi/developers/) (Hindi; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/hi/developers/index.md): Python या cURL से BroadNet Inference को अपना पहला चैट और ट्रांसक्रिप्शन अनुरोध भेजें, फिर चुने गए मॉडल के लिए माइग्रेशन चेकलिस्ट पूरी करें। - [BroadNet Inference के बारे में BroadNet से संपर्क करें](https://broadrouter.com/hi/contact/) (Hindi; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/hi/contact/index.md): BroadNet Inference के बारे में BroadNet तक कैसे पहुँचें: सामान्य ईमेल, तकनीकी सपोर्ट, सेल्स, और प्राइवेट एंडपॉइंट, क्षेत्रीय प्रोसेसिंग या नए मॉडल की पूछताछ। - [अरबी ट्रांसक्रिप्शन API: मॉडल, फ़ॉर्मैट और प्रति मिनट लागत](https://broadrouter.com/hi/arabic-transcription/) (Hindi; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/hi/arabic-transcription/index.md): BroadNet Inference से अरबी ऑडियो ट्रांसक्राइब करें: अरबी और बहुभाषी मॉडल की तुलना, फ़ॉर्मैट और साइज़ सीमा, लागत अनुमान और कॉपी करने लायक कोड। - [प्राइवेसी सारांश: BroadNet Inference और यह वेबसाइट](https://broadrouter.com/hi/privacy/) (Hindi; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/hi/privacy/index.md): BroadNet Inference आपके प्रॉम्प्ट, ऑडियो और जवाबों को कैसे सँभालता है, कौन-सा मेटाडेटा रखता है, इसका छोटा सारांश, साथ में BroadNet की आधिकारिक नीतियों के लिंक। - [BroadNet Inference के मॉडल: DeepSeek, GLM, Qwen, Gemma API](https://broadrouter.com/hi/models/) (Hindi; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/hi/models/index.md): BroadNet Inference के सभी लाइव मॉडल: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL और दो स्पीच-टू-टेक्स्ट मॉडल, ID, कीमतों और कॉन्टेक्स्ट साइज़ के साथ। - [LLM API की कीमतें: अनलिमिटेड प्लान और प्रति-टोकन दरें](https://broadrouter.com/hi/pricing/) (Hindi; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/hi/pricing/index.md): 4 चैट मॉडल की प्रति-टोकन कीमतें, प्रति मिनट ऑडियो कीमतें और US$60 प्रति माह से DeepSeek सेशन प्लान, लागत अनुमान टूल के साथ। - [DeepSeek सेशन प्लान: Full Moon, Orbit और Meteor](https://broadrouter.com/hi/plans/) (Hindi; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/hi/plans/index.md): महीने, हफ़्ते या घंटे के फ़िक्स्ड कीमत वाले DeepSeek V4.1 Flash सेशन, प्रति सेशन एक समय में एक अनुरोध। सेशन प्लान करें और उपयोग के अनुसार भुगतान से तुलना करें। - [DeepSeek V4.1 Flash API: अनलिमिटेड प्लान या प्रति टोकन](https://broadrouter.com/hi/models/deepseek-v4.1-flash/) (Hindi; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/hi/models/deepseek-v4.1-flash/index.md): OpenAI-कम्पैटिबल API से DeepSeek V4.1 Flash, 1M-टोकन कॉन्टेक्स्ट के साथ। इन्फ़रेंस कंटेंट रखा नहीं जाता। प्रति टोकन भुगतान करें या फ़िक्स्ड अनलिमिटेड प्लान लें। - [GLM 5.3 API: कीमत, कॉन्टेक्स्ट और OpenAI-कम्पैटिबल सेटअप](https://broadrouter.com/hi/models/glm-5.3/) (Hindi; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/hi/models/glm-5.3/index.md): OpenAI-कम्पैटिबल एंडपॉइंट से GLM 5.3 इस्तेमाल करें: टूल, स्ट्रक्चर्ड आउटपुट, रीज़निंग, 1M-टोकन कॉन्टेक्स्ट और प्रति-टोकन कीमतें। इन्फ़रेंस कंटेंट रखा नहीं जाता। - [Qwen3.8 27B API: अरबी में टेस्टेड विज़न और टेक्स्ट मॉडल](https://broadrouter.com/hi/models/qwen3.8-27b/) (Hindi; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/hi/models/qwen3.8-27b/index.md): OpenAI-कम्पैटिबल API से Qwen3.8 27B: टेक्स्ट और विज़न, टूल, स्ट्रक्चर्ड आउटपुट, रीज़निंग, 262K कॉन्टेक्स्ट; अरबी और खाड़ी व मिस्री बोलियों में टेस्टेड। - [Gemma 4 26B विज़न API: सस्ता प्रति टोकन या घंटे वाली key](https://broadrouter.com/hi/models/gemma-4-vl-26b/) (Hindi; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/hi/models/gemma-4-vl-26b/index.md): Google के Gemma 4 26B का BroadNet बिल्ड: OpenAI-कम्पैटिबल API से टेक्स्ट और इमेज चैट, शुरुआती क्रेडिट पर या फ़िक्स्ड कीमत वाली घंटे की key के रूप में। - [अरबी स्पीच-टू-टेक्स्ट API, प्रति ऑडियो मिनट कीमत](https://broadrouter.com/hi/models/cohere-transcribe-arabic/) (Hindi; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/hi/models/cohere-transcribe-arabic/index.md): OpenAI-कम्पैटिबल ट्रांसक्रिप्शन एंडपॉइंट से अरबी ऑडियो, वॉइस नोट और कॉल ट्रांसक्राइब करें, प्रति ऑडियो मिनट कीमत पर। इन्फ़रेंस कंटेंट रखा नहीं जाता। - [14 भाषाओं में स्पीच-टू-टेक्स्ट API, प्रति मिनट कीमत](https://broadrouter.com/hi/models/cohere-transcribe/) (Hindi; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/hi/models/cohere-transcribe/index.md): अंग्रेज़ी, स्पैनिश, फ़्रेंच, चीनी और अरबी सहित 14 भाषाओं में ऑडियो ट्रांसक्राइब करें, OpenAI-कम्पैटिबल एंडपॉइंट से, प्रति ऑडियो मिनट कीमत पर। ### Indonesian (id, Bahasa Indonesia): 14 pages, hreflang id - [BroadRouter: API DeepSeek, GLM, Qwen, Gemma dan suara Arab](https://broadrouter.com/id/) (Indonesian; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/id/index.md): Satu API kompatibel OpenAI untuk DeepSeek, GLM, Qwen, Gemma, dan speech-to-text bahasa Arab. Bayar sesuai pemakaian, atau beli sesi DeepSeek harga tetap. - [Panduan cepat BroadNet Inference: API dengan Python dan cURL](https://broadrouter.com/id/developers/) (Indonesian; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/id/developers/index.md): Kirim permintaan chat dan transkripsi pertama Anda ke BroadNet Inference dengan Python atau cURL, lalu ikuti checklist migrasi untuk model pilihan Anda. - [Hubungi BroadNet tentang BroadNet Inference](https://broadrouter.com/id/contact/) (Indonesian; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/id/contact/index.md): Cara menghubungi BroadNet tentang BroadNet Inference: email umum, dukungan teknis, penjualan, serta endpoint privat, pemrosesan regional, atau model lain. - [API transkripsi bahasa Arab: model, format, harga per menit](https://broadrouter.com/id/arabic-transcription/) (Indonesian; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/id/arabic-transcription/index.md): Transkripsi audio bahasa Arab lewat BroadNet Inference: perbandingan model Arab dan multibahasa, format, batas ukuran file, estimasi biaya, dan kode siap salin. - [Ringkasan privasi: BroadNet Inference dan situs ini](https://broadrouter.com/id/privacy/) (Indonesian; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/id/privacy/index.md): Ringkasan singkat cara BroadNet Inference menangani prompt, audio, dan jawaban Anda, metadata apa yang disimpan, serta tautan ke kebijakan resmi BroadNet. - [Model LLM di BroadNet Inference: DeepSeek, GLM, Qwen, Gemma](https://broadrouter.com/id/models/) (Indonesian; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/id/models/index.md): Semua model aktif di BroadNet Inference: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL, dan dua model speech-to-text, dengan ID, harga, dan konteks. - [Harga API LLM: paket unlimited dan tarif per token](https://broadrouter.com/id/pricing/) (Indonesian; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/id/pricing/index.md): Harga per token untuk 4 model chat, harga audio per menit, dan paket sesi DeepSeek mulai US$60 per bulan, lengkap dengan kalkulator biaya. - [Paket sesi DeepSeek: Full Moon, Orbit, dan Meteor](https://broadrouter.com/id/plans/) (Indonesian; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/id/plans/index.md): Sesi DeepSeek V4.1 Flash harga tetap per bulan, minggu, atau jam, satu permintaan dalam satu waktu per sesi. Rencanakan sesi, bandingkan dengan bayar per token. - [API DeepSeek V4.1 Flash: paket unlimited atau per token](https://broadrouter.com/id/models/deepseek-v4.1-flash/) (Indonesian; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/id/models/deepseek-v4.1-flash/index.md): Gunakan DeepSeek V4.1 Flash dengan konteks 1 juta token lewat API kompatibel OpenAI; konten inferensi tidak disimpan. Bayar per token atau paket tetap. - [API GLM 5.3: harga, konteks, dan setup kompatibel OpenAI](https://broadrouter.com/id/models/glm-5.3/) (Indonesian; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/id/models/glm-5.3/index.md): Gunakan GLM 5.3 lewat endpoint kompatibel OpenAI: tool, output terstruktur, penalaran, konteks 1 juta token, dan harga per token. Konten tidak disimpan. - [API Qwen3.8 27B: teks dan gambar, diuji dalam bahasa Arab](https://broadrouter.com/id/models/qwen3.8-27b/) (Indonesian; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/id/models/qwen3.8-27b/index.md): Qwen3.8 27B, API kompatibel OpenAI: teks dan gambar, tool, output terstruktur, penalaran, konteks 262K. Diuji dalam bahasa Arab, dialek Teluk dan Mesir. - [API visi Gemma 4 26B: token murah atau kunci per jam](https://broadrouter.com/id/models/gemma-4-vl-26b/) (Indonesian; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/id/models/gemma-4-vl-26b/index.md): Versi BroadNet dari Gemma 4 26B milik Google: chat teks dan gambar lewat API yang kompatibel dengan OpenAI, dengan kredit awal atau kunci per jam harga tetap. - [API transkripsi audio bahasa Arab: harga per menit](https://broadrouter.com/id/models/cohere-transcribe-arabic/) (Indonesian; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/id/models/cohere-transcribe-arabic/index.md): Ubah audio, voice note, dan panggilan berbahasa Arab menjadi teks lewat endpoint transkripsi kompatibel OpenAI; harga per menit, konten tidak disimpan. - [API speech-to-text 14 bahasa, harga per menit audio](https://broadrouter.com/id/models/cohere-transcribe/) (Indonesian; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/id/models/cohere-transcribe/index.md): Ubah audio dalam 14 bahasa, termasuk Inggris, Spanyol, Prancis, Mandarin, dan Arab, menjadi teks lewat endpoint kompatibel OpenAI dengan harga per menit. ### Filipino (fil, Filipino): 14 pages, hreflang fil - [BroadRouter: DeepSeek, GLM, Qwen, Gemma at Arabic speech API](https://broadrouter.com/fil/) (Filipino; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/fil/index.md): Iisang OpenAI-compatible API para sa DeepSeek, GLM, Qwen, Gemma at Arabic speech-to-text. Pay-as-you-go, o flat-price DeepSeek session para sa agents mo. - [Quickstart ng BroadNet Inference API: Python at cURL](https://broadrouter.com/fil/developers/) (Filipino; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/fil/developers/index.md): Ipadala ang una mong chat at transcription request sa BroadNet Inference gamit ang Python o cURL, saka sundan ang migration checklist para sa model mo. - [Makipag-ugnayan sa BroadNet tungkol sa BroadNet Inference](https://broadrouter.com/fil/contact/) (Filipino; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/fil/contact/index.md): Paano kontakin ang BroadNet tungkol sa BroadNet Inference: email, technical support, sales, at para sa private endpoint, regional processing o ibang model. - [Arabic transcription API: mga model, format at presyo](https://broadrouter.com/fil/arabic-transcription/) (Filipino; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/fil/arabic-transcription/index.md): I-transcribe ang Arabic na audio sa BroadNet Inference: Arabic vs multilingual na model, mga format at size limit, cost estimator at code na makokopya. - [Buod ng privacy: BroadNet Inference at ang website na ito](https://broadrouter.com/fil/privacy/) (Filipino; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/fil/privacy/index.md): Maikling buod kung paano hinahawakan ng BroadNet Inference ang mga prompt, audio at sagot mo, anong metadata ang itinatago, at mga link sa opisyal na patakaran. - [Mga model sa BroadNet Inference: DeepSeek, GLM, Qwen, Gemma](https://broadrouter.com/fil/models/) (Filipino; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/fil/models/index.md): Lahat ng live na model sa BroadNet Inference: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL at dalawang speech-to-text model, may ID, presyo at context. - [Presyo ng LLM API: unlimited na plan at per-token na rate](https://broadrouter.com/fil/pricing/) (Filipino; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/fil/pricing/index.md): Magkano ang LLM API? Per-token na presyo ng 4 chat model, per-minute na presyo ng audio at DeepSeek session plan mula US$60 kada buwan, may cost estimator. - [DeepSeek session plan: Full Moon, Orbit at Meteor](https://broadrouter.com/fil/plans/) (Filipino; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/fil/plans/index.md): Flat-price na DeepSeek V4.1 Flash session kada buwan, linggo o oras, isang request sa bawat pagkakataon. Planuhin ang session at ikumpara sa pay-as-you-go. - [DeepSeek V4.1 Flash API: unlimited na plan o kada token](https://broadrouter.com/fil/models/deepseek-v4.1-flash/) (Filipino; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/fil/models/deepseek-v4.1-flash/index.md): Gamitin ang DeepSeek V4.1 Flash sa OpenAI-compatible API na may 1M-token na context. Hindi itinatago ang inference content. Kada token o unlimited na flat plan. - [GLM 5.3 API: presyo, context at OpenAI-compatible na setup](https://broadrouter.com/fil/models/glm-5.3/) (Filipino; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/fil/models/glm-5.3/index.md): Gamitin ang GLM 5.3 sa OpenAI-compatible endpoint: tools, structured output, reasoning, 1M-token na context at per-token na presyo. Hindi itinatago ang content. - [Qwen3.8 27B API: vision at text model na sinubok sa Arabic](https://broadrouter.com/fil/models/qwen3.8-27b/) (Filipino; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/fil/models/qwen3.8-27b/index.md): Qwen3.8 27B sa OpenAI-compatible API: text at vision, tools, structured output at reasoning, 262K context, sinubok sa Arabic at sa Gulf at Egyptian dialect. - [Gemma 4 26B vision API: murang per token o hourly key](https://broadrouter.com/fil/models/gemma-4-vl-26b/) (Filipino; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/fil/models/gemma-4-vl-26b/index.md): Ang build ng BroadNet ng Gemma 4 26B ng Google: text at image chat sa OpenAI-compatible API, magagamit sa panimulang credit o bilang flat-price na hourly key. - [Arabic speech-to-text API, may presyo kada minuto ng audio](https://broadrouter.com/fil/models/cohere-transcribe-arabic/) (Filipino; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/fil/models/cohere-transcribe-arabic/index.md): I-transcribe ang Arabic na audio, voice note at tawag sa OpenAI-compatible na transcription endpoint, may presyo kada minuto. Hindi itinatago ang content. - [Speech-to-text API sa 14 na wika, may presyo kada minuto](https://broadrouter.com/fil/models/cohere-transcribe/) (Filipino; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/fil/models/cohere-transcribe/index.md): I-transcribe ang audio sa 14 na wika, kasama ang Ingles, Espanyol, Pranses, Chinese at Arabic, sa OpenAI-compatible endpoint na may presyo kada minuto ng audio. ### Malay (ms, Bahasa Melayu): 14 pages, hreflang ms - [BroadRouter: API DeepSeek, GLM, Qwen, Gemma dan pertuturan Arab](https://broadrouter.com/ms/) (Malay; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/ms/index.md): Satu API serasi OpenAI untuk DeepSeek, GLM, Qwen, Gemma dan speech-to-text bahasa Arab. Bayar ikut guna, atau beli sesi DeepSeek harga tetap untuk ejen anda. - [Mula pantas BroadNet Inference API: Python dan cURL](https://broadrouter.com/ms/developers/) (Malay; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/ms/developers/index.md): Hantar permintaan chat dan transkripsi pertama anda ke BroadNet Inference dengan Python atau cURL, kemudian ikuti senarai semak migrasi untuk model anda. - [Hubungi BroadNet mengenai BroadNet Inference](https://broadrouter.com/ms/contact/) (Malay; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/ms/contact/index.md): Cara menghubungi BroadNet tentang BroadNet Inference: e-mel umum, sokongan teknikal, jualan, dan saluran untuk endpoint persendirian atau model lain. - [API transkripsi bahasa Arab: model, format dan kos seminit](https://broadrouter.com/ms/arabic-transcription/) (Malay; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/ms/arabic-transcription/index.md): Transkripsikan audio bahasa Arab melalui BroadNet Inference: perbandingan model Arab dan berbilang bahasa, format, had saiz, penganggar kos dan kod sedia salin. - [Ringkasan privasi: BroadNet Inference dan laman web ini](https://broadrouter.com/ms/privacy/) (Malay; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/ms/privacy/index.md): Ringkasan cara BroadNet Inference mengendalikan prompt, audio dan jawapan anda, metadata yang disimpannya, serta pautan ke dasar rasmi BroadNet. - [Model di BroadNet Inference: DeepSeek, GLM, Qwen, Gemma](https://broadrouter.com/ms/models/) (Malay; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/ms/models/index.md): Semua model aktif di BroadNet Inference: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL dan dua model speech-to-text, dengan ID, harga dan saiz konteks. - [Harga API LLM: pelan tanpa had dan kadar setiap token](https://broadrouter.com/ms/pricing/) (Malay; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/ms/pricing/index.md): Harga setiap token untuk 4 model chat, harga audio seminit dan pelan sesi DeepSeek dari US$60 sebulan, dengan penganggar kos. API serasi OpenAI. - [Pelan sesi DeepSeek: Full Moon, Orbit dan Meteor](https://broadrouter.com/ms/plans/) (Malay; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/ms/plans/index.md): Sesi DeepSeek V4.1 Flash harga tetap secara bulanan, mingguan atau sejam, satu permintaan pada satu masa. Rancang sesi dan bandingkan dengan bayar ikut guna. - [API DeepSeek V4.1 Flash: pelan tanpa had atau setiap token](https://broadrouter.com/ms/models/deepseek-v4.1-flash/) (Malay; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/ms/models/deepseek-v4.1-flash/index.md): Guna DeepSeek V4.1 Flash melalui API serasi OpenAI dengan konteks 1M token. Kandungan inferens tidak disimpan. Bayar setiap token atau beli pelan tanpa had. - [API GLM 5.3: harga, konteks dan persediaan serasi OpenAI](https://broadrouter.com/ms/models/glm-5.3/) (Malay; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/ms/models/glm-5.3/index.md): Guna GLM 5.3 melalui endpoint serasi OpenAI: alatan, output berstruktur, penaakulan, konteks 1M token dan harga setiap token. Kandungan inferens tidak disimpan. - [API Qwen3.8 27B: model teks dan imej, diuji bahasa Arab](https://broadrouter.com/ms/models/qwen3.8-27b/) (Malay; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/ms/models/qwen3.8-27b/index.md): Qwen3.8 27B melalui API serasi OpenAI: teks dan imej, alatan, output berstruktur, penaakulan, konteks 262K, diuji dalam bahasa Arab, dialek Teluk dan Mesir. - [API penglihatan Gemma 4 26B: murah setiap token atau sejam](https://broadrouter.com/ms/models/gemma-4-vl-26b/) (Malay; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/ms/models/gemma-4-vl-26b/index.md): Binaan BroadNet bagi Gemma 4 26B daripada Google: chat teks dan imej melalui API serasi OpenAI, boleh guna kredit permulaan atau kunci sejam harga tetap. - [API speech-to-text bahasa Arab, harga seminit audio](https://broadrouter.com/ms/models/cohere-transcribe-arabic/) (Malay; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/ms/models/cohere-transcribe-arabic/index.md): Transkripsikan audio, nota suara dan panggilan bahasa Arab melalui endpoint transkripsi serasi OpenAI, harga seminit audio. Kandungan inferens tidak disimpan. - [API speech-to-text dalam 14 bahasa, harga seminit](https://broadrouter.com/ms/models/cohere-transcribe/) (Malay; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/ms/models/cohere-transcribe/index.md): Transkripsikan audio dalam 14 bahasa, termasuk Inggeris, Sepanyol, Perancis, Cina dan Arab, melalui endpoint serasi OpenAI dengan harga seminit audio. ### Japanese (ja, 日本語): 14 pages, hreflang ja, Japanese script - [BroadRouter:DeepSeek・GLM・Qwen・Gemma とアラビア語音声の API](https://broadrouter.com/ja/) (Japanese; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/ja/index.md): DeepSeek、GLM、Qwen、Gemma とアラビア語の音声認識を、1 つの OpenAI 互換 API で。従量課金のほか、エージェント向けに定額の DeepSeek セッションも購入できます。 - [BroadNet Inference 開発者クイックスタート:Python と cURL](https://broadrouter.com/ja/developers/) (Japanese; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/ja/developers/index.md): Python または cURL で BroadNet Inference に最初のチャットと文字起こしのリクエストを送り、選んだモデルの移行チェックリストを確認します。 - [BroadNet Inference について BroadNet に問い合わせる](https://broadrouter.com/ja/contact/) (Japanese; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/ja/contact/index.md): BroadNet Inference についての BroadNet への連絡方法:一般窓口、技術サポート、営業、専用エンドポイントや地域処理、一覧にないモデルの相談窓口。 - [アラビア語の音声認識・文字起こし API:モデル、形式、1 分あたりの料金](https://broadrouter.com/ja/arabic-transcription/) (Japanese; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/ja/arabic-transcription/index.md): BroadNet Inference でアラビア語の音声を文字起こし:アラビア語モデルと多言語モデルの比較、対応形式とサイズ上限、コストの見積もり、コピーして使えるコード。 - [プライバシーの要約:BroadNet Inference とこのサイト](https://broadrouter.com/ja/privacy/) (Japanese; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/ja/privacy/index.md): BroadNet Inference がプロンプト、音声、回答をどう扱い、どのメタデータを保持するかの短い要約と、BroadNet の公式ポリシーへのリンク。 - [BroadNet Inference のモデル:DeepSeek・GLM・Qwen・Gemma API](https://broadrouter.com/ja/models/) (Japanese; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/ja/models/index.md): BroadNet Inference で提供中の全モデル:DeepSeek V4.1 Flash、GLM 5.3、Qwen3.8 27B、Gemma 4 VL と 2 つの音声認識モデルの ID、料金、コンテキスト長。 - [LLM API 料金:定額の無制限プランとトークン単価](https://broadrouter.com/ja/pricing/) (Japanese; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/ja/pricing/index.md): チャットモデル 4 個のトークン単価、音声の 1 分あたりの料金、月 US$60 からの DeepSeek セッションプランをまとめ、コスト見積もりツールも用意しています。 - [DeepSeek セッションプラン:Full Moon、Orbit、Meteor](https://broadrouter.com/ja/plans/) (Japanese; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/ja/plans/index.md): 月・週・時間単位で購入できる定額の DeepSeek V4.1 Flash セッション。1 セッションにつき同時に 1 リクエスト。必要なセッション数を計算し、従量課金と比較できます。 - [DeepSeek V4.1 Flash API:無制限プランまたはトークン課金](https://broadrouter.com/ja/models/deepseek-v4.1-flash/) (Japanese; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/ja/models/deepseek-v4.1-flash/index.md): DeepSeek V4.1 Flash を OpenAI 互換 API で。コンテキスト 100 万トークン、推論の内容は保持されません。トークン課金か、定額の無制限プランを選べます。 - [GLM 5.3 API:料金、コンテキスト、OpenAI 互換の設定](https://broadrouter.com/ja/models/glm-5.3/) (Japanese; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/ja/models/glm-5.3/index.md): GLM 5.3 を OpenAI 互換エンドポイントで。ツール呼び出し、構造化出力、推論、コンテキスト 100 万トークン、推論の内容は保持せず、トークン単価は公開済み。 - [Qwen3.8 27B API:アラビア語でテスト済みのビジョン・テキストモデル](https://broadrouter.com/ja/models/qwen3.8-27b/) (Japanese; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/ja/models/qwen3.8-27b/index.md): Qwen3.8 27B を OpenAI 互換 API で。テキストとビジョン、ツール呼び出し、構造化出力、推論、コンテキスト 262K。アラビア語と湾岸・エジプト方言でテスト済み。 - [Gemma 4 26B ビジョン API:低いトークン単価か時間制キー](https://broadrouter.com/ja/models/gemma-4-vl-26b/) (Japanese; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/ja/models/gemma-4-vl-26b/index.md): BroadNet 版の Google Gemma 4 26B:OpenAI 互換 API でテキストと画像のチャット。初回クレジットか、定額の時間制キーで使えます。 - [アラビア語の音声認識 API:音声 1 分単位の料金](https://broadrouter.com/ja/models/cohere-transcribe-arabic/) (Japanese; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/ja/models/cohere-transcribe-arabic/index.md): アラビア語の音声、ボイスメッセージ、通話を OpenAI 互換の文字起こしエンドポイントでテキスト化。料金は音声 1 分単位で、推論の内容は保持されません。 - [14 言語対応の音声認識 API、1 分単位の料金](https://broadrouter.com/ja/models/cohere-transcribe/) (Japanese; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/ja/models/cohere-transcribe/index.md): 英語、スペイン語、フランス語、中国語、アラビア語を含む 14 言語の音声を、OpenAI 互換エンドポイントで文字起こし。料金は音声 1 分単位です。 ### Korean (ko, 한국어): 14 pages, hreflang ko, Hangul - [BroadRouter: DeepSeek, GLM, Qwen, Gemma, 아랍어 음성 API](https://broadrouter.com/ko/) (Korean; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/ko/index.md): DeepSeek, GLM, Qwen, Gemma와 아랍어 음성 인식을 OpenAI 호환 API 하나로. 사용한 만큼 결제하거나, 에이전트용 정액 DeepSeek 세션을 구매하세요. - [BroadNet Inference 개발자 빠른 시작: Python과 cURL](https://broadrouter.com/ko/developers/) (Korean; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/ko/developers/index.md): Python 또는 cURL로 BroadNet Inference에 첫 채팅 요청과 음성 전사 요청을 보내고, 선택한 모델의 마이그레이션 체크리스트를 확인하세요. - [BroadNet Inference 관련 BroadNet 문의](https://broadrouter.com/ko/contact/) (Korean; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/ko/contact/index.md): BroadNet Inference에 관해 BroadNet에 연락하는 방법: 일반 이메일, 기술 지원, 영업, 그리고 전용 엔드포인트, 지역별 처리, 목록에 없는 모델 문의 경로. - [아랍어 음성 인식(STT) API: 모델, 형식, 분당 가격](https://broadrouter.com/ko/arabic-transcription/) (Korean; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/ko/arabic-transcription/index.md): BroadNet Inference로 아랍어 오디오 전사: 아랍어 모델과 다국어 모델 비교, 지원 형식과 크기 한도, 비용 계산기, 바로 복사해 쓰는 코드. - [개인정보 요약: BroadNet Inference와 이 웹사이트](https://broadrouter.com/ko/privacy/) (Korean; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/ko/privacy/index.md): BroadNet Inference가 프롬프트, 오디오, 답변을 어떻게 처리하고 어떤 메타데이터를 보관하는지에 대한 짧은 요약과 BroadNet 공식 정책 링크. - [BroadNet Inference 모델: DeepSeek, GLM, Qwen, Gemma API](https://broadrouter.com/ko/models/) (Korean; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/ko/models/index.md): BroadNet Inference에서 제공 중인 모든 모델: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL과 음성 인식 모델 2개의 ID, 가격, 컨텍스트 크기. - [LLM API 가격: 무제한 정액 플랜과 토큰당 요금](https://broadrouter.com/ko/pricing/) (Korean; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/ko/pricing/index.md): 채팅 모델 4개의 토큰당 가격, 오디오 분당 가격, 월 US$60부터 시작하는 DeepSeek 세션 플랜과 비용 계산기를 한곳에서 확인하세요. - [DeepSeek 세션 플랜: Full Moon, Orbit, Meteor](https://broadrouter.com/ko/plans/) (Korean; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/ko/plans/index.md): 월·주·시간 단위로 구매하는 정액 DeepSeek V4.1 Flash 세션. 세션당 한 번에 요청 1개. 필요한 세션 수를 계산하고 종량제와 비교해 보세요. - [DeepSeek V4.1 Flash API: 무제한 플랜 또는 토큰당 결제](https://broadrouter.com/ko/models/deepseek-v4.1-flash/) (Korean; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/ko/models/deepseek-v4.1-flash/index.md): OpenAI 호환 API로 DeepSeek V4.1 Flash를 사용하세요. 컨텍스트 100만 토큰, 추론 콘텐츠 비보관. 토큰당 결제하거나 무제한 정액 플랜을 구매할 수 있습니다. - [GLM 5.3 API: 가격, 컨텍스트, OpenAI 호환 설정](https://broadrouter.com/ko/models/glm-5.3/) (Korean; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/ko/models/glm-5.3/index.md): OpenAI 호환 엔드포인트로 GLM 5.3을 사용하세요. 도구 호출, 구조화된 출력, 추론, 컨텍스트 100만 토큰, 추론 콘텐츠 비보관, 공개된 토큰당 가격. - [Qwen3.8 27B API: 아랍어로 테스트한 비전·텍스트 모델](https://broadrouter.com/ko/models/qwen3.8-27b/) (Korean; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/ko/models/qwen3.8-27b/index.md): OpenAI 호환 API로 쓰는 Qwen3.8 27B: 텍스트와 비전, 도구 호출, 구조화된 출력, 추론, 컨텍스트 262K. 아랍어와 걸프·이집트 방언으로 테스트했습니다. - [Gemma 4 26B 비전 API: 낮은 토큰 단가 또는 시간제 키](https://broadrouter.com/ko/models/gemma-4-vl-26b/) (Korean; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/ko/models/gemma-4-vl-26b/index.md): BroadNet 버전 Google Gemma 4 26B: OpenAI 호환 API로 텍스트와 이미지 채팅. 초기 크레딧이나 정액 시간제 키로 사용할 수 있습니다. - [아랍어 음성 인식(STT) API, 오디오 1분 단위 가격](https://broadrouter.com/ko/models/cohere-transcribe-arabic/) (Korean; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/ko/models/cohere-transcribe-arabic/index.md): OpenAI 호환 음성 전사 엔드포인트로 아랍어 오디오, 음성 메시지, 통화를 텍스트로 변환하세요. 오디오 1분 단위 가격, 추론 콘텐츠 비보관. - [14개 언어 음성 인식(STT) API, 분 단위 가격](https://broadrouter.com/ko/models/cohere-transcribe/) (Korean; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/ko/models/cohere-transcribe/index.md): 영어, 스페인어, 프랑스어, 중국어, 아랍어를 포함한 14개 언어의 오디오를 OpenAI 호환 엔드포인트로 전사하세요. 오디오 1분 단위 가격입니다. ### Russian (ru, Русский): 14 pages, hreflang ru, Cyrillic script - [BroadRouter: API DeepSeek, GLM, Qwen, Gemma и арабской речи](https://broadrouter.com/ru/) (Russian; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/ru/index.md): Единый OpenAI-совместимый API для DeepSeek, GLM, Qwen, Gemma и распознавания арабской речи. Оплата по факту или сессии DeepSeek по фиксированной цене. - [Быстрый старт BroadNet Inference: Python и cURL](https://broadrouter.com/ru/developers/) (Russian; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/ru/developers/index.md): Отправьте первый запрос к чату и распознаванию речи BroadNet Inference на Python или через cURL, затем пройдите чек-лист миграции для своей модели. - [Связаться с BroadNet по BroadNet Inference](https://broadrouter.com/ru/contact/) (Russian; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/ru/contact/index.md): Как связаться с BroadNet по BroadNet Inference: общая почта, техподдержка, отдел продаж и запрос приватного эндпоинта, региональной обработки или другой модели. - [API распознавания арабской речи: модели, форматы, цена](https://broadrouter.com/ru/arabic-transcription/) (Russian; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/ru/arabic-transcription/index.md): Транскрибация арабского аудио через BroadNet Inference: сравнение арабской и многоязычной моделей, форматы и лимит размера, калькулятор стоимости и готовый код. - [Кратко о конфиденциальности: BroadNet Inference и сайт](https://broadrouter.com/ru/privacy/) (Russian; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/ru/privacy/index.md): Кратко о том, как BroadNet Inference обращается с вашими промптами, аудио и ответами, какие метаданные хранит, и ссылки на официальные политики BroadNet. - [Модели BroadNet Inference: DeepSeek, GLM, Qwen, Gemma](https://broadrouter.com/ru/models/) (Russian; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/ru/models/index.md): Все модели BroadNet Inference: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL и две модели распознавания речи — с ID, ценами и размером контекста. - [Цены на LLM API: безлимитные тарифы и оплата за токены](https://broadrouter.com/ru/pricing/) (Russian; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/ru/pricing/index.md): Цены за токен для 4 чат-моделей, поминутные цены на распознавание речи и тарифы DeepSeek с сессиями от US$60 в месяц, с калькуляторами. - [Тарифы DeepSeek с сессиями: Full Moon, Orbit и Meteor](https://broadrouter.com/ru/plans/) (Russian; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/ru/plans/index.md): Сессии DeepSeek V4.1 Flash по фиксированной цене на месяц, неделю или час, один запрос за раз в каждой сессии. Спланируйте сессии и сравните с оплатой по факту. - [DeepSeek V4.1 Flash API: безлимит или оплата за токены](https://broadrouter.com/ru/models/deepseek-v4.1-flash/) (Russian; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/ru/models/deepseek-v4.1-flash/index.md): DeepSeek V4.1 Flash через OpenAI-совместимый API: контекст 1 млн токенов, контент не хранится. Оплата за токены или безлимитный тариф по фиксированной цене. - [GLM 5.3 API: цена, контекст и OpenAI-совместимая настройка](https://broadrouter.com/ru/models/glm-5.3/) (Russian; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/ru/models/glm-5.3/index.md): GLM 5.3 через OpenAI-совместимый эндпоинт: инструменты, структурированный вывод, рассуждения, контекст 1 млн токенов, контент не хранится, цены опубликованы. - [Qwen3.8 27B API: текст и vision, протестирована на арабском](https://broadrouter.com/ru/models/qwen3.8-27b/) (Russian; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/ru/models/qwen3.8-27b/index.md): Qwen3.8 27B через OpenAI-совместимый API: текст и изображения, инструменты, структурный вывод, рассуждения, 262K контекста. Тесты на арабском и диалектах. - [Gemma 4 26B vision API: низкая цена или почасовой ключ](https://broadrouter.com/ru/models/gemma-4-vl-26b/) (Russian; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/ru/models/gemma-4-vl-26b/index.md): Gemma 4 26B от Google в версии BroadNet: текст и изображения через OpenAI-совместимый API, на стартовом гранте или с почасовым ключом по фиксированной цене. - [API распознавания арабской речи с оплатой за минуту аудио](https://broadrouter.com/ru/models/cohere-transcribe-arabic/) (Russian; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/ru/models/cohere-transcribe-arabic/index.md): Транскрибация арабского аудио, голосовых и звонков через OpenAI-совместимый эндпоинт распознавания речи, с оплатой за минуту аудио; контент не хранится. - [Транскрибация на 14 языках: API с оплатой за минуту аудио](https://broadrouter.com/ru/models/cohere-transcribe/) (Russian; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/ru/models/cohere-transcribe/index.md): Транскрибация аудио на 14 языках, включая английский, испанский, французский, китайский и арабский, через OpenAI-совместимый эндпоинт с оплатой за минуту аудио. ### Vietnamese (vi, Tiếng Việt): 14 pages, hreflang vi - [BroadRouter: API DeepSeek, GLM, Qwen, Gemma, giọng nói Ả Rập](https://broadrouter.com/vi/) (Vietnamese; translation of https://broadrouter.com/; Markdown: https://broadrouter.com/vi/index.md): Một API tương thích OpenAI cho DeepSeek, GLM, Qwen, Gemma và chuyển giọng nói tiếng Ả Rập thành văn bản. Trả theo mức dùng hoặc mua phiên DeepSeek giá cố định. - [Hướng dẫn nhanh BroadNet Inference: Python và cURL](https://broadrouter.com/vi/developers/) (Vietnamese; translation of https://broadrouter.com/developers/; Markdown: https://broadrouter.com/vi/developers/index.md): Gửi request chat và chuyển giọng nói thành văn bản đầu tiên đến BroadNet Inference bằng Python hoặc cURL, rồi làm theo checklist chuyển đổi cho model của bạn. - [Liên hệ BroadNet về BroadNet Inference](https://broadrouter.com/vi/contact/) (Vietnamese; translation of https://broadrouter.com/contact/; Markdown: https://broadrouter.com/vi/contact/index.md): Cách liên hệ BroadNet về BroadNet Inference: email chung, hỗ trợ kỹ thuật, kinh doanh, hỏi về endpoint riêng, xử lý theo khu vực hoặc model chưa niêm yết. - [API chuyển giọng nói tiếng Ả Rập thành văn bản: model và giá](https://broadrouter.com/vi/arabic-transcription/) (Vietnamese; translation of https://broadrouter.com/arabic-transcription/; Markdown: https://broadrouter.com/vi/arabic-transcription/index.md): Chuyển âm thanh tiếng Ả Rập thành văn bản qua BroadNet Inference: so sánh model Ả Rập và đa ngôn ngữ, định dạng, giới hạn file, ước tính chi phí và code mẫu. - [Tóm tắt quyền riêng tư: BroadNet Inference và website này](https://broadrouter.com/vi/privacy/) (Vietnamese; translation of https://broadrouter.com/zero-data-retention/; Markdown: https://broadrouter.com/vi/privacy/index.md): Tóm tắt cách BroadNet Inference xử lý prompt, âm thanh và câu trả lời của bạn, metadata nào được lưu, kèm liên kết đến chính sách chính thức của BroadNet. - [Model trên BroadNet Inference: DeepSeek, GLM, Qwen, Gemma](https://broadrouter.com/vi/models/) (Vietnamese; translation of https://broadrouter.com/models/; Markdown: https://broadrouter.com/vi/models/index.md): Mọi model đang chạy trên BroadNet Inference: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B, Gemma 4 VL và hai model giọng nói, kèm ID, giá và kích thước context. - [Giá API LLM: gói không giới hạn và giá theo token](https://broadrouter.com/vi/pricing/) (Vietnamese; translation of https://broadrouter.com/pricing/; Markdown: https://broadrouter.com/vi/pricing/index.md): Giá theo token cho 4 model chat, giá âm thanh theo phút và gói phiên DeepSeek từ US$60 mỗi tháng, kèm công cụ ước tính chi phí. - [Gói phiên DeepSeek: Full Moon, Orbit và Meteor](https://broadrouter.com/vi/plans/) (Vietnamese; translation of https://broadrouter.com/plans/; Markdown: https://broadrouter.com/vi/plans/index.md): Phiên DeepSeek V4.1 Flash giá cố định theo tháng, tuần hoặc giờ, mỗi phiên một request tại một thời điểm. Lập kế hoạch phiên và so sánh với trả theo mức dùng. - [API DeepSeek V4.1 Flash: không giới hạn hoặc trả theo token](https://broadrouter.com/vi/models/deepseek-v4.1-flash/) (Vietnamese; translation of https://broadrouter.com/models/deepseek-v4.1-flash/; Markdown: https://broadrouter.com/vi/models/deepseek-v4.1-flash/index.md): Dùng DeepSeek V4.1 Flash qua API tương thích OpenAI, context 1 triệu token, không lưu nội dung. Trả theo token hoặc mua gói không giới hạn giá cố định. - [API GLM 5.3: giá, context và cấu hình tương thích OpenAI](https://broadrouter.com/vi/models/glm-5.3/) (Vietnamese; translation of https://broadrouter.com/models/glm-5.3/; Markdown: https://broadrouter.com/vi/models/glm-5.3/index.md): Dùng GLM 5.3 qua endpoint tương thích OpenAI: công cụ, đầu ra có cấu trúc, suy luận, context 1 triệu token, không lưu nội dung, giá theo token công khai. - [API Qwen3.8 27B: model văn bản và vision thử với tiếng Ả Rập](https://broadrouter.com/vi/models/qwen3.8-27b/) (Vietnamese; translation of https://broadrouter.com/models/qwen3.8-27b/; Markdown: https://broadrouter.com/vi/models/qwen3.8-27b/index.md): Qwen3.8 27B qua API tương thích OpenAI: văn bản và hình ảnh, công cụ, đầu ra có cấu trúc, suy luận, context 262K. Đã thử với tiếng Ả Rập và phương ngữ. - [API vision Gemma 4 26B: giá token thấp hoặc key theo giờ](https://broadrouter.com/vi/models/gemma-4-vl-26b/) (Vietnamese; translation of https://broadrouter.com/models/gemma-4-vl-26b/; Markdown: https://broadrouter.com/vi/models/gemma-4-vl-26b/index.md): Bản Gemma 4 26B của Google do BroadNet vận hành: chat văn bản và hình ảnh qua API tương thích OpenAI, dùng tín dụng ban đầu hoặc key theo giờ giá cố định. - [API chuyển giọng nói Ả Rập thành văn bản, giá theo phút](https://broadrouter.com/vi/models/cohere-transcribe-arabic/) (Vietnamese; translation of https://broadrouter.com/models/cohere-transcribe-arabic/; Markdown: https://broadrouter.com/vi/models/cohere-transcribe-arabic/index.md): Chuyển âm thanh, tin nhắn thoại và cuộc gọi tiếng Ả Rập thành văn bản qua endpoint tương thích OpenAI, giá theo phút âm thanh và không lưu nội dung. - [API giọng nói thành văn bản 14 ngôn ngữ, giá theo phút](https://broadrouter.com/vi/models/cohere-transcribe/) (Vietnamese; translation of https://broadrouter.com/models/cohere-transcribe/; Markdown: https://broadrouter.com/vi/models/cohere-transcribe/index.md): Chuyển âm thanh 14 ngôn ngữ thành văn bản, gồm tiếng Anh, Tây Ban Nha, Pháp, Trung và Ả Rập, qua endpoint tương thích OpenAI, giá theo phút âm thanh. ## Translated articles, guides and glossary: 17 languages, 4998 pages Translated from English with AI assistance; each page links its English original, which is authoritative and may be newer. Prices, model IDs, code and URLs are the same as in English. Every page has a Markdown twin (append index.md). ### Arabic (ar, العربية): 294 translated pages - [Articles in Arabic](https://broadrouter.com/ar/blog/) - [تغيّرات أسعار LLM API في أكتوبر 2026: 3 ارتفاعات ولا تخفيض](https://broadrouter.com/ar/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 تغييرًا في صفحات الأسعار لدى 5 من 7 مزوّدي LLM API في أكتوبر 2026: 3 ارتفاعات، ولا تخفيض، و9 نماذج بأسعار جديدة، و11 أُزيلت. مع روابط الأرشيف لكلٍّ منها. - [كم يكلّف تشغيل واحد لوكيل برمجة: الخطوات ونمو السياق والتخزين المؤقت](https://broadrouter.com/ar/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): لماذا ترتفع تكلفة الخطوة في تشغيل وكيل البرمجة كلما طال، وكيف يغيّر التخزين المؤقت المجموع، ومتى تتفوق الجلسة الثابتة على الدفع لكل رمز. حساب مفصّل للرموز. - [تشكيل النص العربي بنموذج LLM: الأمر والتحقق ومعدل DER](https://broadrouter.com/ar/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): أعِد التشكيل إلى النص العربي عبر API لنموذج LLM: أمر يحفظ كل حرف، وتحقق يرفض أي تعديل، وقياس بمعدلي DER وWER، ومتى تختار CATT بدلًا منه. - [التدقيق الإملائي والنحوي للنص العربي بنموذج LLM: دليل عملي](https://broadrouter.com/ar/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): صحّح الأخطاء الإملائية والنحوية في النص العربي عبر API لنموذج LLM: أمر بأقل تعديل، وحماية من إعادة الصياغة، وتصنيف التعديلات، وقياس بأسلوب M2. - [محضر اجتماع عربي من التسجيلات: مسار عمل عبر API](https://broadrouter.com/ar/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): حوّل تسجيلات الاجتماعات العربية إلى محاضر فيها القرارات والمهام: فرّغ الصوت نصيًا، ونظّفه، ولخّصه بـLLM إلى JSON، ثم تحقّق من الأسماء والتواريخ. - [استخراج الكيانات المسماة من النص العربي بنموذج LLM: JSON والتواريخ](https://broadrouter.com/ar/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): استخرج الأسماء والمؤسسات والأماكن والتواريخ الهجرية من النص العربي إلى JSON بنموذج LLM: أمر، وفحص مقاطع يتجاهل التشكيل، وتحويل التواريخ، وشيفرة للتقييم. - [مقارنة واجهات OCR العربي: Google وAzure وTextract وMistral ونماذج LLM](https://broadrouter.com/ar/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): أي واجهات OCR تدعم العربية رسميًا، المطبوعة والمكتوبة بخط اليد، وماذا تعيد كلٌّ منها، وكم تكلّف لكل 1,000 صفحة، مقارنةً بواجهات نماذج اللغة البصرية. - [ترجمة نصية عربية (SRT) من API لتحويل الصوت إلى نص: ما تحتاجه](https://broadrouter.com/ar/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): ما يحتاجه ملف SRT، وأي واجهات API لتحويل الصوت إلى نص تعيد طوابع زمنية أو SRT للعربية، ومسار عمل مُختبر يبني الترجمة النصية من API تعيد نصًا فقط. - [تصنيف النصوص العربية بنموذج LLM: المشاعر والنية وفرز التذاكر](https://broadrouter.com/ar/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): صنّف المراجعات والرسائل وتذاكر الدعم العربية عبر API لنموذج LLM: تصميم الفئات، وأمر يراعي اللهجات، ومخرجات JSON، واختبار macro F1، والتكلفة عند الحجم الكبير. - [توحيد النص العربي للبحث وRAG: ما الذي تغيّره وما الذي تتركه](https://broadrouter.com/ar/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): التشكيل، وصور الألف، والتاء المربوطة، وكراسي الهمزة، والحروف الفارسية، والأرقام، وUnicode: أي قواعد توحيد النص العربي تفيد البحث، وأيها تضر، مع شيفرة Python مُختبرة. - [تلخيص النصوص العربية بنماذج LLM: الأوامر والطول والتقييم](https://broadrouter.com/ar/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): لخّص المقالات والتقارير ورسائل اللهجات العربية عبر LLM API: صيغ الأوامر، وضبط عدد الكلمات، وفحص الأرقام، ولماذا يعطي ROUGE الافتراضي للعربية صفرًا. - [بدائل Artificial Analysis: أين تقارن نماذج LLM في 2026](https://broadrouter.com/ar/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): مقارنة بدائل Artificial Analysis: Arena وEpoch AI وVals AI وتصنيفات OpenRouter وHugging Face وبيانات الأسعار، حسب ما يقيسه كلٌّ منها ومتى تستخدمه. - [عتاد GPU لدى BroadNet وطريقة تقديم النماذج، كما تصفها BroadNet](https://broadrouter.com/ar/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): ما تقوله BroadNet عن العتاد وراء BroadNet Inference: ثلاث طبقات من وحدات GPU من B300 إلى أجهزة لدى العملاء، والإنتاجية لكل GPU، وما تعلّمته من اختباراتها. - [تخزين استجابات LLM API مؤقتًا لخفض التكلفة: تطابق تام بـ Python](https://broadrouter.com/ar/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): ذاكرة مؤقتة مُختبرة للتطابق التام لاستجابات LLM API بـ Python: ما يدخل في المفتاح، وما لا يُخزَّن أبدًا، ومدد الصلاحية، والخصوصية، وقياس نسبة الإصابة قبل البناء. - [Claude Code مع النماذج المفتوحة: الوكلاء ونقاط النهاية الأصلية والحدود](https://broadrouter.com/ar/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): يتحدث Claude Code صيغة Anthropic Messages API. كيف تشغّله على نماذج مفتوحة عبر وكيل LiteLLM أو نقطة نهاية أصلية، وما الذي يتعطل في الطريق. - [المراجعة البشرية جزء من سير عمل الذكاء الاصطناعي للمستندات](https://broadrouter.com/ar/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): صمّم خطوة المراجعة البشرية في خط استخراج المستندات بالذكاء الاصطناعي: قواعد التحقق، وما يُحال إلى شخص، وعينات تدقيق عشوائية، وسجل تدقيق، ونسبة المراجعة. - [قانون حماية البيانات الشخصية المصري (151 لسنة 2020) وLLM API](https://broadrouter.com/ar/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): إرسال بيانات شخصية لمصريين إلى LLM API أو API لتحويل الصوت إلى نص في الخارج: ما يقوله القانون 151 لسنة 2020 والقرار 816 لسنة 2025 عن النقل والتراخيص والخرق. - [قبل الضبط الدقيق: حدّد المشكلة التي تريد حلّها](https://broadrouter.com/ar/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): متى تحلّ هندسة الأوامر أو RAG أو التقييم المشكلة فلا يلزم الضبط الدقيق، وما بيانات التدريب التي تحتاجها، وكيف تختبر نموذجًا مضبوطًا قبل أن تطلبه. - [إصلاح JSON غير الصالح من LLM: الأسباب والإصلاح وإعادة المحاولة](https://broadrouter.com/ar/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): لماذا تُعيد نماذج اللغة JSON معطوبًا (أسوار الشيفرة، والنص المحيط، والقطع عند max_tokens، والفواصل الزائدة)، ودالة Python مُختبرة تصلح الآمن وترفض الباقي. - [كم مرة تتغيّر أسعار LLM API؟ 3 سنوات من صفحات الأسعار المؤرشفة](https://broadrouter.com/ar/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): قرأنا 308 صفحات أسعار مؤرشفة لدى 11 مزوّدًا لـ LLM API من 2023 إلى 2026. لم يتغيّر سعر 80% من النماذج قط، وأجرت DeepInfra نسبة 75% من كل تغييرات الأسعار المُعلنة. - [كيف تخفّض تكلفة LLM API؟ قائمة من 10 خطوات مرتّبة](https://broadrouter.com/ar/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): خفّض فاتورة LLM API بالترتيب الصحيح: القياس، وتخزين البادئة مؤقتًا، واختصار الأوامر، وتقييد الإخراج، واختيار النموذج المناسب، ومتى تربح الخطة الثابتة. - [التحويل التلقائي في LLM API: الانتقال إلى مزوّد احتياطي في شيفرتك](https://broadrouter.com/ar/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): كيف تنفّذ التحويل التلقائي بين مزوّدي LLM المتوافقين مع OpenAI: أي الأخطاء تستدعيه، وقاطع الدائرة، وربط معرّفات النماذج، والبث المتدفق، وتكافؤ التكلفة والخصوصية. - [تخفيضات أسعار LLM API من 2023 إلى 2026: أكبر الانخفاضات بالمصادر](https://broadrouter.com/ar/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): أكبر تخفيضات أسعار LLM API من 2023 إلى 2026، لكلٍّ منها رابطا الأرشيف قبل التغيير وبعده: o3 وCodestral وGemini 1.5، وارتفاعات قلّما انتبه لها أحد. - [هل تنخفض أسعار LLM API فقط؟ 52 ارتفاعًا وكيف تخطط ميزانيتك لها](https://broadrouter.com/ar/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): نعم، ترتفع أسعار LLM API. وجدنا 52 ارتفاعًا في السعر المُعلن للنموذج نفسه في صفحات أسعار مؤرشفة من 2024 إلى 2026، بوسيط 40%، مع كل المصادر وقاعدة للميزانية. - [حدود المعدل في LLM API حسب المزود: مقارنة RPM وTPM والتزامن](https://broadrouter.com/ar/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): حدود المعدل في الفئة الأولى لـ LLM API لدى 13 مزودًا من وثائقهم: OpenAI وAnthropic وGemini وGroq وDeepSeek وFireworks وغيرهم، مع شرح RPM وTPM. - [مقارنة SLA لـ LLM API: بماذا تتعهد Azure وBedrock وGoogle وOpenAI؟](https://broadrouter.com/ar/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): بماذا يتعهد كل مزوّد LLM API كتابيًا: نسبة الإتاحة، وما يُعدّ انقطاعًا، وأرصدة الخدمة، وأهداف زمن الاستجابة، كما قرأناها من صفحات اتفاقيات مستوى الخدمة. - [معالجة دفعات LLM دون Batch API: مشغّل قابل للاستئناف](https://broadrouter.com/ar/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): شغّل من 10,000 إلى 1,000,000 أمر عبر API لإكمال المحادثات دون Batch API: حدود التزامن، وضبط المعدل، وإعادة المحاولة، ونقاط حفظ للاستئناف، وحساب الوقت. - [خريطة الأمان: كيف يتعامل 39 نموذج LLM مع البيانات الحساسة](https://broadrouter.com/ar/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): دراسة من BroadNet شملت 39 نموذج LLM من 14 مختبرًا في بيانات الاعتماد والبيانات الشخصية والصحية والمالية: ما انتقل وما لم ينتقل، وتسرّبات استدعاء الأدوات والاستدلال. - [كيف تختبر أمان نموذج LLM مع البيانات الحساسة: طريقة BroadNet](https://broadrouter.com/ar/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): طريقة BroadNet المنشورة لاختبار تسريب نموذج LLM لبيانات الاعتماد أو الشخصية أو الصحية أو المالية قبل النشر: الفئات، والتسريب مقابل الإغفال، و3 تشغيلات أو أكثر. - [الانتقال من OpenAI API: شرح عملي للشيفرة بـ Python وNode وcURL](https://broadrouter.com/ar/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): تعديلات الشيفرة للانتقال من OpenAI إلى واجهة API متوافقة معها: إعداد العميل، ومن Responses إلى Chat Completions، والأدوات، والبث المتدفق، والتضمينات، واختبار سريع. - [النماذج التي قيّمتها BroadNet واستبعدتها، ولماذا](https://broadrouter.com/ar/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): تنشر BroadNet النماذج التي اختبرتها ورفضتها مع الأسباب: نماذج محادثة رسبت في فحص أمان بيانات الاعتماد، ونماذج كلام وOCR وتضمين خسرت المقارنة. - [كيف تصمّم منتج ذكاء اصطناعي متعدد اللغات](https://broadrouter.com/ar/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): صمّم منتج ذكاء اصطناعي للعربية والإنجليزية والإسبانية: لغة الواجهة مقابل لغة النموذج، والاتجاه من اليمين لليسار، والأرقام والتواريخ، وكشف اللغة والتكلفة. - [تكلفة API النماذج المفتوحة مقابل المغلقة: ماذا تقول بيانات الأسعار؟](https://broadrouter.com/ar/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): هل النماذج مفتوحة الأوزان أرخص عبر API من GPT وClaude وGemini؟ أسعار لكل رمز (توكن) من صفحات المزودين، والتفاوت بين المزودين، وتكلفة محسوبة لكل مستخدم. - [تاريخ أسعار API لدى OpenAI وAnthropic وGoogle من 2023 إلى 2026](https://broadrouter.com/ar/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): أسعار GPT وClaude وGemini API جنبًا إلى جنب من صفحات أسعار مؤرشفة من 2023 إلى 2026: أيّ السلاسل صارت أرخص وأيّها صارت أغلى، ولكلّ سعر رابط في الأرشيف. - [عناوين base URL لواجهات API المتوافقة مع OpenAI: فحص 14 مزوّدًا](https://broadrouter.com/ar/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): قيمة base_url لـ 14 واجهة LLM API متوافقة مع OpenAI، مأخوذة من وثائق كل مزوّد، مع نقاط النهاية المدرجة والفروق التي تُفشل التبديل المباشر. - [OpenClaw وHermes Agent على نموذج رخيص: الإعداد وحساب الرموز](https://broadrouter.com/ar/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): اربط OpenClaw أو Hermes Agent بنقطة نهاية متوافقة مع OpenAI، واعرف لماذا تستهلك الفحوص المجدولة الرموز (توكن)، وخفّض الفاتورة بسياق أخف ونماذج أرخص. - [نقطة النهاية الخاصة تبدأ بحجم عمل واضح](https://broadrouter.com/ar/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): ما تحدّده قبل طلب نقطة نهاية LLM خاصة: حجم العمل، والرموز في الثانية، والتزامن، والسياق، والمنطقة، والترخيص، والميزانية، ثم كيف تختبرها. - [حقن الأوامر في وكلاء البرمجة: دفاعات تصمد فعلًا](https://broadrouter.com/ar/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): من أين يصل حقن الأوامر إلى وكيل البرمجة، والدفاعات متعددة الطبقات التي يوثّقها المزوّدون والباحثون: العزل، وتقييد الاتصال الخارجي، وقوائم السماح، والموافقة. - [حجم المقطع والتداخل في RAG: أثرهما على الاسترجاع والتكلفة](https://broadrouter.com/ar/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): كيف يغيّر حجم المقطع والتداخل دقة استرجاع RAG وحجم الفهرس وتكلفة التضمين والرموز التي يرسلها كل استعلام إلى النموذج، مع نتائج منشورة وأرقام محسوبة. - [ضبط max_tokens لنماذج الاستدلال المنطقي: طريقة لتحديد الميزانية](https://broadrouter.com/ar/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): لماذا تعيد نماذج الاستدلال المنطقي ردودًا فارغة عند قيمة max_tokens منخفضة، وكيف تحدد الميزانية من مئينات مقيسة ونافذة السياق وإعادة محاولة متصاعدة. - [تسجيل الصوت في المتصفح وإرساله إلى API لتحويل الصوت إلى نص](https://broadrouter.com/ar/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): سجّل الميكروفون باستخدام MediaRecorder، وحوّل WebM أو MP4 من Safari إلى WAV بتردد 16 kHz، ثم أرسله إلى API التفريغ الصوتي عبر خادم صغير يحتفظ بالمفتاح. - [تقليل رموز الأمر: تخفيضات مقيسة لأوامر النظام والأدوات والبيانات](https://broadrouter.com/ar/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): توفير مقيس في الرموز (توكن) عند تقليص أوامر النظام وأمثلة few-shot ومخططات الأدوات وصيغ البيانات وسجل المحادثة والتشكيل العربي، محسوب بمقسّمات رموز حقيقية. - [النموذج نفسه لدى مزود آخر: كم تتفاوت أسعار LLM API؟](https://broadrouter.com/ar/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): في 119 نموذجًا يبيعها أكثر من مزود، يبلغ أغلى سعر قائمة في الوسيط 1.76 ضعف أرخص سعر. بيانات من صفحات أسعار 41 مزودًا، نوفمبر 2026. - [نظام حماية البيانات الشخصية في السعودية والإمارات وLLM API: قائمة تحقق](https://broadrouter.com/ar/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): إرسال بيانات شخصية إلى LLM API من السعودية أو الإمارات: ماذا تقول قواعد النقل خارج المملكة، وماذا تسأل المزوّد، ومتى تهمّ المعالجة داخل الدولة. - [نص أم صوت أم رؤية؟ ابدأ من المهمة](https://broadrouter.com/ar/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): اختر بين محادثة نصية أو إدخال صور أو تحويل الصوت إلى نص حسب المهمة، مع النماذج والحدود التي توثّقها BroadNet، وما لا تقدّمه مثل تحويل النص إلى كلام. - [تحويل صوت طويل يتجاوز 25 MB إلى نص: التقسيم والإرسال والدمج](https://broadrouter.com/ar/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): كيف تفرّغ تسجيلات تتجاوز حد حجم الملف في API تحويل الصوت إلى نص: صغّر الملف، وقسّمه عند فترات الصمت بـ ffmpeg، وأرسل الأجزاء بالتوازي، ثم اجمع النص. - [ترجمة نصوص التطبيق إلى العربية بنموذج LLM: المتغيرات وصيغ الجمع](https://broadrouter.com/ar/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): ترجم ملفات i18n JSON إلى العربية عبر API لنموذج LLM: جمّد المتغيرات، وولّد صيغ الجمع العربية الست، وتحقق من صيغة ICU في CI، واعزل الأسماء المُدرجة. - [ترجمة الصوت العربي إلى نص إنجليزي عبر API: مسار مُختبَر](https://broadrouter.com/ar/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): حوّل التسجيلات العربية إلى نص إنجليزي باستدعاءين: تفريغ مع language=ar ثم ترجمة بنموذج LLM، مع فحص الأرقام والتقسيم والتكلفة والتقييم. - [متاح الآن، أو عند الطلب، أو للبحث: كيف تقرأ كتالوج النماذج بوضوح](https://broadrouter.com/ar/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): كيف تميّز نماذج الكتالوج التي تستطيع استدعاءها اليوم، والتي تحتاج إلى طلب، والتي اختُبرت فقط، مع كتالوج BroadNet ونقطة /v1/models مثالًا. - [Uno في vLLM: ما الذي قاسته دراسة التقديم لدى BroadNet؟](https://broadrouter.com/ar/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): دمجت BroadNet محوّل المسوّدة الذاتية Uno في vLLM وقاسته على H100: إنتاجية 2.58x مقارنة بـvLLM العادي عند تدفق واحد، و1.74x عند 32. المنهج والحدود. - [Uno على Gemma 4 26B: دراسة BroadNet عن نماذج مزيج الخبراء](https://broadrouter.com/ar/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): درّبت BroadNet أول محوّل Uno للمسوّدة الذاتية لنموذج مزيج خبراء، هو Gemma 4 26B A4B. ما الذي قاسته، ومتى تُجدي المسوّدة الذاتية، ومعرّف DOI. - [ميزانية رموز الوكيل: أين تذهب رموز حلقة الوكيل](https://broadrouter.com/ar/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): أمر النظام، ومخططات الأدوات، وقراءة الملفات، والسجل، والاستدلال المنطقي: كيف تنفق حلقة الوكيل الرموز، وكيف تقيس كل جزء من بيانات الاستخدام، وماذا تقلّص أولًا. - [معدل خطأ الكلمات (WER) في التعرّف على الكلام العربي: شرح مبسّط](https://broadrouter.com/ar/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): ما معدل خطأ الكلمات WER، وكيف تحسبه، ولماذا يتأثر WER العربي بشدة بالتشكيل وتوحيد الإملاء واللهجات. مع شيفرة Python وأرقام حقيقية. - [تحويل الصوت إلى نص باللهجات العربية: خليجي ومصري وشامي ومغاربي](https://broadrouter.com/ar/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): لماذا يتعثر تحويل الصوت إلى نص مع اللهجات الخليجية والمصرية والشامية والمغاربية، وكيف تبني مجموعة اختبار لكل لهجة تكشف أداء المزود على تسجيلاتك. - [الترجمة بين العربية والإنجليزية بنماذج LLM: الأوامر والمسارد والتقييم](https://broadrouter.com/ar/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): ترجم من العربية إلى الإنجليزية وبالعكس عبر API لنموذج LLM: أوامر تحفظ النبرة، ومسارد للمصطلحات الثابتة، والأرقام والأسماء، وكيف تقيّم جودة الترجمة. - [اللهجات الخليجية والمصرية والشامية: كيف تتعامل معها نماذج LLM](https://broadrouter.com/ar/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): كيف تتعامل نماذج LLM مع اللهجات الخليجية والمصرية والشامية: أين تخطئ، وكيف تكتب الأوامر للهجة، وكيف تختبرها، وما تذكره BroadNet عن Qwen3.8 27B. - [OCR العربي أم نماذج اللغة البصرية؟ قراءة المستندات والفواتير العربية](https://broadrouter.com/ar/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): متى تستخدم التعرف الضوئي على النص (OCR) ومتى نموذج لغة بصريًا للفواتير والإيصالات والنماذج العربية: فخاخ الدقة، ومخرجات JSON، وتكلفة الصفحة، وخط التحقق. - [هندسة الأوامر بالعربية: المستوى اللغوي والاتجاه والأرقام والنص المختلط](https://broadrouter.com/ar/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): اكتب أوامر تحصل بها على عربية جيدة من نموذج LLM: اختر المستوى اللغوي واللهجة، وتحكّم في الأرقام والتواريخ، ونظّف عرض RTL، وتعامل مع مزج العربية بالإنجليزية. - [RAG للعربية: التطبيع والتقطيع والاسترجاع الذي ينجح فعلًا](https://broadrouter.com/ar/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): التوليد المعزّز بالاسترجاع على مستندات عربية: وحّد الألف والياء والتاء المربوطة، وقطّع النص بذكاء، واختر تضمينات متعددة اللغات، وأضف البحث بالكلمات. - [API تحويل الصوت العربي إلى نص: كيف تختار الأنسب في 2026؟](https://broadrouter.com/ar/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): كيف تختار API لتحويل الصوت العربي إلى نص: معدل الخطأ في الكلمات، واللهجات، وحدود الملفات، وسعر ساعة الصوت، والخصوصية، مع خطة لاختبار تسجيلاتك. - [بنية بوت واتساب عربي: من الرسائل الصوتية إلى الردود](https://broadrouter.com/ar/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): بنية عملية لمساعد واتساب عربي: webhook وطابور وتحويل الرسائل الصوتية إلى نص ورد LLM وقواعد السياسة والتكلفة، مع شيفرة Python قابلة للتعديل. - [العربيزي ونماذج LLM: التعامل مع الفرانكو والأرقام مثل 3 و7](https://broadrouter.com/ar/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): العربيزي كتابة العربية بحروف لاتينية وأرقام مثل 3 لحرف ع و7 لحرف ح. كيف تكشفه، وتحوّله إلى الحروف العربية بنموذج LLM، وتردّ عليه بالطريقة الصحيحة. - [أفضل نموذج LLM لـ Aider: كيف تختار النموذج وتربطه](https://broadrouter.com/ar/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): كيف تختار أفضل نموذج LLM لـ Aider: الالتزام بصيغة التعديل، وحجم السياق، والتكلفة بالساعة، واختبار لساعة على مستودعك، وإعداد OPENAI_API_BASE بالضبط. - [أفضل نموذج مفتوح لوكلاء البرمجة: اختر حسب الأدوات والسياق والتكلفة](https://broadrouter.com/ar/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): GLM 5.3 أم DeepSeek V4.1 Flash أم Qwen3.8 27B أم Gemma 4 لوكيل البرمجة لديك؟ اختر حسب استدعاء الأدوات وحجم السياق والاستدلال المنطقي وتكلفة المهمة، مع الحساب. - [رموز الإدخال المخزّنة مؤقتًا: كيف تخفّض فاتورة LLM API](https://broadrouter.com/ar/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): الإدخال المخزّن مؤقتًا يكلّف جزءًا من سعر الإدخال العادي. كيف يُحتسب التخزين المؤقت للأوامر، وكيف تقرأ cached_tokens، وكيف تبني الأوامر لتستفيد منه. - [أرخص API لنماذج الرؤية: التكلفة لكل صورة ولكل صفحة بالحساب](https://broadrouter.com/ar/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): كم يكلّف استدعاء نموذج رؤية لكل صورة ولكل صفحة مستند: حساب الرموز لـ Gemma 4 وQwen3.8 27B وDeepSeek V4.1 Flash، وأين تكون DeepInfra أرخص. - [كيف تختار مزوّد LLM API: قائمة تحقق من 20 نقطة](https://broadrouter.com/ar/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): عشرون سؤالًا تطرحها قبل اختيار LLM API: الأسعار، والحدود، والتوافق مع OpenAI، والخصوصية، والنماذج، والدعم، ووقت التشغيل، مع ورقة تقييم وخطة اختبار. - [اختيار LLM API للغة العربية: ما الذي تختبره؟](https://broadrouter.com/ar/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): كيف تختار نموذج لغة لتطبيقات عربية: فهم اللهجات، وجودة الكتابة، والترجمة، وقراءة المستندات والصور العربية، والتكلفة، مع مجموعة اختبار تبنيها في يوم. - [Cline أم Roo Code أم Aider للفرق على النماذج المفتوحة (2026)](https://broadrouter.com/ar/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): مقارنة بين Cline وRoo Code (المؤرشف الآن، ويستمر باسم Zoo Code) وAider فيما يهم مع النماذج المفتوحة: استدعاء الأدوات، والسياق، وMCP، وgit، والموافقات. - [كيف تحسب عدد الرموز قبل إرسال طلب إلى LLM API؟](https://broadrouter.com/ar/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): قدّر رموز الأمر حسب اللغة والنموذج قبل استدعاء LLM API: لماذا تختلف المُرمِّزات، وقياسات فعلية بخمس لغات، وهوامش آمنة لقيمة max_tokens. - [10 طرق لخفض تكلفة LLM في وكلاء البرمجة، مع التوفير المتوقع](https://broadrouter.com/ar/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): عشر طرق عملية لخفض ما ينفقه Cline أو Aider أو opencode أو وكيل البرمجة الخاص بك على الرموز، لكل منها حساب التوفير وطريقة قياسه. - [بدائل DeepSeek API: خمس طرق لاستخدام نماذج DeepSeek](https://broadrouter.com/ar/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): استخدام نماذج DeepSeek عبر الواجهة الرسمية أو منصة تجميع أو كتالوج سحابي أو خوادمك أو BroadNet Inference: مقارنة السعر والخصوصية والحدود. - [سعر DeepSeek V4.1 Flash API: لكل رمز ومخزّن مؤقتًا وبالخطط](https://broadrouter.com/ar/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): سعر DeepSeek V4.1 Flash API لكل مليون رمز (توكن) إدخال ومخزّن مؤقتًا وإخراج، والخطط الشهرية والأسبوعية وبالساعة، ومقارنة بأسعار DeepSeek وDeepInfra وFireworks. - [كيف تقدّر تكلفة LLM API قبل إطلاق ميزة جديدة](https://broadrouter.com/ar/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): نموذج تكلفة لميزة ذكاء اصطناعي: المستخدمون × الإجراءات × الاستدعاءات × الرموز، مع التخزين المؤقت والاستدلال المنطقي وإعادة المحاولة وحدود المعدل، ومثال شهري. - [كيف تقيّم نموذج LLM قبل الانتقال إلى مزوّد آخر](https://broadrouter.com/ar/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): ابنِ مجموعة تقييم صغيرة من حركة بياناتك الفعلية، وقيّمها بفحوص برمجية وبنموذج LLM مُقيِّم مُعالَج التحيّز، وقارن الجودة والتكلفة وزمن الاستجابة قبل الانتقال. - [استخراج بيانات الفواتير والإيصالات والنماذج إلى JSON بنموذج رؤية](https://broadrouter.com/ar/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): مسار إنتاجي للفواتير والإيصالات والنماذج: تجهيز الصور، وأمر بمخطط JSON، والتحقق بـ Pydantic، وفحوص حسابية، وإعادة المحاولة، وطابور مراجعة بشرية. - [LLM API بسعر ثابت أم الدفع لكل رمز؟ حساب نقطة التعادل](https://broadrouter.com/ar/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): متى تصبح خطة LLM API بسعر ثابت أرخص من الدفع لكل رمز (توكن)؟ أربع فواتير شهرية محسوبة لنموذج DeepSeek V4.1 Flash، وحجم التعادل، وطريقة قياس استخدامك. - [رصيد LLM API المجاني: إلى أي مدى يكفي الرصيد المبدئي فعلًا؟](https://broadrouter.com/ar/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): ماذا يشتري رصيد LLM API المجاني فعلًا؟ رموز ودقائق صوت ومفاتيح بالساعة من رصيد BroadNet المبدئي، مقارنةً بالطبقات المجانية في Gemini وOpenRouter. - [اللائحة العامة لحماية البيانات (GDPR) وLLM API: 15 سؤالًا لمزوّدك](https://broadrouter.com/ar/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): قائمة تحقق للائحة العامة لحماية البيانات (GDPR) في LLM API: اتفاقية المعالجة، والمعالجون الفرعيون، والاحتفاظ، والتدريب، والنقل الدولي، والإخطار بالخرق، مع المواد. - [GLM 5.3 أم DeepSeek V4.1 Flash: أيهما تستخدم ومتى؟](https://broadrouter.com/ar/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): مقارنة GLM 5.3 وDeepSeek V4.1 Flash في السعر والسياق والأدوات والإخراج المنظّم والصور والاستدلال المنطقي، مع تكلفة أحجام عمل محسوبة وقاعدة بسيطة للاختيار. - [كيف تعالج الخطأ 429 في LLM API: الانتظار المتزايد والتذبذب والمحدِّدات](https://broadrouter.com/ar/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): الفرق بين 429 و5xx، والانتظار الأُسّي المتزايد مع التذبذب الكامل، ومحدِّدات التزامن والطوابير لواجهات LLM، مع شيفرة Python وNode جاهزة للإنتاج. - [مفتاح LLM API بالساعة: سعر ثابت لكل ساعة، بالشرح](https://broadrouter.com/ar/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): مفتاح الساعة من BroadNet يمنحك ساعة واحدة من Gemma 4 VL 26B على تدفق واحد بسعر ثابت. ما الذي يشمله، وما حدوده، ولمن يناسب، ومتى يصبح مجديًا. - [الاحتفاظ بالبيانات في LLM API: OpenAI وAnthropic وGemini وMistral](https://broadrouter.com/ar/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): ما تقوله OpenAI وAnthropic وGemini API وMistral وBroadNet رسميًا عن تخزين أوامر API والتدريب عليها، منقولًا من صفحاتها في نوفمبر 2026. - [أمان مفاتيح LLM API: التخزين والتدوير وقوائم عناوين IP المسموحة](https://broadrouter.com/ar/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): أبعِد مفاتيح LLM API عن الشيفرة والسجلات، وخصّص مفتاحًا لكل تطبيق، ودوّرها بانتظام، واقصرها على عناوين IP خوادمك، وألغِ المفتاح المسرَّب خلال ثوانٍ. - [شرح أسعار LLM API: الإدخال والمخزّن مؤقتًا والإخراج والاستدلال](https://broadrouter.com/ar/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): كيف تُحتسب أسعار LLM API لكل رمز (توكن): الإدخال، والإدخال المخزّن مؤقتًا، والإخراج، ورموز الاستدلال، والصور، والصوت بالدقيقة، مع أمثلة من جدول أسعار منشور. - [إرسال البيانات الحساسة إلى LLM API: الحجب والتقليل وسجلات أقل](https://broadrouter.com/ar/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): كيف ترسل بيانات العملاء أو الصحة أو المال إلى LLM API بأمان: صنّف الحقول، واحجبها بعناصر نائبة قابلة للاسترجاع، واختر عدم الاحتفاظ بالبيانات، ونظّف سجلاتك. - [التحكم في إنفاق LLM API: الميزانيات وحدود المفاتيح والحماية من الحلقات](https://broadrouter.com/ar/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): أوقف تضخّم فاتورة LLM API: رصيد مسبق الدفع، ومفتاح لكل تطبيق، وقوائم عناوين IP المسموحة، وتسجيل الاستخدام، وميزانيات في الشيفرة، وسقوف للحلقات، ومفاتيح بسعر ثابت. - [تسجيل استخدام LLM API: التكلفة لكل مستخدم وميزة وحجب البيانات](https://broadrouter.com/ar/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): سجّل كل استدعاء لـ LLM API بالرموز والتكلفة وزمن الاستجابة ومعرّف الطلب، ووزّع الإنفاق على المستخدمين والميزات، واحجب البيانات الشخصية قبل الكتابة. مع شيفرة Python. - [تكلفة LLM لكل مستخدم: اقتصاديات الوحدة لميزة ذكاء اصطناعي في SaaS](https://broadrouter.com/ar/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): احسب تكلفة ميزة الذكاء الاصطناعي لكل مستخدم نشط شهريًا، ولماذا يهمّ المستخدمون الكثيفون أكثر من المتوسط، وكيف تسعّر باقات SaaS لتغطي فاتورة النموذج. - [كيف يسعّر مزودو الاستدلال النماذج المفتوحة؟ مقارنة بين 6 مزودين](https://broadrouter.com/ar/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): كيف تسعّر OpenRouter وTogether AI وFireworks وGroq وDeepInfra وBroadNet النماذج المفتوحة: لكل رمز (توكن)، والمخزّن مؤقتًا، والدفعات، وساعات GPU، والخطط الثابتة. - [زمن استجابة LLM: زمن أول رمز مقابل الرموز في الثانية والزمن الكلي](https://broadrouter.com/ar/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): ما الذي يقيسه زمن أول رمز (TTFT) والرموز في الثانية وزمن الاستجابة الكلي، وكيف تقيس كلًّا منها من جهة العميل بالشيفرة، وأي حل يحسّن أي رقم. - [شرح البث في LLM: مقاطع SSE والفروقات (delta) والاستخدام والإلغاء](https://broadrouter.com/ar/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): كيف يعمل بث إكمالات المحادثة عبر الشبكة، وكيف تعيد بناء الإجابة من الفروقات، وتقرأ الاستخدام من المقطع الأخير، وتلغي البث بسلاسة في Python وNode وcurl. - [السياق الطويل أم RAG: متى تُغني مليون رمز عن قاعدة المتجهات؟](https://broadrouter.com/ar/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): متى تُغني نافذة سياق بمليون رمز عن قاعدة بيانات المتجهات، ومتى يظل RAG أفضل، مقارنةً بتكلفة الاستعلام وزمن الاستجابة ودقة الاسترجاع والصيانة. - [إدارة السياق للوكلاء طويلي التشغيل مع نافذة بمليون رمز](https://broadrouter.com/ar/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): كيف تُبقي جلسات الوكيل الطويلة دقيقة ومعقولة التكلفة مع نافذة بمليون رمز (توكن): سجل يُلحق به فقط، وأوامر تناسب التخزين المؤقت، ومتى تضغط السياق وماذا تحذف أولًا. - [الوكلاء طويلو التشغيل: مهلات وإعادة محاولات ونقاط حفظ تصمد](https://broadrouter.com/ar/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): أبقِ وكيل LLM يعمل لساعات: مهلة لكل استدعاء، وبث لاكتشاف التوقف، وإعادة محاولة آمنة، وآثار جانبية للأدوات لا تتكرر، ونقطة حفظ بعد كل خطوة. - [استخدام خوادم MCP مع نماذج متوافقة مع OpenAI (أي مزوّد)](https://broadrouter.com/ar/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): يعمل MCP مع أي نموذج لغوي: فالعميل يحوّل أدوات MCP إلى استدعاءات دوال. كيف تستخدم MCP مع النماذج المفتوحة في Cline وContinue وopencode وn8n أو شيفرة Python. - [متى تستحق نافذة سياق بمليون رمز (ومتى لا تستحق)؟](https://broadrouter.com/ar/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): ما الذي يتسع له 1,048,576 رمزًا، وكم يكلّف على DeepSeek V4.1 Flash، وكيف يغيّر الاستدلال المنطقي والتخزين المؤقت الحساب، ومتى يظل RAG الخيار الأفضل. - [تكلفة الأنظمة متعددة الوكلاء: الدفع لكل رمز أم الجلسات الثابتة](https://broadrouter.com/ar/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): تستهلك الأنظمة متعددة الوكلاء أضعاف رموز الوكيل الواحد. حساب الجلسات في الخطط الثابتة مقابل الدفع لكل رمز، والمضاعِفات الخفية، ومتى يفوز كلٌّ منهما. - [API تحويل الصوت إلى نص بعدة لغات: 14 لغة عبر نقطة واحدة](https://broadrouter.com/ar/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): حوّل الصوت إلى نص بالإنجليزية والإسبانية والفرنسية والصينية والعربية و9 لغات أخرى عبر نقطة واحدة متوافقة مع OpenAI بسعر الدقيقة، واعرف متى تختار النموذج العربي. - [وكيل n8n AI Agent على نموذج LLM رخيص: الإعداد والاختيار والتكلفة](https://broadrouter.com/ar/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): ابنِ مسارات عمل n8n AI Agent على نموذج منخفض التكلفة متوافق مع OpenAI: أي نموذج تختار للأدوات، والإعدادات التي تقلّل الرموز، وحساب التكلفة لكل 10,000 تشغيل. - [النماذج المفتوحة مقابل المغلقة للشركات: التكلفة والتحكم والمخاطر](https://broadrouter.com/ar/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): نماذج LLM مفتوحة الأوزان أم مغلقة لشركتك؟ قارن التكلفة والتحكم والخصوصية وفجوة الجودة والارتهان للمورّد، مع ثلاث طرق لتشغيل النماذج المفتوحة ومتى تفوز كل منها. - [الانتقال إلى واجهة API متوافقة مع OpenAI: قائمة تحقق للترحيل](https://broadrouter.com/ar/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): تغيير base_url هو الجزء السهل. ما يتعطل فعلًا: معرّفات النماذج، ورموز الاستدلال المنطقي، والاستخدام في البث، والأخطاء، وحدود المعدل، والمفاتيح. قائمة تحقق مع الكود. - [OpenRouter أم خطة بسعر ثابت لوكلاء الذكاء الاصطناعي: أيهما أرخص؟](https://broadrouter.com/ar/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): يحتسب OpenRouter لكل رمز مع رسوم على شراء الرصيد، والخطة الثابتة تحتسب لكل جلسة. حساب تكلفة الوكلاء الدائمين، ورسوم OpenRouter وحدوده، وأيهما يناسبك. - [كيف تشغّل وكلاء البرمجة طوال الليل دون مراقبة، بأمان](https://broadrouter.com/ar/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): إعداد عملي لتشغيل وكلاء البرمجة دون مراقبة: العزل، ونقاط حفظ git، وملفات المهام، وحدود الخطوات، والمراجعة الصباحية، وجلسات الخطط بسعر ثابت طوال الليل. - [كيف تبني الأوامر لتُخزَّن بادئتها مؤقتًا؟](https://broadrouter.com/ar/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): كيف ترتّب أوامر LLM لتبقى البادئة المخزّنة مؤقتًا متطابقة: أمر النظام، ومخططات الأدوات، والمستندات، وسجل يُضاف إليه فقط، مع اختبار يثبت أن البادئة يُعاد استخدامها. - [رموز الاستدلال المنطقي: لماذا تدفع مقابل مخرجات أكثر مما تراه؟](https://broadrouter.com/ar/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): تحتسب نماذج الاستدلال المنطقي التفكير المخفي رموزَ إخراج (توكن). كيف تقيسه، وكم يكلّف لكل نموذج، وكيف تخفّضه بـ reasoning_effort وmax_tokens وتوجيه المهام. - [كيف تشغّل وكلاء برمجة متوازيين على LLM API بسعر ثابت](https://broadrouter.com/ar/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): الوكلاء الفرعيون والمنسّقون وعمّال الدفعات على DeepSeek V4.1 Flash غير المحدود: كم جلسة تحتاج، وكيف توزعها على المفاتيح، وكيف تضع سقفًا للتداخل. - [الاستضافة الذاتية لنموذج LLM أم API؟ التكلفة الحقيقية ونقطة التعادل](https://broadrouter.com/ar/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): أسعار استئجار GPU من الصفحات الرسمية، ومعادلة تكلفة المليون رمز (توكن) على GPU خاص بك، ونسبة الاستغلال اللازمة لتتفوق على API، وتكاليف التشغيل المنسية. - [أسعار API تحويل الصوت إلى نص: مقارنة سعر الدقيقة في 2026](https://broadrouter.com/ar/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): سعر الدقيقة ولكل 1,000 ساعة لدى OpenAI وGoogle وAWS وDeepgram وAssemblyAI وSpeechmatics وGroq وBroadNet Transcribe من الصفحات الرسمية، مع الإضافات. - [وضع JSON أم json_schema: مخرجات منظَّمة يمكن الاعتماد عليها](https://broadrouter.com/ar/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): وضع JSON يضمن JSON صالحًا؛ وjson_schema يضمن البنية التي تحددها. كيف تطلب كلًّا منهما من النماذج المتوافقة مع OpenAI، وتتحقق بـ Pydantic، وتصلح المخرجات الخاطئة. - [عدد الرموز لكل كلمة في العربية والإنجليزية: لماذا تكلّف العربية أكثر؟](https://broadrouter.com/ar/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): لماذا تستهلك العربية غالبًا رموزًا (توكن) أكثر من الإنجليزية للمعنى نفسه، ولماذا يضلّل مقياس الرموز لكل كلمة، وشيفرة لقياس نسبتك من حقل usage. - [استدعاء الأدوات في النماذج المفتوحة: المخططات والتوازي والإعادة](https://broadrouter.com/ar/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): استدعاء الدوال على نماذج مفتوحة متوافقة مع OpenAI: نصائح للمخططات، والاستدعاءات المتوازية، وخمس طرق يفشل بها الاستدعاء، وحلقة Python تتحقق وتعيد المحاولة. - [كيف تفرّغ تسجيلات مراكز الاتصال نصيًا: خط معالجة جاهز للعربية](https://broadrouter.com/ar/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): افصل المكالمات الاستريو حسب القناة، وقطّع أدوار الكلام، وفرّغ العربية والإنجليزية، واحجب البيانات الشخصية ولخّص كل مكالمة بنموذج LLM. خط معالجة Python مع التكاليف. - [النماذج المفتوحة في وكلاء البرمجة: Aider وCline وopencode وCodex](https://broadrouter.com/ar/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): وجّه Aider وCline وRoo Code وopencode وContinue وCodex CLI إلى أي نقطة نهاية متوافقة مع OpenAI. أسماء الحقول الدقيقة لكل أداة، والإعدادات التي تتعطل. - [الرسائل الصوتية إلى نص للشركات: من WhatsApp وTelegram إلى تذاكر](https://broadrouter.com/ar/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): حوّل رسائل العملاء الصوتية من WhatsApp Business وبوتات Telegram إلى نصوص وتذاكر دعم: التنزيل، والتفريغ الصوتي، واستخراج الحقول بنموذج LLM، والتوجيه. - [خطة LLM أسبوعية للمشاريع القصيرة: المدار مقابل الساعة والشهر](https://broadrouter.com/ar/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): متى تتفوق خطة DeepSeek V4.1 Flash غير المحدودة لأسبوع على جلسة بالساعة أو خطة شهرية أو الدفع لكل رمز، مع الحسابات للهاكاثونات والسباقات وعمليات الترحيل. - [ما معنى «غير محدود» فعلًا في خطط API نماذج اللغة؟](https://broadrouter.com/ar/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): عدّاد الرموز، ونوافذ الـ 5 ساعات، والسقوف الأسبوعية، وحدود المعدل والتزامن: سبعة أسئلة تكشف هل الخطة «غير المحدودة» غير محدودة فعلًا لوكيلك. - [بديل Whisper API: متى تنتقل، وكيف تختبر البديل](https://broadrouter.com/ar/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): متى يكون البديل المستضاف لواجهة OpenAI Whisper API منطقيًا: معدل الخطأ في الكلمات على تسجيلاتك، واللغات، والسعر للساعة، والحدود، والخصوصية، مع شيفرة اختبار. - [عدم الاحتفاظ بالبيانات في LLM API: قائمة تحقق من 10 أسئلة](https://broadrouter.com/ar/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): ماذا يجب أن يعني «عدم الاحتفاظ بالبيانات» في LLM API: الأوامر والمخرجات والسجلات والتدريب والمعالجون الفرعيون والتدقيق. عشرة أسئلة قبل إرسال بياناتك. - [Guides in Arabic](https://broadrouter.com/ar/guides/) - [استخدام OpenAI Python SDK مع BroadNet Inference](https://broadrouter.com/ar/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): وجّه حزمة OpenAI Python SDK الرسمية إلى BroadNet Inference بتغيير base_url فقط: التثبيت وضبط المفتاح وأول طلب وبث الإجابة. - [استخدام OpenAI Node.js SDK مع BroadNet Inference](https://broadrouter.com/ar/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): وجّه حزمة OpenAI SDK الرسمية لـ Node.js وTypeScript إلى BroadNet Inference بضبط baseURL: التثبيت وأول طلب وبث الرموز وتسجيل معرّفات الطلبات. - [جرّب BroadNet Inference في متصفحك دون حساب](https://broadrouter.com/ar/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): اختبر تحويل الكلام العربي إلى نص وقراءة المستندات والمحادثة مع Gemma 4 في ساحة تجربة BroadNet دون تسجيل أو بطاقة، ثم انقل الطلب نفسه إلى الـ API. - [أول طلب إلى BroadNet Inference باستخدام curl](https://broadrouter.com/ar/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): نفّذ أول استدعاءاتك لـ BroadNet Inference عبر curl: إكمال محادثة، وإجابة ببث متدفق، وتحويل صوت إلى نص، ثم اقرأ الاستجابة والاستهلاك والأخطاء. - [البث المتدفق لإكمال المحادثات والتعامل مع انقطاع الاتصال](https://broadrouter.com/ar/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): كيف يعمل البث المتدفق في BroadNet Inference: أحداث Server-Sent Events، وعلامة [DONE]، واستهلاك الرموز في آخر جزء، وما العمل حين ينتهي البث مبكرًا. - [لماذا يعيد نموذج الاستدلال إجابة فارغة؟](https://broadrouter.com/ar/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): تفكّر نماذج DeepSeek V4.1 Flash وGLM 5.3 وQwen3.8 27B قبل أن تجيب، ويُحتسب هذا الاستدلال ضمن max_tokens. إليك كيف تتجنب الردود الفارغة. - [رموز أخطاء BroadNet Inference ومتى تعيد المحاولة](https://broadrouter.com/ar/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): ماذا يعني كل رمز حالة HTTP من BroadNet Inference API، وأيها تعيد محاولته مع انتظار متزايد، وأيها تصلحه أولًا، وماذا ترسل إلى الدعم مع معرّف الطلب. - [اربط مفاتيح API بخوادمك عبر قوائم عناوين IP المسموح بها](https://broadrouter.com/ar/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): احمِ مفاتيح BroadNet Inference بقوائم عناوين IP المسموح بها، وافصل مفاتيح الاستدلال عن بيانات اعتماد البوابة، وألغِ أي مفتاح مسرّب خلال ثوانٍ. مع أمثلة curl. - [كيف تعمل خطط LLM API غير المحدودة مع الوكلاء](https://broadrouter.com/ar/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): خطط DeepSeek V4.1 Flash غير المحدودة بلا عدّاد رموز ولا حد أسبوعي. إليك كيف تعمل الجلسات والوكلاء المتوازون والمفاتيح والفوترة فعليًا. - [تشغيل Aider على DeepSeek V4.1 Flash غير المحدود](https://broadrouter.com/ar/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): اربط وكيل البرمجة Aider بـ BroadNet Inference عبر إعداده المتوافق مع OpenAI، وشغّله على خطة DeepSeek V4.1 Flash غير محدودة بسعر ثابت. - [استخدام LangChain مع BroadNet Inference](https://broadrouter.com/ar/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): اربط ChatOpenAI في LangChain بـ BroadNet Inference عبر base_url وapi_key، واعرف ما الذي يحتفظ به LangChain وما الذي يُسقطه من نماذج الاستدلال المنطقي. - [استخدام Cline مع BroadNet Inference (مزوّد OpenAI Compatible)](https://broadrouter.com/ar/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): اضبط إضافة Cline في VS Code على BroadNet Inference عبر مزوّد OpenAI Compatible، مع الحقول الدقيقة ومعرّفات النماذج وأحجام السياق وحلول المشكلات الشائعة. - [استخدام Roo Code مع BroadNet Inference (مزوّد OpenAI Compatible)](https://broadrouter.com/ar/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): أُوقف Roo Code في مايو 2026. إليك إعدادات التثبيتات الحالية والنسخ المشتقة، وإلى أين تنتقل بعده (Cline أو Zoo Code) مع BroadNet Inference. - [استخدام Continue مع BroadNet Inference (config.yaml وapiBase)](https://broadrouter.com/ar/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): أضف نماذج BroadNet Inference إلى إضافة Continue في بيئة التطوير عبر provider openai وapiBase في config.yaml، مع قدرة tool_use وطول السياق. - [إضافة BroadNet Inference إلى opencode كمزوّد مخصّص](https://broadrouter.com/ar/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): أضف BroadNet Inference إلى وكيل البرمجة opencode كمزوّد مخصّص متوافق مع OpenAI في opencode.jsonc، مع أسماء بديلة للنماذج وقدراتها وحدودها. - [ربط Open WebUI بـ BroadNet Inference](https://broadrouter.com/ar/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): أضف BroadNet Inference إلى Open WebUI كاتصال متوافق مع OpenAI، وأدرج معرّفات النماذج في القائمة المسموح بها، وعالج مشكلات التحقق والردود الفارغة وحد المعدل. - [إضافة BroadNet Inference إلى LibreChat كنقطة نهاية مخصّصة](https://broadrouter.com/ar/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): أضف BroadNet Inference إلى LibreChat كنقطة نهاية مخصّصة متوافقة مع OpenAI في librechat.yaml، مع المفتاح في .env وقائمة النماذج ونموذج العناوين والحلول. - [استخدام BroadNet Inference في n8n (بيانات اعتماد OpenAI وBase URL)](https://broadrouter.com/ar/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): وجّه بيانات اعتماد OpenAI في n8n إلى BroadNet Inference عبر حقل Base URL، ثم استخدم عقدة OpenAI Chat Model بالمعرّف مع إيقاف خيار Responses API. - [استخدام Vercel AI SDK مع BroadNet Inference](https://broadrouter.com/ar/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): استدعِ BroadNet Inference من Vercel AI SDK عبر createOpenAICompatible من الحزمة @ai-sdk/openai-compatible: generateText وstreamText والأدوات ومستوى الاستدلال. - [API لتحويل الصوت العربي إلى نص خطوة بخطوة](https://broadrouter.com/ar/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): حوّل الصوت العربي إلى نص بنموذج BroadNet Transcribe Arabic عبر نقطة نهاية متوافقة مع OpenAI: الصيغ، وحقل language، وأمثلة Python وcurl، والسعر. - [توجيه LiteLLM إلى BroadNet Inference (عبر SDK والوكيل)](https://broadrouter.com/ar/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): استدعِ BroadNet Inference عبر LiteLLM بالبادئة openai/ قبل اسم النموذج وبالمعامل api_base، في Python SDK أو في config.yaml للوكيل، مع حدود معدل مطابقة. - [تشغيل OpenAI Agents SDK على BroadNet Inference](https://broadrouter.com/ar/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): استخدم OpenAI Agents SDK لـ Python مع BroadNet Inference عبر AsyncOpenAI وOpenAIChatCompletionsModel، وأوقف التتبّع الذي يحتاج إلى مفتاح OpenAI. - [تحويل رسائل WhatsApp الصوتية إلى نص عبر API](https://broadrouter.com/ar/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): حوّل رسائل WhatsApp الصوتية (.ogg) أو تسجيلات الهاتف (.m4a) إلى WAV بـ ffmpeg، ثم فرّغها إلى نص بالعربية أو بـ 13 لغة أخرى باستدعاء API واحد. - [تفريغ الرسائل الصوتية من WhatsApp إلى نص في n8n](https://broadrouter.com/ar/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): مسار عمل في n8n يستقبل رسالة WhatsApp الصوتية، وينزّلها من Meta، ويحوّلها بـ ffmpeg، ثم يفرّغها إلى نص بالعربية أو بـ 13 لغة أخرى. - [قراءة الفواتير والإيصالات العربية بنموذج رؤية](https://broadrouter.com/ar/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): أرسل صورة فاتورة أو إيصال إلى نموذج رؤية على BroadNet Inference واحصل على الإجمالي والتاريخ واسم المورّد بصيغة JSON، مع أسئلة بالعربية أو الإنجليزية. - [استخدام Zed مع BroadNet Inference كمزوّد متوافق مع OpenAI](https://broadrouter.com/ar/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): أضف BroadNet Inference إلى محرر Zed كمزوّد متوافق مع OpenAI في settings.json، مع معرّفات النماذج وحدود الرموز والقدرات ومتغير المفتاح. - [استخدام Kilo Code مع BroadNet Inference (مزوّد OpenAI Compatible)](https://broadrouter.com/ar/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): اربط Kilo Code بـ BroadNet Inference كمزوّد مخصّص من نوع OpenAI Compatible، من نافذة VS Code أو kilo.json، مع معرّفات النماذج والحدود وخيارات الأدوات. - [إضافة BroadNet Inference إلى goose كمزوّد مخصّص](https://broadrouter.com/ar/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): اربط وكيل الذكاء الاصطناعي goose بـ BroadNet Inference كمزوّد مخصّص من نوع OpenAI Compatible، من goose Desktop أو goose configure أو ملف JSON. - [استخدام Pydantic AI مع BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/ar/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): شغّل وكلاء Pydantic AI على BroadNet Inference باستخدام OpenAIChatModel وOpenAICompatibleProvider، مع استدعاء الأدوات والمخرجات المنظَّمة وضبط ملف تعريف النموذج. - [استخدام CrewAI مع BroadNet Inference كنقطة نهاية OpenAI مخصّصة](https://broadrouter.com/ar/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): وجّه وكلاء CrewAI إلى BroadNet Inference عبر LLM(custom_openai=True, base_url=...)، ثم اختر النماذج للأدوات والسياق الطويل والخطة ثابتة السعر. - [استخدام LlamaIndex مع BroadNet Inference عبر OpenAILike](https://broadrouter.com/ar/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): اربط LlamaIndex بـ BroadNet Inference عبر الصنف OpenAILike، واضبط api_base وcontext_window وis_chat_model وis_function_calling_model بالشكل الصحيح. - [إضافة BroadNet Inference إلى Dify عبر إضافة OpenAI-API-compatible](https://broadrouter.com/ar/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): اربط Dify بـ BroadNet Inference عبر إضافة مزوّد النماذج OpenAI-API-compatible، مع قيم الحقول الدقيقة وأحجام السياق وإعدادات استدعاء الدوال. - [استخدام Jan مع BroadNet Inference كمزوّد مخصّص متوافق مع OpenAI](https://broadrouter.com/ar/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): أضف BroadNet Inference إلى تطبيق Jan لسطح المكتب كمزوّد مخصّص متوافق مع OpenAI، وأضف معرّفات النماذج يدويًا، وفعّل الأدوات والرؤية لكل نموذج. - [ربط SillyTavern بـ BroadNet Inference (Custom OpenAI-compatible)](https://broadrouter.com/ar/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): اربط SillyTavern بـ BroadNet Inference عبر مصدر Custom (OpenAI-compatible) في Chat Completion، مع معرّفات النماذج الصحيحة ومعالجة الموجّه وحدود الرموز. - [استخدام Spring AI مع BroadNet Inference (OpenAI base-url)](https://broadrouter.com/ar/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): وجّه OpenAI chat starter في Spring AI إلى BroadNet Inference عبر spring.ai.openai.base-url ومعرّف نموذج من BroadNet وmax-tokens، ثم استدعه من ChatClient. - [استدعاء BroadNet Inference من Go بحزمة openai-go](https://broadrouter.com/ar/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): استخدم حزمة openai-go v3 الرسمية مع BroadNet Inference عبر option.WithBaseURL: إكمال المحادثات، والبث مع بيانات الاستخدام، ومعالجة الأخطاء في Go. - [بناء وكلاء LangGraph على BroadNet Inference](https://broadrouter.com/ar/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): شغّل StateGraph أو create_agent في LangGraph على BroadNet Inference عبر ChatOpenAI وbase_url ومعرّف نموذج من BroadNet، مع الأدوات والبث والذاكرة. - [استخدام خطوط Haystack ووكلائه مع BroadNet Inference](https://broadrouter.com/ar/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): وجّه OpenAIChatGenerator في Haystack إلى BroadNet Inference عبر api_base_url ومعرّف نموذج BroadNet ومفتاح Secret، للخطوط (pipelines) والأدوات والتدفق والوكلاء. - [استخدام Semantic Kernel مع BroadNet Inference](https://broadrouter.com/ar/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): اربط Microsoft Semantic Kernel بـ BroadNet Inference في Python عبر عميل AsyncOpenAI، أو في C# عبر نقطة نهاية مخصّصة، للمحادثة والإضافات والوكلاء. - [استخدام وكلاء Mastra مع BroadNet Inference](https://broadrouter.com/ar/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): شغّل وكلاء Mastra على BroadNet Inference بكائن model يضبط id وurl وapiKey، مع البادئة broadnet/ المضاعفة التي يحتاجها معرّف النموذج. - [مخرجات منظَّمة باستخدام Instructor على BroadNet Inference](https://broadrouter.com/ar/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): استخرج كائنات Pydantic من BroadNet Inference باستخدام Instructor، عبر from_openai أو from_provider مع base_url مخصّص، واختر وضع Mode المناسب. - [استخدام DSPy مع BroadNet Inference (اضبط model_type="chat")](https://broadrouter.com/ar/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): اضبط dspy.LM على BroadNet Inference بالبادئة openai/ وapi_base وmodel_type="chat"، لأن DSPy 3.5 يرسل نماذج OpenAI إلى Responses API. - [تقييم نماذج BroadNet باستخدام promptfoo](https://broadrouter.com/ar/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): شغّل تقييمات promptfoo ومقارنات النماذج جنبًا إلى جنب على BroadNet Inference باستخدام المزوّد openai:chat وapiBaseUrl وapiKeyEnvar، مع المُقيِّمات. - [تشغيل smolagents على BroadNet Inference عبر OpenAIModel](https://broadrouter.com/ar/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): اربط smolagents من Hugging Face بـ BroadNet Inference عبر OpenAIModel وapi_base، لتشغيل CodeAgent على أي نموذج أو ToolCallingAgent على GLM 5.3 وQwen3.8 27B. - [ربط Flowise بـ BroadNet Inference (عقدة OpenAI Custom Model)](https://broadrouter.com/ar/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): استخدم BroadNet Inference في مسارات Flowise عبر عقدة OpenAI Custom Model، مع Base Path وبيانات اعتماد OpenAI API واسم نموذج BroadNet مكتوب يدويًا. - [ربط Langflow بـ BroadNet Inference (مكوّن OpenAI)](https://broadrouter.com/ar/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): استخدم BroadNet Inference في Langflow عبر حقل OpenAI API Base في مكوّن OpenAI واسم نموذج مكتوب يدويًا، ولماذا لا يُدرجه مزوّد OpenAI Compatible. - [Glossary entries in Arabic](https://broadrouter.com/ar/glossary/) - [وكيل الذكاء الاصطناعي (AI agent)](https://broadrouter.com/ar/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): وكيل الذكاء الاصطناعي نظام يوجّه فيه النموذج اللغوي خطواته بنفسه نحو هدف: يقرر أي الأدوات يستدعي، ويقرأ النتائج، ويكرّر حتى يُنتج إجابة نهائية أو يبلغ حدًّا. أما البرنامج الثابت الذي يستدعي النموذج عند نقاط محددة فيُسمّى عادةً سير عمل (workflow). - [مفتاح API (API key)](https://broadrouter.com/ar/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): مفتاح API هو سلسلة نصية سرية تعرّف العميل الذي يستدعي الواجهة وتخوّله. تتوقعه واجهات LLM API عادةً في الترويسة Authorization: Bearer، وكل من يملك المفتاح يستطيع إرسال طلبات تُحتسب تكلفتها على الحساب المالك له. - [التشكيل (الحركات العربية، Arabic diacritics)](https://broadrouter.com/ar/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): التشكيل هو علامات صغيرة تُكتب فوق الحروف العربية أو تحتها لبيان الحركات القصيرة والحروف المضعّفة وغياب الحركة. وهو اختياري: تخلو منه معظم النصوص العربية، ومنها الأخبار، فيستنتج القرّاء والبرمجيات الحركات من السياق. - [تحديد اللهجة العربية (Arabic dialect identification)](https://broadrouter.com/ar/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): تحديد اللهجة العربية هو مهمة معرفة الصيغة العربية التي كُتب بها نص أو قيل بها تسجيل، مثل المصرية أو الخليجية أو الشامية أو المغاربية أو الفصحى الحديثة، على مستوى المنطقة أو البلد أو المدينة. ويُستخدم لتوجيه الرسائل واختيار الأوامر وتقسيم مجموعات التقييم. - [توحيد النص العربي (Arabic text normalization)](https://broadrouter.com/ar/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): توحيد النص العربي هو الخطوة التي تحوّل الصيغ الإملائية المختلفة للكلمة نفسها إلى صيغة واحدة قبل المطابقة: حذف التشكيل والتطويل، وتوحيد أشكال الألف، وغالبًا تحويل ى إلى ي وة إلى ه. ويُطبَّق على مفاتيح البحث والتقييم، لا على النص الذي تعرضه. - [العربيزي (الفرانكو آراب، Arabizi)](https://broadrouter.com/ar/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): العربيزي هي العربية المكتوبة بشكل غير رسمي بحروف لاتينية، مع أرقام للأصوات العربية التي لا مقابل لها في اللاتينية، مثل 3 بدل ع. وهي شائعة في المحادثات ووسائل التواصل، وغالبًا ما تكون باللهجات، ولا إملاء معياريًا لها، وكثيرًا ما تُمزج بالإنجليزية أو الفرنسية. - [AWQ (التكميم الواعي بالتفعيلات)](https://broadrouter.com/ar/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (التكميم الواعي بالتفعيلات) طريقة تُطبَّق بعد التدريب لضغط أوزان نموذج اللغة إلى دقة منخفضة، غالبًا أعداد صحيحة بـ 4 بتات. وتستخدم إحصاءات التفعيلات من مجموعة معايرة صغيرة لتحديد أهم قنوات الأوزان، ثم تكبّرها كي تفقد دقة أقل. ولا تحتاج إلى الانتشار العكسي. - [العنوان الأساسي (Base URL)](https://broadrouter.com/ar/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): العنوان الأساسي (base URL) هو الجذر الثابت لعنوان الواجهة، مثل https://api.example.com/v1، الذي يضعه العميل قبل كل مسار نقطة نهاية، مثل /chat/completions. وفي حزم SDK المتوافقة مع OpenAI، يكفي تغيير العنوان الأساسي، مع مفتاح API ومعرّف النموذج، لتوجيه الشيفرة نفسها إلى مزوّد آخر. - [واجهة الدفعات (Batch API)](https://broadrouter.com/ar/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): واجهة الدفعات (Batch API) هي نقطة نهاية غير متزامنة لنماذج اللغة تستقبل طلبات كثيرة دفعة واحدة، غالبًا في ملف، وتعالجها خلال مهلة محددة بدل المعالجة الفورية، ثم تُرجع النتائج معًا. ويسعّر مزوّدون مثل OpenAI وAnthropic طلبات الدفعات بخصم مقارنةً بواجهاتهم القياسية. - [BM25 (ترتيب نتائج البحث بالكلمات المفتاحية)](https://broadrouter.com/ar/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (أو Okapi BM25) هي دالة ترتيب للبحث بالكلمات المفتاحية تمنح المستند درجة بحسب عدد مرات ظهور كلمات الاستعلام فيه، ومدى ندرة هذه الكلمات في المجموعة كلها، وطول المستند. وهي مقياس التشابه الافتراضي في Lucene وElasticsearch. - [ترميز أزواج البايتات (BPE)](https://broadrouter.com/ar/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): ترميز أزواج البايتات (BPE) خوارزمية تجزئة إلى وحدات دون الكلمة، تبدأ من الحروف أو البايتات المفردة، ثم تدمج مرارًا أكثر زوج متجاور تكرارًا في مدخل جديد في المفردات، حتى تبلغ حجم المفردات المستهدف. ومعظم مجزّئات نماذج اللغة الحالية تستخدم BPE أو صيغته على مستوى البايت. - [رموز الإدخال المخزّنة مؤقتًا (Cached input tokens)](https://broadrouter.com/ar/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): رموز الإدخال المخزّنة مؤقتًا هي رموز من الأمر (prompt) عالجها مزوّد نماذج اللغة مؤخرًا ويستطيع إعادة استخدامها من ذاكرة مؤقتة بدل حسابها من جديد، وهي عادةً بادئة متكررة مثل أمر النظام أو مستند. ويحتسبها المزوّدون بسعر أقل من رموز الإدخال الجديدة. - [سلسلة التفكير (Chain-of-thought، CoT)](https://broadrouter.com/ar/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): سلسلة التفكير (CoT) أسلوب في كتابة الأوامر يُعرض فيه على نموذج اللغة، أو يُطلب منه، أن يكتب خطوات استدلال وسيطة قبل إجابته النهائية. وهي تحسّن الدقة في المسائل متعددة الخطوات كالحساب والمنطق، وصارت نماذج الاستدلال المنطقي تولّدها تلقائيًا. - [معدل الخطأ في الحروف (CER)](https://broadrouter.com/ar/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): معدل الخطأ في الحروف (CER) يقيس دقة التفريغ الصوتي أو OCR على مستوى الحرف: عدد الاستبدالات والحذوفات والإضافات اللازمة لتحويل المخرجات إلى النص المرجعي، مقسومًا على عدد حروف النص المرجعي. والأقل أفضل، والصفر يعني تطابقًا تامًا. - [Chat Completions API (واجهة إكمال المحادثة)](https://broadrouter.com/ar/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Chat Completions API هي نقطة نهاية HTTP لدى OpenAI، أي POST /v1/chat/completions، تستقبل اسم نموذج وقائمة رسائل موسومة بأدوارها وتُرجع الرسالة التالية من النموذج. وصارت بنية طلبها واستجابتها الصيغة المشتركة التي يعيد استخدامها المزوّدون المتوافقون مع OpenAI وحزم SDK. - [قالب المحادثة (Chat template)](https://broadrouter.com/ar/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): قالب المحادثة هو قاعدة التنسيق، المكتوبة غالبًا بلغة Jinja والمرفقة بمجزّئ النموذج، التي تحوّل قائمة رسائل المحادثة بأدوارها إلى تسلسل الرموز الواحد الذي يتوقعه النموذج، بما فيه رموز التحكم الخاصة. واستخدام قالب خاطئ قد يضعف إجابات النموذج بشدة. - [التقطيع إلى أجزاء (Chunking في RAG)](https://broadrouter.com/ar/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): التقطيع (chunking) هو الخطوة في سلسلة الاسترجاع التي تقسم المستندات إلى قطع أصغر تسمّى الأجزاء (chunks) قبل تضمينها وفهرستها. ويحدد حجم الجزء مقدار النص في كل قطعة، ويكرر التداخل بعض النص بين الأجزاء المتجاورة كي لا تضيع الأفكار المقطوعة عند الحد الفاصل. - [التناوب اللغوي (Code-switching)](https://broadrouter.com/ar/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): التناوب اللغوي هو استخدام لغتين أو أكثر، أو صيغتين من اللغة نفسها، في المحادثة أو الجملة الواحدة، كلامًا أو كتابةً. ومن أمثلته أن يمزج الناطقون بالعربية كلمات إنجليزية في لهجتهم، أو أن ينتقلوا بين العربية الفصحى الحديثة ولهجتهم. - [وكيل البرمجة بالذكاء الاصطناعي (AI coding agent)](https://broadrouter.com/ar/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): وكيل البرمجة بالذكاء الاصطناعي أداة برمجية يعمل فيها نموذج لغة على قاعدة شيفرة في حلقة: يقرأ الملفات ويعدّلها، وينفّذ أوامر كالاختبارات، ويقرأ المخرجات، ثم يقرّر الخطوة التالية حتى تنتهي المهمة أو يُبلغ حدّ معيّن. - [التزامن (Concurrency)](https://broadrouter.com/ar/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): التزامن (concurrency) هو عدد الطلبات التي تكون لدى العميل قيد التنفيذ في الوقت نفسه. وكثيرًا ما تضع واجهات LLM API سقفًا له لكل مفتاح أو خطة، بمعزل عن عدد الطلبات في الدقيقة، لأن كل طلب مفتوح يشغل قدرة من وحدة GPU حتى تكتمل استجابته. - [ضغط السياق (Context compaction)](https://broadrouter.com/ar/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): ضغط السياق هو تلخيص محادثة تقترب من حد نافذة السياق ومتابعة الجلسة من ذلك الملخص، فيحتفظ الوكيل بالقرارات والحالة الأساسية ويتخلص من السجل الخام، كمخرجات الأدوات القديمة. - [هندسة السياق (Context engineering)](https://broadrouter.com/ar/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): هندسة السياق هي عملية اختيار مجموعة الرموز (التوكنات) التي يراها نموذج اللغة في كل طلب والحفاظ عليها، من أمر النظام وتعريفات الأدوات والسجل والمستندات المسترجعة والملاحظات، حتى يحصل النموذج على أصغر قدر من المعلومات عالية الدلالة يحتاجه لإنجاز المهمة جيدًا. - [تآكل السياق (Context rot)](https://broadrouter.com/ar/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): تآكل السياق هو الفقدان التدريجي للدقة والموثوقية الذي تُظهره نماذج اللغة كلما طالت مدخلاتها، حتى حين تبقى المهمة نفسها وتكون المدخلات ضمن نافذة السياق بفارق كبير. وقد انتشر المصطلح بفضل تقرير تقني أصدرته Chroma عام 2025 واختبر 18 نموذجًا. - [نافذة السياق (Context window)](https://broadrouter.com/ar/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): نافذة السياق هي الحد الأقصى لعدد الرموز (التوكنات) التي يستطيع نموذج اللغة التعامل معها في طلب واحد. وتشمل الأمر (prompt) وسجل المحادثة وتعريفات الأدوات والصور والمخرجات المولَّدة، بما فيها رموز الاستدلال المنطقي. وكل ما لا يتسع لها يجب قصّه أو تلخيصه أو توزيعه على عدة طلبات. - [إقامة البيانات (Data residency)](https://broadrouter.com/ar/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): إقامة البيانات هي اشتراط أو التزام بأن تُخزَّن البيانات وتُعالَج في دولة أو منطقة محددة. وفي واجهة نموذج لغوي (LLM API) تشمل مكان معالجة الأوامر أثناء الاستدلال ومكان أي محتوى مخزّن، وكثيرًا ما تستثني بيانات الحساب والفوترة. - [اللهجات العربية (Dialectal Arabic)](https://broadrouter.com/ar/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): اللهجات العربية هي العربية المحكية يوميًا في منطقة أو مدينة، مثل المصرية أو الخليجية أو الشامية أو المغاربية. وتختلف عن العربية الفصحى الحديثة في الأصوات والمفردات والقواعد، ولا إملاء معياريًا لها، وتُكتب على الإنترنت بشكل متزايد، ما يجعلها أصعب على نماذج اللغة. - [تمييز المتحدثين (Speaker diarization)](https://broadrouter.com/ar/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): تمييز المتحدثين (speaker diarization) هو مهمة تقسيم التسجيل الصوتي إلى مقاطع حسب المتحدث، للإجابة عن سؤال "من تكلّم ومتى". ويَسِم كل مقطع كلام بمعرّف متحدث مجهول مثل SPEAKER_00، ويُدمج عادةً مع النص المفرّغ ليُظهر من قال كل سطر. - [التقطير المعرفي (Distillation)](https://broadrouter.com/ar/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): التقطير، أو التقطير المعرفي (knowledge distillation)، أسلوب تدريب يتعلم فيه نموذج طالب أصغر إعادة إنتاج مخرجات نموذج معلّم أكبر، مثل توزيعاته الاحتمالية أو إجاباته المولَّدة، فيحتفظ الطالب بكثير من جودة المعلّم بتكلفة تشغيل أقل. - [اتفاقية معالجة البيانات (DPA)](https://broadrouter.com/ar/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): اتفاقية معالجة البيانات (DPA) هي العقد بين المتحكم والمعالج الذي يحكم طريقة تعامل المعالج مع البيانات الشخصية. تشترطها المادة 28(3) من GDPR وتحدد محتواها: تعليمات موثقة، والسرية، والأمن، وقواعد المعالجين الفرعيين، وحذف البيانات أو إعادتها، والتدقيق. - [التضمينات (Embeddings)](https://broadrouter.com/ar/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): التضمينات (embeddings) قوائم من الأعداد العشرية (متجهات) يُنتجها نموذج لتمثيل قطعة نص أو صورة أو أي مدخل آخر. وتحصل المدخلات المتشابهة في المعنى على متجهات متقاربة، فيمكن استخدام المسافة بين المتجهات لترتيب نتائج البحث أو تجميع المستندات أو إيجاد العناصر المرتبطة. - [التقييم (Eval)](https://broadrouter.com/ar/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): التقييم (eval، اختصار evaluation) اختبار قابل للتكرار لنموذج لغوي أو أمر: مجموعة ثابتة من المدخلات، ووصف لما تبدو عليه المخرجات الجيدة، ومصحِّح (شيفرة أو إنسان أو نموذج آخر) يمنح كل مخرج درجة حتى يمكن المقارنة بين الإصدارات. - [الانتظار الأُسّي (Exponential backoff)](https://broadrouter.com/ar/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): الانتظار الأُسّي (exponential backoff) استراتيجية لإعادة المحاولة ينتظر فيها العميل مدة أطول بعد كل محاولة فاشلة، وعادةً يضاعف التأخير حتى حد أقصى، قبل أن يحاول مجددًا. وإضافة تذبذب عشوائي (jitter) إلى كل انتظار توزّع محاولات العملاء الكثيرين كي لا تصل إلى الخادم في اللحظة نفسها. - [التلقين بأمثلة قليلة (Few-shot prompting)](https://broadrouter.com/ar/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): التلقين بأمثلة قليلة (few-shot prompting) أسلوب يتضمن فيه الأمر عددًا صغيرًا من أمثلة المدخلات والمخرجات المحلولة قبل المدخل الحقيقي، فيستنتج نموذج اللغة منها المهمة والتنسيق. لا تتغير أوزان النموذج؛ ويعني zero-shot عدم استخدام أمثلة، وone-shot استخدام مثال واحد. - [الضبط الدقيق (Fine-tuning)](https://broadrouter.com/ar/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): الضبط الدقيق (fine-tuning) هو مواصلة تدريب نموذج لغوي مدرَّب مسبقًا على مجموعة بيانات أصغر خاصة بمهمة، مثل محادثات نموذجية، فتتغير أوزانه ويؤدي تلك المهمة أو ذلك الأسلوب بموثوقية أكبر دون الحاجة إلى تعليمات أو أمثلة طويلة في كل أمر. - [سبب الانتهاء (finish_reason)](https://broadrouter.com/ar/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason (سبب الانتهاء) حقل في كل خيار ضمن استجابة إكمال المحادثة يوضّح سبب توقف النموذج عن التوليد. وفي صيغة OpenAI قيمه stop وlength وtool_calls وcontent_filter، إضافةً إلى function_call المُهمَلة. وتعبّر Messages API لدى Anthropic عن الفكرة نفسها بالحقل stop_reason. - [خطة LLM API بسعر ثابت (Flat-rate LLM plan)](https://broadrouter.com/ar/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): خطة LLM بسعر ثابت هي خطة تحتسب سعرًا ثابتًا لفترة، كالشهر، مقابل الوصول إلى واجهة API لنموذج لغة، بدل الاحتساب لكل رمز (توكن). وتحدّ هذه الخطط عادةً شيئًا آخر، مثل التزامن أو السرعة أو النماذج أو نوافذ الاستخدام، لذا تهم الحدود بقدر ما يهم السعر. - [FP8](https://broadrouter.com/ar/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 صيغة أعداد بفاصلة عائمة من 8 بت تُستخدم في التعلم العميق، ولها ترميزان: E4M3 بأربع بتات للأس وثلاث للجزء العشري، وE5M2 بخمس بتات للأس واثنتين للجزء العشري. وهي تخفض الذاكرة إلى النصف مقارنةً بصيغ 16 بت، وتعمل أصليًا على وحدات GPU الحديثة. - [عقوبة التكرار وعقوبة الحضور (Frequency/presence penalty)](https://broadrouter.com/ar/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): عقوبة التكرار (frequency penalty) وعقوبة الحضور (presence penalty) إعدادان لأخذ العينات يخفضان درجة الرموز الموجودة أصلًا في المخرجات. تُطبَّق عقوبة الحضور مرة واحدة بمجرد ظهور الرمز، وتزداد عقوبة التكرار مع كل تكرار. وفي واجهة OpenAI يتراوح كلاهما بين -2.0 و2.0، والقيمة الافتراضية 0. - [اللائحة العامة لحماية البيانات (GDPR)](https://broadrouter.com/ar/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): اللائحة العامة لحماية البيانات (GDPR، اللائحة (EU) 2016/679) هي قانون الاتحاد الأوروبي لمعالجة البيانات الشخصية. تُطبَّق منذ 25 مايو 2018، وتشمل الجهات القائمة في الاتحاد وبعض الجهات خارجه. وفي واجهات LLM API تحكم عقود المعالجين والنقل الدولي وتقليل البيانات. - [GGUF](https://broadrouter.com/ar/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF صيغة ملفات ثنائية لتخزين نماذج اللغة بغرض الاستدلال عبر بيئات تشغيل مبنية على GGML مثل llama.cpp. يحوي ملف GGUF واحد الأوزان، المكمَّمة غالبًا، مع البيانات الوصفية اللازمة لتحميل النموذج، ومنها تفاصيل المُرمِّز والبنية، ويمكن ربطه بالذاكرة لتحميل سريع. - [ساعة GPU (GPU-hour)](https://broadrouter.com/ar/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): ساعة GPU هي استخدام معالج رسوميات واحد لمدة ساعة. يفوتر مزوّدو السحابة مثيلات GPU المستأجرة بالساعة أو بالثانية، ويستخدم الباحثون ساعات GPU للتعبير عن حجم الحوسبة التي احتاجتها مهمة ما. وتدفع ثمن وقت الجهاز، سواء كان مشغولًا أم لا. - [الإسناد إلى المصادر (Grounding)](https://broadrouter.com/ar/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): الإسناد إلى المصادر (grounding) هو ربط مخرجات نموذج اللغة بمصادر معلومات محددة قابلة للتحقق، مثل المستندات المقدَّمة أو نتائج البحث أو سجلات قواعد البيانات، وتوجيهه إلى الإجابة منها. ويمكن تتبّع الإجابات المسنَدة إلى مصدرها، وهي أقل عرضة لاحتواء حقائق مختلَقة. - [ضوابط الحماية (Guardrails)](https://broadrouter.com/ar/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): ضوابط الحماية (guardrails) فحوص يجريها التطبيق حول استدعاء نموذج اللغة لإبقائه ضمن حدود مرسومة. ضوابط الإدخال تفحص ما يدخل إلى النموذج، وضوابط الإخراج تتحقق مما يخرج منه، وضوابط الأدوات تتحقق من الإجراءات قبل تنفيذها. والفحص الفاشل يحجب الطلب أو يعيد كتابته أو يضع عليه علامة. - [الهلوسة (Hallucination)](https://broadrouter.com/ar/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): الهلوسة مخرجات من النموذج اللغوي تبدو معقولة لكنها خاطئة أو مختلَقة أو غير مدعومة بالمدخلات التي أُعطيها، مثل مرجع مخترع أو رقم خاطئ أو واجهة API لا وجود لها. وسببها أن النموذج يتنبأ بالنص المرجّح ولا يتحقق من الحقائق. - [الهمزة (Hamza)](https://broadrouter.com/ar/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): الهمزة (ء) حرف عربي يمثّل الوقفة الحنجرية. تُكتب منفردة أو على كرسي: الألف (أ إ) أو الواو (ؤ) أو الياء (ئ)، أما آ فتجمع الألف والمدّة. ولأن النص المكتوب غير منتظم في رسمها، كثيرًا ما توحّد محركات البحث وأدوات تقييم الكلام أشكال الألف إلى ا. - [البحث الهجين (Hybrid search)](https://broadrouter.com/ar/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): البحث الهجين (hybrid search) طريقة استرجاع تشغّل البحث بالكلمات المفتاحية، مثل BM25، والبحث المتجهي عبر التضمينات على الاستعلام نفسه، ثم تدمج القائمتين المرتبتين في قائمة واحدة، غالبًا بدمج الرتب التبادلي (RRF) أو بمجموع موزون للدرجات بعد تطبيعها. - [مفتاح عدم التكرار (Idempotency key)](https://broadrouter.com/ar/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): مفتاح عدم التكرار (idempotency key) قيمة فريدة، غالبًا معرّف عشوائي مثل UUID، يرفقها العميل بطلب يغيّر الحالة، كدفعة أو عملية شراء. فإذا أُرسل الطلب نفسه مرة أخرى بالمفتاح نفسه، أعاد الخادم النتيجة الأصلية بدل تنفيذ الإجراء مرتين. - [رموز الصور (Image tokens)](https://broadrouter.com/ar/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): رموز الصور (image tokens) هي رموز الإدخال التي يستخدمها نموذج لغة بصري لتمثيل صورة في الطلب. يحسبها المزوّدون من حجم الصورة، غالبًا بتقسيمها إلى رقع أو مربعات ثابتة الحجم بعد تغيير مقاسها، ويحاسبون عليها كرموز إدخال النص، فتكلّف الصور الأكبر أكثر. - [مزوّد الاستدلال (Inference provider)](https://broadrouter.com/ar/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): مزوّد الاستدلال هو شركة تشغّل نماذج تعلّم الآلة، وغالبًا نماذج LLM مفتوحة الأوزان، على وحدات GPU تملكها أو تستأجرها، وتبيع الوصول إليها عبر API، يُسعَّر عادةً لكل رمز أو لكل طلب أو لكل وحدة زمنية. ويتولى التشغيل والتوسّع والعتاد. - [رموز الإدخال والإخراج (Input and output tokens)](https://broadrouter.com/ar/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): رموز الإدخال هي الرموز (التوكنات) التي يقرؤها نموذج اللغة في الطلب، ومنها أمر النظام وسجل المحادثة وتعريفات الأدوات والصور. ورموز الإخراج هي الرموز التي يولّدها في الرد، بما فيها الاستدلال المنطقي. وتسعّر واجهات LLM API كلًّا منهما على حدة عادةً، والإخراج أغلى في الغالب. - [INT4](https://broadrouter.com/ar/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 صيغة أعداد صحيحة من 4 بتات تمثّل 16 مستوى. وفي استدلال نماذج LLM تعني غالبًا تكميم الأوزان وحدها بـ 4 بتات (W4A16): تُخزَّن الأوزان بصيغة INT4 مع معاملات مقياس وتُوسَّع إلى 16 بتًا أثناء الحساب، فتستهلك نحو ربع ذاكرة أوزان الـ 16 بتًا. - [قائمة عناوين IP المسموحة (IP allow list)](https://broadrouter.com/ar/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): قائمة عناوين IP المسموحة (IP allow list) ضابط أمني يقبل الطلبات من عناوين IP أو نطاقات CIDR المدرجة فقط ويرفض كل ما عداها. وفي واجهات API تحدّ من ضرر تسرّب المفتاح، لأن المفتاح لا يعمل من شبكات غير مدرجة في القائمة. - [كسر القيود (Jailbreak) في نماذج LLM](https://broadrouter.com/ar/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): كسر القيود (jailbreak) مُدخل مصمَّم ليجعل نموذج اللغة يتجاهل تدريبه على السلامة أو القواعد التي وضعها التطبيق الذي ينشره، فيُنتج محتوى أو ينفّذ إجراءات كان سيرفضها عادةً. وتعدّ OWASP كسر القيود شكلًا من أشكال حقن الأوامر (prompt injection). - [وضع JSON (JSON mode)](https://broadrouter.com/ar/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): وضع JSON (JSON mode) إعداد في LLM API، هو response_format {"type": "json_object"} في الواجهات على طريقة OpenAI، يجعل رد النموذج JSON صالحًا قابلًا للتحليل. وبخلاف المخرجات المنظَّمة مع JSON Schema، لا يضمن المفاتيح أو الأنواع أو القيم التي يحتويها الكائن. - [مخطط JSON (JSON Schema)](https://broadrouter.com/ar/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema (مخطط JSON) مفردات مكتوبة بصيغة JSON لوصف بنية بيانات JSON والتحقق منها: الأنواع والحقول الإلزامية والقيم المسموحة والتداخل. والإصدار الحالي هو 2020-12. وتستخدمه واجهات LLM API لتعريف معاملات الأدوات والبنية الدقيقة التي يجب أن تتبعها المخرجات المنظَّمة. - [ذاكرة KV المؤقتة (KV cache)](https://broadrouter.com/ar/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): ذاكرة KV المؤقتة (KV cache) هي الذاكرة التي يخزّن فيها نموذج لغة من نوع transformer مفاتيح الانتباه وقيمه المحسوبة لكل رمز عالجه، فيعيد كل رمز جديد استخدامها بدل إعادة حساب التسلسل كله. وتكبر مع طول السياق، وكثيرًا ما تحدّ من عدد الطلبات التي تخدمها وحدة GPU. - [زمن الاستجابة (Latency)](https://broadrouter.com/ar/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): زمن الاستجابة (latency)، في واجهات API للنماذج اللغوية، هو المدة التي يستغرقها الطلب من لحظة إرساله حتى استلام آخر رمز من المخرجات. وهو مجموع زمن أول رمز وزمن التوليد، ويُذكر عادةً بالمئينات مثل p50 وp95 لا بالمتوسط. - [نموذج اللغة حَكَمًا (LLM-as-a-judge)](https://broadrouter.com/ar/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): نموذج اللغة حَكَمًا (LLM-as-a-judge) طريقة تقييم يمنح فيها نموذج لغة درجات لمخرجات نماذج أخرى، إما بتقييم إجابة واحدة وفق معايير مكتوبة أو باختيار الأفضل بين إجابتين. وهو يقارب الحكم البشري بتكلفة أقل كثيرًا، لكن له تحيّزات معروفة، كتفضيل الإجابة الأولى أو الأطول. - [الاختبار المعياري لنماذج LLM (LLM benchmark)](https://broadrouter.com/ar/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): الاختبار المعياري لنماذج LLM (LLM benchmark) مجموعة ثابتة، عامة في الغالب، من المهام مع طريقة لاحتساب الدرجات، تُستخدم لقياس نماذج اللغة ومقارنتها في مهارة واحدة، كأسئلة المعرفة أو البرمجة أو التعرّف على الكلام. وتُظهر درجته القدرة النسبية في ذلك الاختبار، لا أداء النموذج في عملك أنت. - [موجّه LLM (LLM router)](https://broadrouter.com/ar/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): موجّه LLM (LLM router) برنامج أو خدمة تقع بين التطبيق وواجهة أو أكثر من واجهات نماذج اللغة، وتقرر لكل طلب أي نموذج أو مزوّد يتولاه، وفق قواعد مثل التكلفة أو زمن الاستجابة أو التوفر أو نوع المهمة. - [اللوغاريتمات الاحتمالية (Logprobs)](https://broadrouter.com/ar/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): logprobs هي الاحتمالات اللوغاريتمية التي أعطاها النموذج اللغوي للرموز التي ولّدها. والواجهة التي تُرجعها تُظهر، لكل رمز في المخرجات، مدى احتماله في تقدير النموذج، ويمكنها اختياريًا أن تُظهر الرموز البديلة الأرجح في الموضع نفسه. - [LoRA (التكيّف منخفض الرتبة)](https://broadrouter.com/ar/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (التكيّف منخفض الرتبة، low-rank adaptation) طريقة ضبط دقيق موفّرة للمعاملات، تجمّد أوزان النموذج المدرَّب مسبقًا وتدرّب زوجًا من المصفوفات الصغيرة منخفضة الرتبة لطبقات مختارة. ويمكن حفظ التحديث المتعلَّم محوّلًا (adapter) صغيرًا أو دمجه في الأوزان، دون إضافة زمن إلى الاستدلال. - [الضياع في المنتصف (Lost in the middle)](https://broadrouter.com/ar/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): الضياع في المنتصف (lost in the middle) هو الميل الملاحَظ لدى نماذج اللغة إلى استخدام المعلومات ذات الصلة على أفضل وجه حين تظهر قرب بداية المُدخل الطويل أو نهايته، وعلى أسوأ وجه حين تقع في وسطه. وقد سُمّي في دراسة أجراها Liu وآخرون عام 2023. - [max_tokens (الحد الأقصى لرموز المخرجات)](https://broadrouter.com/ar/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens معامل في واجهات LLM API يحدد أقصى عدد من الرموز يمكن أن يولّده النموذج في استجابة واحدة. وحين يبلغ التوليد هذا الحد يتوقف، وتُعلَّم الاستجابة بأنها مقطوعة، مثلًا بضبط finish_reason على "length". - [MCP (بروتوكول سياق النموذج)](https://broadrouter.com/ar/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (بروتوكول سياق النموذج، Model Context Protocol) بروتوكول مفتوح مبني على JSON-RPC 2.0 يوحّد طريقة اتصال تطبيقات LLM بالأدوات والبيانات الخارجية. يعرض خادم MCP الأدوات والموارد والأوامر، ويكتشفها عميل MCP داخل التطبيق المضيف وينقل استدعاءات الأدوات من النموذج إلى الخادم. - [مزيج الخبراء (Mixture of experts، MoE)](https://broadrouter.com/ar/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): مزيج الخبراء (mixture of experts، MoE) تصميم للشبكات العصبية تحوي فيه الطبقة شبكات فرعية خبيرة كثيرة، ويرسل موجّه متعلَّم كل رمز إلى عدد قليل منها فقط. ولنموذج اللغة من نوع MoE عدد معاملات كلي كبير، لكنه يستخدم مجموعة فرعية نشطة صغيرة لكل رمز. - [إيقاف دعم النموذج (Model deprecation)](https://broadrouter.com/ar/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): إيقاف دعم النموذج (model deprecation) هو إعلان مزوّد API أن نموذجًا ما سيتوقف تقديمه، مع تحديد تاريخ للسحب أو الإيقاف ونموذج بديل موصى به عادةً. وبعد ذلك التاريخ تفشل الطلبات الموجّهة إلى النموذج، لذا يجب أن تنتقل التطبيقات إلى معرّف نموذج آخر قبله. - [النموذج الاحتياطي (Model fallback)](https://broadrouter.com/ar/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): النموذج الاحتياطي (model fallback) نمط للموثوقية يُعاد فيه طلب LLM الفاشل على نموذج أو مزوّد مختلف من قائمة مرتبة. وقد يكون المحفّز خطأً أو حدّ معدل أو خطأ في طول السياق أو رفضًا بسبب سياسة المحتوى. ويعمل داخل عميلك أو في بوابة أمام الواجهات. - [ترخيص النموذج (Model license)](https://broadrouter.com/ar/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): ترخيص النموذج هو الشروط القانونية التي يجوز بموجبها استخدام أوزان النموذج وشيفرته ومخرجاته وتعديلها ومشاركتها. وتستخدم النماذج مفتوحة الأوزان تراخيص برمجيات قياسية مثل Apache 2.0 أو MIT، أو تراخيص خاصة بالنماذج قد تضيف شروطًا، مثل قيود على الخدمات التجارية الكبيرة. - [العربية الفصحى الحديثة (Modern Standard Arabic، MSA)](https://broadrouter.com/ar/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): العربية الفصحى الحديثة (MSA) هي الصيغة المعيارية من العربية المستخدمة في الكتابة والأخبار والتعليم والخطاب الرسمي في أنحاء العالم العربي. وليست لغة الحياة اليومية الأم لأحد: يتحدث الناس لهجات إقليمية، كالمصرية أو الخليجية، وينتقلون إلى الفصحى في المواقف الرسمية. - [النظام متعدد الوكلاء (Multi-agent system)](https://broadrouter.com/ar/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): النظام متعدد الوكلاء هو تطبيق LLM يعمل فيه عدة وكلاء، لكل منهم أمره وأدواته وسياقه، على أجزاء من مهمة واحدة. ومن التصاميم الشائعة نمط المنسّق والعاملين: وكيل رئيسي يخطط، ويطلق وكلاء عاملين بالتوازي، ويجمع نتائجهم. - [الذكاء الاصطناعي متعدد الوسائط (Multimodal AI)](https://broadrouter.com/ar/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): الذكاء الاصطناعي متعدد الوسائط هو نموذج أو نظام يتعامل مع أكثر من نوع من البيانات، مثل النص والصور والفيديو والصوت. فقد يقرأ النموذج متعدد الوسائط صورة وسؤالًا معًا، أو يستقبل كلامًا ويجيب نصًا. والنموذج اللغوي البصري هو أشهر أنواعه. - [التعرّف على الكيانات المسماة (NER)](https://broadrouter.com/ar/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): التعرّف على الكيانات المسماة (NER) هو مهمة إيجاد الأسماء في النص ووسم كل منها بنوع، مثل شخص أو منظمة أو مكان. تَسِم الأنظمة التقليدية كل رمز، أما نماذج LLM فيُطلب منها غالبًا إرجاع قائمة بالكيانات بصيغة JSON، تُطابَق بعد ذلك مع النص المصدر. - [التعرّف الضوئي على الحروف (OCR)](https://broadrouter.com/ar/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): التعرّف الضوئي على الحروف (OCR) تقنية تكتشف النص المطبوع أو المكتوب بخط اليد في صورة أو مستند ممسوح أو ملف PDF وتحوّله إلى حروف قابلة للقراءة آليًا. يُخرج OCR التقليدي النص وموضعه في الصفحة، أما النماذج اللغوية البصرية فتقرأ النص وتجيب عن أسئلة حوله مباشرة. - [Ogg Opus (صيغة الرسائل الصوتية في واتساب)](https://broadrouter.com/ar/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus هو صوت مرمَّز بترميز Opus (RFC 6716) ومحفوظ داخل حاوية Ogg (RFC 7845). صُمّم للكلام وللصوت التفاعلي بمعدلات بت منخفضة، وهو الصيغة التي يستخدمها واتساب للرسائل الصوتية، عادةً في ملفات .ogg أو .opus. - [النموذج مفتوح الأوزان (Open-weight model)](https://broadrouter.com/ar/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): النموذج مفتوح الأوزان نموذج تعلّم آلي تُنشر أوزانه المدرَّبة ليحمّلها أي أحد ويشغّلها، عادةً بموجب ترخيص له شروط. ويختلف عن الذكاء الاصطناعي مفتوح المصدر بالكامل، الذي يُصدر أيضًا شيفرة التدريب ومعلومات مفصّلة عن بيانات التدريب، ليتمكن الآخرون من دراسة النموذج وإعادة إنتاجه. - [واجهة API متوافقة مع OpenAI (OpenAI-compatible API)](https://broadrouter.com/ar/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): واجهة API المتوافقة مع OpenAI هي واجهة HTTP تقبل صيغ الطلبات والاستجابات نفسها التي تستخدمها OpenAI API، مثل POST /v1/chat/completions، فتستطيع حزم OpenAI SDK وأدواتها استخدامها بتغيير العنوان الأساسي (base URL) ومفتاح API واسم النموذج فقط. - [استدعاءات الأدوات المتوازية (Parallel tool calls)](https://broadrouter.com/ar/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): استدعاءات الأدوات المتوازية هي أن يعيد النموذج أكثر من استدعاء أداة في استجابة واحدة، مثل قراءة ثلاثة ملفات دفعة واحدة. وفي صيغة OpenAI تصل على شكل عدة عناصر في tool_calls، وضبط parallel_tool_calls على false يقصر النموذج على صفر أو استدعاء واحد في كل دور. - [الدفع حسب الاستخدام (Pay-as-you-go)](https://broadrouter.com/ar/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): الدفع حسب الاستخدام هو نموذج تسعير تدفع فيه مقابل ما تستخدمه فقط، مقيسًا بوحدات مثل الرموز (التوكنات) أو الطلبات أو دقائق الصوت، دون اشتراك أو حد أدنى للالتزام. وتحتسب معظم واجهات LLM API تكلفتها بهذه الطريقة، من رصيد مدفوع مسبقًا أو بفاتورة دورية. - [نظام حماية البيانات الشخصية (PDPL) في السعودية والإمارات](https://broadrouter.com/ar/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) هو الاسم الشائع لنظام حماية البيانات الشخصية في السعودية، الصادر بالمرسوم الملكي م/19 وتطبّقه سدايا (SDAIA)، وللمرسوم بقانون اتحادي رقم 45 لسنة 2021 في الإمارات. وكلاهما ينظّم معالجة البيانات الشخصية ويضع شروطًا لنقلها إلى الخارج، بما في ذلك الأوامر المرسلة إلى واجهات LLM API أجنبية. - [معلومات التعريف الشخصية (PII)](https://broadrouter.com/ar/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): معلومات التعريف الشخصية (PII) هي معلومات يمكن استخدامها لتمييز هوية شخص أو تتبّعها، مثل الاسم أو رقم الهوية أو السجل البيومتري، إضافةً إلى معلومات أخرى مرتبطة بذلك الشخص أو قابلة للربط به. ومصطلح GDPR الأوسع هو البيانات الشخصية. وفي أعمال LLM يزيل حجب PII هذه المعلومات من الأوامر قبل أن تغادر أنظمتك. - [مرحلتا الملء المسبق وفك الترميز (Prefill and decode)](https://broadrouter.com/ar/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): الملء المسبق (prefill) وفك الترميز (decode) هما مرحلتا الاستدلال في نماذج LLM. في الملء المسبق يعالج النموذج الأمر كله بالتوازي ويملأ ذاكرة KV المؤقتة؛ وفي فك الترميز يولّد الرد رمزًا تلو الآخر، مستفيدًا من تلك الذاكرة. يحدد الملء المسبق في الغالب زمن أول رمز، ويحدد فك الترميز سرعة الإخراج. - [الرصيد المدفوع مسبقًا (Prepaid credit)](https://broadrouter.com/ar/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): الرصيد المدفوع مسبقًا هو مال يُضاف إلى حساب API قبل الاستخدام. يُخصم من الرصيد ثمن كل طلب، وعندما ينفد الرصيد تتوقف الطلبات المحتسبة بالاستهلاك حتى تشحن رصيدك. وهو يحصر الإنفاق في حدود ما دفعته. - [هندسة الأوامر (Prompt engineering)](https://broadrouter.com/ar/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): هندسة الأوامر (prompt engineering) هي عملية كتابة التعليمات المقدَّمة إلى النموذج اللغوي وتنظيمها واختبارها، حتى يُنتج باستمرار مخرجات تفي بمتطلباتك. وتشمل رسائل الأدوار، والتعليمات الواضحة، والأمثلة، والتنسيق، وقياس النتائج على حالات اختبار. - [حقن الأوامر (Prompt injection)](https://broadrouter.com/ar/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): حقن الأوامر هجوم يغيّر فيه نصٌّ يُقدَّم إلى نموذج لغوي، من مستخدم أو مخبّأ في صفحة ويب أو بريد إلكتروني أو ملف أو نتيجة أداة، سلوكَ النموذج خلافًا لما أراده المطوّر. وتضعه OWASP في المرتبة الأولى (LLM01:2025) في قائمتها Top 10 لتطبيقات LLM. - [التكميم (Quantization)](https://broadrouter.com/ar/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): التكميم (quantization) هو تخزين أرقام الشبكة العصبية، وأوزانها أساسًا، بصيغة أقل دقة مثل 8 بتات أو 4 بتات بدل 16 بتًا، فيستهلك النموذج ذاكرة أقل ويعمل غالبًا أسرع. والطرق الجيدة تُبقي فقدان الدقة صغيرًا. - [RAG (التوليد المعزّز بالاسترجاع)](https://broadrouter.com/ar/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG (التوليد المعزّز بالاسترجاع) نمط يسترجع فيه التطبيق أولًا مقاطع متصلة بالسؤال من مخزن خارجي، مثل فهرس بحث أو فهرس متجهات، ثم يضيفها إلى الأمر ليجيب النموذج اللغوي من ذلك النص لا من ذاكرته وحدها. - [حد المعدل (Rate limit)](https://broadrouter.com/ar/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): حد المعدل هو سقف تفرضه واجهة API على مقدار استخدام العميل لها خلال فترة زمنية، مثل عدد الطلبات في الدقيقة أو الرموز في الدقيقة. وتُرفض الطلبات التي تتجاوزه، غالبًا بالخطأ HTTP 429 Too Many Requests، إلى أن يعود الاستخدام تحت الحد. - [ReAct (الاستدلال ثم الفعل)](https://broadrouter.com/ar/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (الاستدلال والفعل، reason and act) نمط تلقين قدّمه Yao وآخرون عام 2022، يمزج فيه النموذج اللغوي بين مسارات الاستدلال والإجراءات، مثل البحث أو استدعاء الأدوات. يقرأ النموذج كل ملاحظة، ويحدّث خطته، ويكرر ذلك حتى يستطيع تقديم إجابة نهائية. - [رموز الاستدلال المنطقي (Reasoning tokens)](https://broadrouter.com/ar/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): رموز الاستدلال المنطقي هي رموز (توكنات) يولّدها نموذج الاستدلال ليحلّ المسألة قبل أن يكتب إجابته الظاهرة. تُحتسب عادةً رموز إخراج، وتدخل ضمن حد الإخراج ونافذة السياق، وقد تخفيها الواجهة أو تلخّصها. - [مهلة الطلب (Request timeout)](https://broadrouter.com/ar/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): مهلة الطلب (request timeout) هي أطول مدة ينتظرها عميل HTTP ردَّ الخادم قبل أن يتخلى عن الطلب. وعادةً يقسمها العملاء إلى مهلة اتصال لفتح الاتصال ومهلة قراءة لانتظار البيانات. وتحتاج استدعاءات LLM إلى مهلات قراءة سخية لأن الإجابات الطويلة تستغرق وقتًا. - [إعادة الترتيب (Reranking)](https://broadrouter.com/ar/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): إعادة الترتيب مرحلة استرجاع ثانية، يقيّم فيها نموذج أدق لكنه أبطأ، عادةً cross-encoder، كل زوج من الاستعلام والمستند المرشح القادم من بحث أوّلي، ثم يعيد ترتيبها، فلا يصل إلى النموذج اللغوي إلا القليل الأكثر صلة. - [النص من اليمين إلى اليسار (RTL)](https://broadrouter.com/ar/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): النص من اليمين إلى اليسار (RTL) هو كتابة، كالعربية أو العبرية، تُعرض من اليمين إلى اليسار. تخزّنه الحواسيب بترتيب القراءة المنطقي، وتحدّد خوارزمية Unicode ثنائية الاتجاه ترتيب العرض عندما يختلط بنص من اليسار إلى اليمين أو بأرقام أو أسماء نماذج أو شيفرة. - [معدل أخذ العينات (Sample rate)](https://broadrouter.com/ar/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): معدل أخذ العينات هو عدد عينات الصوت الملتقطة في الثانية، ويُقاس بالهرتز (Hz). تتوقع نماذج التعرّف على الكلام عادةً صوتًا أحادي القناة بمعدل 16,000 Hz (16 kHz). ويحدد معدل العينات، مع عمق البت وعدد القنوات، حجم الملف الصوتي غير المضغوط في الدقيقة. - [معامل البذرة (seed)](https://broadrouter.com/ar/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): معامل البذرة (seed) عدد صحيح ترسله مع الطلب لتحاول الواجهة أخذ عينات الرموز نفسها حين يتكرر الطلب نفسه. وتصفه OpenAI بأنه أفضل جهد ممكن: فالحتمية غير مضمونة، ويُظهر الحقل system_fingerprint متى تغيّر إعداد الخادم الخلفي. - [الاستدلال بلا خوادم (Serverless inference)](https://broadrouter.com/ar/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): الاستدلال بلا خوادم طريقة لتشغيل نماذج تعلّم الآلة تبدأ فيها المنصة موارد الحوسبة خلف نقطة النهاية وتوسّعها وتوقفها تلقائيًا، وغالبًا حتى الصفر عند الخمول، وتحاسبك على الاستخدام لا على أجهزة محجوزة. والثمن هو احتمال تأخير البدء البارد. - [اتفاقية مستوى الخدمة (SLA)](https://broadrouter.com/ar/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): اتفاقية مستوى الخدمة (SLA) التزام تعاقدي من المزوّد بمستوى خدمة قابل للقياس، عادةً نسبة توفر شهرية، مع تعويض محدد، عادةً أرصدة خدمة، إذا لم يتحقق. وتحدد الاتفاقية ما يُعدّ توقفًا، وما هو مستثنى، وكيفية المطالبة. - [فك الترميز التخميني (Speculative decoding)](https://broadrouter.com/ar/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): فك الترميز التخميني تقنية استدلال يقترح فيها نموذج مسوّدة صغير وسريع عدة رموز تالية، ثم يتحقق منها النموذج الكبير المستهدف في مرور أمامي واحد ويُبقي المقبول منها. وهي تسرّع التوليد دون أن تغيّر توزيع المخرجات. - [تحويل الصوت إلى نص (Speech-to-text)](https://broadrouter.com/ar/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): تحويل الكلام إلى نص، ويسمّى أيضًا التعرّف التلقائي على الكلام (ASR)، هو تقنية تحوّل الصوت المنطوق إلى نص مكتوب. تستقبل واجهة تحويل الصوت إلى نص ملفًا صوتيًا أو بثًا وتُرجع نصًا مفرّغًا، يُسعَّر عادةً لكل دقيقة صوت. وتُقاس الدقة بمعدل الخطأ في الكلمات، وتتفاوت بحسب اللكنة واللهجة وجودة التسجيل. - [حد الإنفاق (Spend limit)](https://broadrouter.com/ar/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): حد الإنفاق سقف لما يجوز أن ينفقه حساب API أو مشروع أو مفتاح في فترة ما، عادةً شهر. وحين يبلغه الاستخدام، يرفض المزوّد الطلبات اللاحقة، غالبًا بخطأ HTTP 429 أو 400، حتى تبدأ الفترة من جديد أو يُرفع الحد. - [ملفات الترجمة SRT (SRT subtitles)](https://broadrouter.com/ar/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) صيغة نصية بسيطة للترجمة المصاحبة للفيديو. تتكون كل تسمية من رقم تسلسلي، ووقت بداية ونهاية بالشكل hh:mm:ss,mmm --> hh:mm:ss,mmm، وسطر نص أو أكثر، ثم سطر فارغ. وWebVTT صيغة ويب مشابهة يستخدمها الفيديو في HTML. - [الأحداث التي يرسلها الخادم (SSE)](https://broadrouter.com/ar/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): الأحداث التي يرسلها الخادم (server-sent events أو SSE) معيار ويب يتيح للخادم دفع تدفق من الأحداث النصية إلى العميل عبر استجابة HTTP واحدة طويلة الأمد بنوع المحتوى text/event-stream. وكل حدث سطر حقل أو أكثر، مثل data:، ينتهي بسطر فارغ. - [تسلسلات التوقف (Stop sequences)](https://broadrouter.com/ar/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): تسلسلات التوقف هي نصوص ترسلها مع الطلب فيتوقف النموذج عن التوليد فور أن يُنتج أحدها. تسمّي OpenAI الحقل stop وتقبل فيه 4 تسلسلات كحد أقصى، وتسمّيه Anthropic باسم stop_sequences. وعادةً لا يُضمَّن النص المطابق في المخرجات المُعادة. - [البث المتدفق (Streaming)](https://broadrouter.com/ar/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): البث المتدفق (streaming) طريقة لإرجاع استجابة النموذج اللغوي على أجزاء صغيرة أثناء توليدها، بدل إرجاعها ردًا واحدًا في النهاية. وتبث الواجهات المتوافقة مع OpenAI عبر أحداث يرسلها الخادم (server-sent events) حين يضبط الطلب stream على true. - [المخرجات المنظَّمة (Structured output)](https://broadrouter.com/ar/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): المخرجات المنظَّمة هي ميزة في LLM API تقيّد رد النموذج بصيغة يحدّدها المطوّر، عادةً JSON Schema، لكي تحلّلها الشيفرة بموثوقية. وهي أكثر صرامة من وضع JSON (JSON mode)، الذي يضمن JSON صالحًا لكن لا يضمن بنية معيّنة. - [الوكيل الفرعي (Subagent)](https://broadrouter.com/ar/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): الوكيل الفرعي وكيل منفصل يشغّله الوكيل الرئيسي لتنفيذ مهمة واحدة محددة. يعمل بنافذة سياق وأمر نظام وصلاحيات أدوات خاصة به، وينجز العمل، ثم لا يُرجع إلى الوكيل الرئيسي إلا نتيجة أو ملخصًا، فتبقى المحادثة الرئيسية صغيرة. - [المعالج الفرعي (Subprocessor)](https://broadrouter.com/ar/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): المعالج الفرعي طرف ثالث يستعين به المعالج لتنفيذ معالجة البيانات الشخصية نيابةً عن المتحكم. وفي واجهة LLM API قد يشمل المعالجون الفرعيون مستضيفي GPU ومزوّدي استدلال شركاء وسعة احتياطية. وبموجب المادة 28 من GDPR يحتاجون إلى تفويض المتحكم وإلى التزامات حماية البيانات نفسها. - [أمر النظام (System prompt)](https://broadrouter.com/ar/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): أمر النظام (system prompt) هو نص التعليمات الذي يرسله التطبيق إلى النموذج اللغوي قبل المحادثة، عادةً في رسالة بالدور system أو developer. وهو يحدد دور النموذج وقواعده ونبرته وصيغة مخرجاته، ويُفترض أن يلتزم به النموذج إذا تعارض مع رسائل المستخدم. - [التاء المربوطة (ة، Taa marbuta)](https://broadrouter.com/ar/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): التاء المربوطة (ة، U+0629) حرف عربي لا يظهر إلا في آخر الكلمة، ويدل غالبًا على تأنيث الأسماء والصفات. وشكلها شكل الهاء (ه، U+0647) مع نقطتين، فيختلط الحرفان في النص المكتوب، ولهذا تحوّل كثير من أدوات توحيد النص للبحث ة إلى ه. - [درجة الحرارة (Temperature)](https://broadrouter.com/ar/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): درجة الحرارة (temperature) معامل لأخذ العينات يتحكم في مدى عشوائية مخرجات النموذج اللغوي. يقسم النموذج درجات الرموز على قيمة الحرارة قبل تحويلها إلى احتمالات: فالقيم المنخفضة تجعل الرمز الأرجح هو المهيمن، والقيم المرتفعة تسطّح التوزيع فتُختار الرموز الأقل احتمالًا أكثر. - [تحويل النص إلى كلام (TTS)](https://broadrouter.com/ar/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): تحويل النص إلى كلام (TTS)، أو تركيب الكلام، يحوّل النص المكتوب إلى صوت منطوق. تستقبل واجهة TTS نصًا وصوتًا وصيغة مخرجات مثل MP3 أو WAV، وتُرجع كلامًا مولَّدًا. وهو عكس تحويل الصوت إلى نص، الذي يحوّل الصوت إلى نص. - [الإنتاجية (Throughput)](https://broadrouter.com/ar/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): الإنتاجية (throughput) في تقديم نماذج LLM هي إجمالي العمل الذي ينجزه النظام في وحدة الزمن عبر كل الطلبات المتزامنة، وتُقاس عادةً برموز الإخراج في الثانية أو الطلبات في الثانية. وهي تختلف عن سرعة التدفق الواحد، أي مدى سرعة وصول الرموز إلى طلب واحد. - [زمن أول رمز (TTFT)](https://broadrouter.com/ar/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): زمن أول رمز (TTFT) هو المدة من إرسال طلب إلى النموذج اللغوي حتى وصول أول رمز من المخرجات. ويشمل زمن الشبكة، والانتظار في طابور الخادم، والمعالجة المسبقة (prefill)، وهي الخطوة التي يقرأ فيها النموذج الأمر كاملًا، لذا يزداد مع طول الأمر والضغط على الخادم. - [الطوابع الزمنية في التفريغ الصوتي (Timestamps)](https://broadrouter.com/ar/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): الطوابع الزمنية في تحويل الصوت إلى نص هي أوقات بداية ونهاية، بالثواني، تُرفق بأجزاء من النص المفرّغ. تغطي طوابع المقاطع عبارة أو جملة؛ وتغطي طوابع الكلمات كل كلمة. وهي لازمة للترجمة المصاحبة للفيديو، وللبحث داخل التسجيلات، ولمطابقة النص مع المتحدثين. - [الرمز (التوكن، Token)](https://broadrouter.com/ar/glossary/token/) (translation of https://broadrouter.com/glossary/token/): الرمز (التوكن) هو وحدة النص التي يقرؤها نموذج اللغة ويكتبها: كلمة، أو جزء من كلمة، أو علامة ترقيم، أو تسلسل بايتات ينتجه مجزّئ النموذج (tokenizer). وتَعُدّ واجهات LLM API الاستهلاك بالرموز، وتضع حدودها بها، وتحتسب تكلفتها بها عادةً. - [المجزّئ (Tokenizer)](https://broadrouter.com/ar/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): المجزّئ (tokenizer) هو المكوّن الذي يحوّل النص إلى سلسلة معرّفات رموز عددية يقرؤها نموذج اللغة، ويحوّل المعرّفات المولَّدة إلى نص من جديد. ولكل نموذج مجزّئه الخاص، فقد يُنتج النص نفسه عددًا مختلفًا من الرموز (التوكنات) في نماذج مختلفة. - [الرموز في الثانية (TPS)](https://broadrouter.com/ar/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): الرموز في الثانية (TPS) هي معدل توليد النموذج اللغوي لرموز الإخراج، وتُقاس عادةً لطلب واحد بعد وصول أول رمز. وهي تحدّد سرعة تدفق الرد بعد بدئه. وحين تُجمع عبر كل الطلبات المتزامنة تصبح إنتاجية النظام؛ أما الانتظار قبل أول رمز فيُقاس منفصلًا. - [استدعاء الأدوات (Tool calling)](https://broadrouter.com/ar/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): استدعاء الأدوات، ويُسمّى أيضًا استدعاء الدوال (function calling)، ميزة في واجهات LLM API يُرجع فيها النموذج طلبًا منظَّمًا لتشغيل دالة عرّفها المطوّر، مع معاملات بصيغة JSON. ويشغّل التطبيق الدالة ويعيد النتيجة؛ فالنموذج لا يشغّل الشيفرة بنفسه أبدًا. - [tool_choice (اختيار الأداة)](https://broadrouter.com/ar/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice معامل في الطلب يتحكم في استخدام الأدوات. في Chat Completions من OpenAI قد تكون قيمته auto (يقرر النموذج)، أو required (يجب أن يستدعي أداة أو أكثر)، أو none (بلا أدوات)، أو اسم دالة لفرض استدعائها. والخيارات المقابلة لدى Anthropic هي auto وany وtool وnone. - [Top-p (أخذ العينات من النواة)](https://broadrouter.com/ar/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Top-p، ويُسمّى أيضًا أخذ العينات من النواة (nucleus sampling)، طريقة فك ترميز يختار فيها النموذج اللغوي الرمز التالي من أصغر مجموعة من الرموز الأرجح فقط، تلك التي يبلغ مجموع احتمالاتها عتبة p مثل 0.9، ويستبعد الذيل قليل الاحتمال. - [توحيد Unicode (Unicode normalization)](https://broadrouter.com/ar/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): توحيد Unicode هو تحويل النص إلى صيغة معيارية بحيث تستخدم السلاسل التي تبدو متطابقة نقاط الترميز نفسها. والصيغ الأربع هي NFD وNFC وNFKD وNFKC؛ وصيغتا K تطويان أيضًا محارف التوافق، مثل أشكال العرض العربية والحروف المركّبة، إلى حروف عادية. - [كشف النشاط الصوتي (VAD)](https://broadrouter.com/ar/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): كشف النشاط الصوتي (VAD) خطوة لمعالجة الإشارة تصنّف إطارات قصيرة من الصوت إلى كلام أو غير كلام. تستخدمه أنظمة الكلام لتخطي الصمت، وتقسيم التسجيلات الطويلة عند فترات التوقف، ومعرفة متى أنهى المتحدث كلامه، وتجنّب تفريغ ضوضاء الخلفية. - [قاعدة بيانات المتجهات (Vector database)](https://broadrouter.com/ar/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): قاعدة بيانات المتجهات مخزن بيانات يحفظ متجهات التضمين، ومعها غالبًا نصها المصدر وبياناتها الوصفية، ويجيب عن استعلامات أقرب الجيران: تعطيها متجه استعلام فتُرجع المتجهات المخزّنة الأشبه به، عادةً بمسافة جيب التمام أو المسافة الإقليدية أو الضرب الداخلي. - [نموذج الرؤية واللغة (Vision-language model)](https://broadrouter.com/ar/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): نموذج الرؤية واللغة (VLM) نموذج يستقبل الصور والنص معًا مدخلاتٍ ويُنتج نصًا، فيستطيع وصف الصور والإجابة عن أسئلة حولها وقراءة المستندات والرسوم البيانية ولقطات الشاشة. وتُسمّى هذه النماذج أيضًا نماذج image-text-to-text أو النماذج متعددة الوسائط. - [vLLM](https://broadrouter.com/ar/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM مكتبة مفتوحة المصدر للاستدلال بالنماذج اللغوية الكبيرة وتقديمها، طُوّرت أولًا في مختبر Sky Computing بجامعة UC Berkeley. تدير ذاكرة KV المؤقتة (KV cache) بتقنية PagedAttention، وتجمّع الطلبات الواردة باستمرار، وتدعم صيغ تكميم كثيرة، وتوفّر خادم HTTP متوافقًا مع OpenAI. - [حجم المفردات (Vocabulary size)](https://broadrouter.com/ar/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): حجم المفردات هو عدد الرموز (التوكنز) المختلفة في مجزّئ النموذج اللغوي، لكل منها معرّف وتضمين خاص. والمفردات الأكبر قد ترمّز النص الشائع برموز أقل، لكنها تكبّر طبقتي التضمين والإخراج، ولا تفيد إلا مع النص الذي دُرّبت عليه المفردات. - [ملف WAV](https://broadrouter.com/ar/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format، أو WAVE) حاوية صوتية مبنية على صيغة RIFF من Microsoft وIBM. تحمل عادةً عينات PCM خطية غير مضغوطة، فيكون حجمها متوقَّعًا: معدل العينات مضروبًا في عدد بايتات العينة مضروبًا في عدد القنوات، لكل ثانية. - [معدل الخطأ في الكلمات (WER)](https://broadrouter.com/ar/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): معدل الخطأ في الكلمات (WER) هو مقياس الدقة المعياري للتعرّف على الكلام: عدد استبدالات الكلمات وحذوفاتها وإدراجاتها اللازمة لتحويل النص المفرّغ إلى النص المرجعي الصحيح، مقسومًا على عدد كلمات المرجع. والأقل أفضل. - [عدم الاحتفاظ بالبيانات (Zero data retention)](https://broadrouter.com/ar/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): عدم الاحتفاظ بالبيانات (ZDR) هو سياسة للتعامل مع البيانات يعالج فيها مزوّد الذكاء الاصطناعي الأوامر والمخرجات لخدمة الطلب فقط، ولا يخزّن هذا المحتوى بعده في السجلات أو الذاكرة المؤقتة أو بيانات التدريب. وتظل السجلات الخالية من المحتوى، مثل أعداد الرموز والتوقيتات، تُحفظ عادةً. ### Spanish (es, Español): 294 translated pages - [Articles in Spanish](https://broadrouter.com/es/blog/) - [Precios de API de LLM en octubre de 2026: 3 subidas, ninguna rebaja](https://broadrouter.com/es/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 cambios en las páginas de precios de 5 de 7 proveedores de API de LLM en octubre de 2026: 3 subidas, ninguna rebaja, 9 modelos con precio nuevo y 11 retirados. - [Cuánto cuesta ejecutar un agente de IA: pasos, contexto y caché](https://broadrouter.com/es/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): Por qué cada paso de un agente de programación cuesta más cuanto más dura la ejecución, cómo cambia la caché el total y cuándo conviene una sesión plana. - [Añadir diacríticos al texto árabe con un LLM: prompt, controles y DER](https://broadrouter.com/es/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): Restaure los diacríticos árabes (tashkeel) con una API de LLM: un prompt que conserva cada letra, un control que rechaza ediciones, DER y WER, y cuándo usar CATT. - [Corrección gramatical y ortográfica del árabe con un LLM: cómo hacerla](https://broadrouter.com/es/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Corrija la ortografía y la gramática del árabe con una API de LLM: prompt de edición mínima, protección contra reescrituras, ediciones por tipo y evaluación M2. - [Actas de reuniones en árabe a partir de grabaciones: pipeline con API](https://broadrouter.com/es/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Convierta grabaciones de reuniones en árabe en actas con decisiones y tareas: transcriba, limpie, resuma con un LLM en JSON y verifique nombres y fechas. - [Extracción de entidades nombradas en árabe con un LLM: JSON y fechas](https://broadrouter.com/es/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Extraiga a JSON nombres, organizaciones, lugares y fechas hégira en árabe con un LLM: prompt, verificación que ignora diacríticos, conversión de fechas y evaluación. - [Comparativa de OCR árabe: Google, Azure, Textract, Mistral y LLM](https://broadrouter.com/es/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): Qué API de OCR admiten oficialmente el árabe, impreso y manuscrito, qué devuelve cada una y cuánto cuestan 1,000 páginas, comparadas con API de LLM de visión. - [Subtítulos en árabe (SRT) desde una API de voz a texto: qué necesita](https://broadrouter.com/es/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): Qué necesita un archivo SRT, qué APIs de voz a texto devuelven marcas de tiempo o SRT para árabe y un pipeline probado que crea subtítulos con una API solo de texto. - [Clasificar texto en árabe con un LLM: sentimiento, intención, triaje](https://broadrouter.com/es/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): Clasifique reseñas, mensajes y tickets en árabe con una API de LLM: diseño de etiquetas, prompt por dialecto, salida JSON, pruebas con macro F1 y costo a escala. - [Normalización de texto árabe para búsqueda y RAG: qué cambiar y qué no](https://broadrouter.com/es/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Tashkeel, alef, ta marbuta, hamza, letras persas, dígitos y Unicode: qué reglas de normalización del árabe ayudan a la búsqueda, cuáles perjudican y código Python. - [Resumir textos en árabe con un LLM: prompts, extensión y evaluación](https://broadrouter.com/es/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): Resuma artículos, informes y mensajes en dialecto árabe con una API de LLM: prompts, control de palabras, revisión de cifras y por qué ROUGE da 0 al árabe. - [Alternativas a Artificial Analysis: dónde comparar LLM en 2026](https://broadrouter.com/es/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Alternativas a Artificial Analysis: Arena, Epoch AI, Vals AI, rankings de OpenRouter, Hugging Face y datos de precios, según lo que mide cada una y cuándo usarla. - [El hardware GPU de BroadNet y cómo sirve los modelos, según BroadNet](https://broadrouter.com/es/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): Qué dice BroadNet sobre el hardware de BroadNet Inference: tres niveles de GPU, de B300 a equipos en el cliente, rendimiento por GPU y lo aprendido en sus pruebas. - [Caché de respuestas de APIs de LLM para reducir costos en Python](https://broadrouter.com/es/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Caché de coincidencia exacta probada para respuestas de APIs de LLM en Python: qué va en la clave, qué no cachear, TTL, privacidad y cómo medir la tasa de aciertos. - [Claude Code con modelos abiertos: proxies, endpoints nativos y límites](https://broadrouter.com/es/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code habla la Messages API de Anthropic. Cómo ejecutarlo con modelos abiertos mediante un proxy LiteLLM o un endpoint nativo, y qué falla por el camino. - [Integre la revisión humana en su flujo de IA documental](https://broadrouter.com/es/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): Diseñe el paso de revisión humana de un pipeline de IA documental: reglas de validación, qué enviar a una persona, muestras de auditoría, registro y tasa. - [Ley de protección de datos de Egipto (Ley 151 de 2020) y API de LLM](https://broadrouter.com/es/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Enviar datos personales de egipcios a una API de LLM o de voz a texto extranjera: qué dicen la Ley 151 de 2020 y la Resolución 816 de 2025 sobre transferencias. - [Antes del fine-tuning: defina el problema que quiere resolver](https://broadrouter.com/es/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): Cuándo el prompting, RAG o la evaluación resuelven el problema sin fine-tuning, qué datos de entrenamiento necesita y cómo probar un fine-tune antes de pedirlo. - [Cómo corregir JSON inválido de un LLM: causas, reparación y reintento](https://broadrouter.com/es/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): Por qué un LLM devuelve JSON roto (bloques de código, texto extra, cortes por max_tokens, comas finales) y una función Python probada que solo repara lo seguro. - [¿Cada cuánto cambian los precios de las API de LLM? 3 años de archivo](https://broadrouter.com/es/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): 308 páginas de precios archivadas de 11 proveedores de API de LLM (2023–2026): el 80 % de los modelos nunca cambió de precio y DeepInfra hizo el 75 % de los cambios. - [Cómo reducir los costos de una API de LLM: 10 pasos, en orden](https://broadrouter.com/es/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): Reduzca su factura de API de LLM en orden: medir, cachear el prefijo, recortar prompts, limitar la salida, elegir modelo y saber cuándo gana la tarifa plana. - [Failover de API de LLM: pase a un segundo proveedor desde su código](https://broadrouter.com/es/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): Cómo hacer failover entre proveedores de LLM compatibles con OpenAI: qué errores lo activan, circuit breaker, mapeo de IDs de modelo, streaming, costo y privacidad. - [Rebajas de precios de API de LLM 2023–2026: las mayores, con fuentes](https://broadrouter.com/es/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): Las mayores rebajas de precios de API de LLM de 2023 a 2026, con enlaces de Internet Archive de antes y después: o3, Codestral, Gemini 1.5 y subidas poco conocidas. - [¿Las API de LLM solo bajan de precio? 52 subidas y su presupuesto](https://broadrouter.com/es/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): Los precios de las API de LLM sí suben: 52 subidas del mismo modelo en páginas archivadas de 2024 a 2026, mediana del 40 %. Fuentes y regla de presupuesto. - [Límites de API de LLM por proveedor: RPM, TPM y concurrencia](https://broadrouter.com/es/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): Límites del nivel de entrada de 13 proveedores de API de LLM según su documentación: OpenAI, Anthropic, Gemini, Groq, DeepSeek y otros, y qué significan RPM y TPM. - [SLA de APIs de LLM: qué prometen Azure, Bedrock, Google y OpenAI](https://broadrouter.com/es/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): A qué se compromete por escrito cada proveedor de API de LLM: porcentaje de disponibilidad, qué cuenta como caída, créditos de servicio y objetivos de latencia. - [Procesamiento por lotes con LLM sin Batch API: un runner reanudable](https://broadrouter.com/es/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Procese de 10,000 a 1,000,000 de prompts con una API de chat completions sin Batch API: concurrencia, ritmo, reintentos, checkpoints y cálculo de tiempos. - [El Safety Map: cómo tratan 39 LLM los datos sensibles](https://broadrouter.com/es/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet evaluó 39 LLM de 14 laboratorios con credenciales, PII, datos de salud y financieros: qué se transfirió, qué no, y las fugas por tool calls y razonamiento. - [Cómo evaluar la seguridad de un LLM con datos sensibles (BroadNet)](https://broadrouter.com/es/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): Método de BroadNet para comprobar si un LLM filtra credenciales, PII, datos de salud o financieros antes del despliegue: categorías, fuga vs omisión, 3+ ejecuciones. - [Migrar desde la API de OpenAI: el código en Python, Node y cURL](https://broadrouter.com/es/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): Los cambios de código para pasar de OpenAI a una API compatible: cliente, de Responses a Chat Completions, herramientas, streaming, embeddings y una prueba rápida. - [Modelos que BroadNet evaluó y descartó, y por qué](https://broadrouter.com/es/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet publica los modelos que probó y rechazó, con motivos: modelos de chat que fallaron su control de credenciales y modelos de voz, OCR y embeddings superados. - [Cómo diseñar una experiencia de IA para varios idiomas](https://broadrouter.com/es/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): Diseñe un producto de IA para usuarios en árabe, inglés y español: idioma de la interfaz vs. del modelo, RTL y texto mixto, números, fechas, detección y costo. - [Costo de API de modelos abiertos vs. cerrados: lo que dicen los datos](https://broadrouter.com/es/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): ¿Los modelos de pesos abiertos son más baratos por API que GPT, Claude o Gemini? Precios por token de cada proveedor, diferencias entre ellos y costo por usuario. - [Historial de precios de API: OpenAI vs Anthropic vs Google, 2023–2026](https://broadrouter.com/es/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): Precios de las API de GPT, Claude y Gemini, lado a lado, en páginas archivadas de 2023 a 2026: qué líneas se abarataron y cuáles se encarecieron, con su archivo. - [URL base de APIs compatibles con OpenAI: 14 proveedores verificados](https://broadrouter.com/es/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): La base_url de 14 APIs de LLM compatibles con OpenAI, tomada de la documentación de cada proveedor, con sus endpoints y los detalles que impiden un cambio directo. - [OpenClaw y Hermes Agent con un modelo barato: configuración y tokens](https://broadrouter.com/es/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): Conecte OpenClaw o Hermes Agent a un endpoint compatible con OpenAI, entienda por qué las tareas programadas gastan tokens y reduzca la factura. - [Un endpoint privado empieza con una carga de trabajo clara](https://broadrouter.com/es/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): Qué definir antes de pedir un endpoint privado de LLM: carga de trabajo, tokens por segundo, concurrencia, contexto, región, licencia y presupuesto, y cómo probarlo. - [Inyección de prompts en agentes de código: defensas que sí resisten](https://broadrouter.com/es/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): Por dónde llega la inyección de prompts a un agente de código y las defensas por capas de proveedores e investigadores: sandbox, red limitada, listas y aprobación. - [Tamaño de chunk y solapamiento en RAG: efecto en recuperación y costo](https://broadrouter.com/es/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): Cómo el tamaño de chunk y el solapamiento cambian el recall de RAG, el índice, el costo de embeddings y los tokens por consulta al LLM, con pruebas publicadas. - [Cómo configurar max_tokens en modelos de razonamiento: un método](https://broadrouter.com/es/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Por qué los modelos de razonamiento devuelven respuestas vacías con un max_tokens bajo, y cómo fijarlo con percentiles, la ventana de contexto y reintentos. - [Grabar audio en el navegador y enviarlo a una API de voz a texto](https://broadrouter.com/es/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): Grabe el micrófono con MediaRecorder, convierta WebM o el MP4 de Safari a WAV de 16 kHz y envíelo a una API de transcripción con un servidor que guarda la clave. - [Reducir tokens del prompt: recortes medidos en sistema, tools y datos](https://broadrouter.com/es/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): Ahorro de tokens medido al recortar prompts del sistema, few-shot, esquemas de tools, formatos de datos, historial y diacríticos árabes, con tokenizadores reales. - [Mismo modelo, distinto proveedor: cuánto varían los precios de las API](https://broadrouter.com/es/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): En 119 modelos que venden varios proveedores, el precio de lista más caro es una mediana de 1.76 veces el más barato. Datos de 41 proveedores, noviembre de 2026. - [PDPL de Arabia Saudita y EAU y API de LLM: checklist de transferencias](https://broadrouter.com/es/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Datos personales hacia una API de LLM desde Arabia Saudita o EAU: las reglas de transferencia de la PDPL, qué preguntar y cuándo importa la inferencia local. - [¿Texto, voz o visión? Empiece por la tarea](https://broadrouter.com/es/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): Elija chat de texto, entrada de imágenes o transcripción de audio según la tarea, con los modelos y límites que documenta BroadNet y lo que no ofrece, como TTS. - [Transcribir audio largo de más de 25 MB: dividir, enviar y unir](https://broadrouter.com/es/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): Cómo transcribir grabaciones que superan el límite de una API de voz a texto: reduzca el archivo, divídalo en silencios con ffmpeg, envíe en paralelo y una el texto. - [Traducir textos de apps al árabe con un LLM: placeholders y plurales](https://broadrouter.com/es/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): Traduzca archivos JSON de i18n al árabe con una API de LLM: congele placeholders, genere las seis formas de plural, valide ICU en CI y aísle los nombres insertados. - [Traducir audio en árabe a texto en inglés por API: pipeline probado](https://broadrouter.com/es/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): Pase grabaciones en árabe a texto en inglés con dos llamadas: transcriba con language=ar y traduzca con un LLM, con control de cifras, división, costos y evaluación. - [Disponible, bajo solicitud o investigación: lea bien un catálogo](https://broadrouter.com/es/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): Cómo distinguir qué modelos de un catálogo puede llamar hoy, cuáles requieren una consulta y cuáles solo se probaron, con el catálogo de BroadNet y /v1/models. - [Uno en vLLM: qué midió el estudio de serving de BroadNet](https://broadrouter.com/es/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet integró en vLLM el adaptador de autoborrador Uno y lo midió en una H100: 2.58x el rendimiento de vLLM normal con un stream y 1.74x con 32. Método y límites. - [Uno en Gemma 4 26B: el estudio de BroadNet sobre mezcla de expertos](https://broadrouter.com/es/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet entrenó el primer adaptador de autoborrador Uno para un modelo de mezcla de expertos, Gemma 4 26B A4B. Qué midió, cuándo compensa el autoborrador y el DOI. - [Presupuesto de tokens de un agente: adónde van los tokens del bucle](https://broadrouter.com/es/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): Prompt de sistema, esquemas de herramientas, archivos, historial y razonamiento: cómo gasta tokens el bucle de un agente, cómo medir cada parte y qué recortar. - [Tasa de error de palabras (WER) en el reconocimiento de voz en árabe](https://broadrouter.com/es/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): Qué es la tasa de error de palabras, cómo se calcula y por qué el WER en árabe depende tanto de diacríticos, ortografía y dialectos. Con Python y cifras reales. - [Voz a texto en dialectos árabes: Golfo, egipcio, levantino y magrebí](https://broadrouter.com/es/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Por qué la voz a texto falla con el árabe del Golfo, egipcio, levantino y magrebí, y cómo armar un conjunto de prueba por dialecto para evaluar a un proveedor. - [Traducción árabe–inglés con LLM: prompts, glosarios y evaluación](https://broadrouter.com/es/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): Traduzca del árabe al inglés y viceversa con una API de LLM: prompts que conservan el tono, glosarios de términos fijos, números y nombres, y cómo evaluar calidad. - [Árabe del Golfo, egipcio y levantino: los dialectos y los LLM](https://broadrouter.com/es/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): Cómo manejan los LLM el árabe del Golfo, egipcio y levantino: dónde fallan, cómo pedir un dialecto, cómo probarlo y qué reporta BroadNet de Qwen3.8 27B. - [OCR árabe frente a LLM de visión: leer documentos y facturas en árabe](https://broadrouter.com/es/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): Cuándo usar OCR clásico y cuándo un LLM de visión para facturas, recibos y formularios en árabe: trampas de precisión, salida JSON, costo por página y validación. - [Ingeniería de prompts en árabe: registro, RTL, cifras y texto mixto](https://broadrouter.com/es/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): Prompts para obtener buen árabe de un LLM: elija registro y dialecto, controle cifras y fechas, cuide el formato RTL y gestione la mezcla de árabe e inglés. - [RAG en árabe: normalización, chunking y recuperación que funciona](https://broadrouter.com/es/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): RAG sobre documentos en árabe: normalice alef, ya y ta marbuta, divida en chunks con criterio, elija embeddings multilingües y añada búsqueda por palabras clave. - [API de voz a texto en árabe: cómo elegir una en 2026](https://broadrouter.com/es/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): Cómo elegir una API de transcripción en árabe: tasa de error de palabras, dialectos, límites de archivo, precio por hora de audio, privacidad y un plan de pruebas. - [Chatbot de WhatsApp en árabe: de las notas de voz a las respuestas](https://broadrouter.com/es/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): Arquitectura funcional de un asistente de WhatsApp en árabe: webhook, cola, transcripción de notas de voz, respuesta con LLM, políticas y costos, con Python. - [Arabizi y LLM: cómo manejar el franco-árabe con números como 3 y 7](https://broadrouter.com/es/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): El arabizi escribe el árabe con letras latinas y dígitos como 3 para ع y 7 para ح. Cómo detectarlo, convertirlo a escritura árabe con un LLM y responder bien. - [El mejor LLM para Aider: cómo elegir un modelo y conectarlo](https://broadrouter.com/es/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Cómo elegir el mejor LLM para Aider: formato de edición, contexto, costo por hora, una prueba de una hora en su repo y la configuración de OPENAI_API_BASE. - [Mejor modelo abierto para agentes: herramientas, contexto y costo](https://broadrouter.com/es/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): ¿GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B o Gemma 4 para su agente de programación? Elija por herramientas, contexto, razonamiento y costo por tarea, con cálculos. - [Tokens de entrada en caché: cómo reducir la factura de una API de LLM](https://broadrouter.com/es/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): La entrada en caché cuesta una fracción de la entrada normal. Cómo se factura el prompt caching, cómo leer cached_tokens en usage y cómo estructurar los prompts. - [La API de LLM de visión más barata: costo por imagen y por página](https://broadrouter.com/es/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): Cuánto cuesta un LLM de visión por imagen y por página: cálculo de tokens para Gemma 4, Qwen3.8 27B y DeepSeek V4.1 Flash, y dónde DeepInfra es más barato. - [Cómo elegir un proveedor de API de LLM: lista de 20 puntos](https://broadrouter.com/es/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): Veinte preguntas antes de elegir una API de LLM: precios, límites, compatibilidad con OpenAI, privacidad, modelos, soporte y disponibilidad, con hoja de puntuación. - [Cómo elegir una API de LLM para árabe: qué probar](https://broadrouter.com/es/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): Cómo elegir un LLM para apps en árabe: dialectos, redacción, traducción, documentos e imágenes en árabe, costo y un conjunto de prueba que se arma en un día. - [Cline vs Roo Code vs Aider para equipos con modelos abiertos (2026)](https://broadrouter.com/es/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (ya archivado, continuado como Zoo Code) y Aider comparados en lo que importa con modelos abiertos: tool calls, contexto, MCP, git y aprobaciones. - [Cómo contar tokens antes de enviar una petición a un LLM](https://broadrouter.com/es/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): Estime los tokens del prompt por idioma y modelo antes de llamar a una API LLM: por qué difieren los tokenizadores, cinco idiomas medidos y márgenes para max_tokens. - [10 formas de reducir el costo de LLM de un agente de programación](https://broadrouter.com/es/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Diez formas concretas de reducir lo que Cline, Aider, opencode o su propio agente de programación gasta en tokens, cada una con el cálculo del ahorro y cómo medirlo. - [Alternativas a la API de DeepSeek: 5 formas de llamar a sus modelos](https://broadrouter.com/es/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): Cómo llamar a los modelos de DeepSeek con la API oficial, un agregador, una nube, GPU propias o BroadNet Inference: precio, privacidad y límites de cada opción. - [Precio de la API de DeepSeek V4.1 Flash: por token, en caché y planes](https://broadrouter.com/es/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): Precio de la API de DeepSeek V4.1 Flash por 1M de tokens de entrada, en caché y de salida, planes de tarifa plana y comparación con DeepSeek, DeepInfra y Fireworks. - [Cómo estimar el costo de una API de LLM antes de lanzar una función](https://broadrouter.com/es/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): Un modelo de costos paso a paso para una función de IA: usuarios × acciones × llamadas × tokens, con caché, razonamiento, reintentos, límites y una tabla mensual. - [Cómo evaluar un LLM antes de cambiar de proveedor](https://broadrouter.com/es/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): Cree un pequeño conjunto de evaluación con tráfico real, puntúelo con código y un LLM juez sin sesgos, y compare calidad, costo y latencia antes de migrar. - [Extraer datos de facturas, recibos y formularios a JSON con visión](https://broadrouter.com/es/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): Flujo de producción para facturas, recibos y formularios: imágenes, prompt con esquema JSON, Pydantic, controles aritméticos, reintentos y revisión humana. - [API de LLM con tarifa plana o pago por token: el punto de equilibrio](https://broadrouter.com/es/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): ¿Cuándo sale más barato un plan de LLM con tarifa plana que pagar por token? Cuatro facturas mensuales de DeepSeek V4.1 Flash, el equilibrio y cómo medir el suyo. - [API de LLM gratis: hasta dónde llega realmente un crédito inicial](https://broadrouter.com/es/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): Qué compra de verdad un crédito gratis de API de LLM: tokens, minutos de audio y claves por hora con el crédito de BroadNet, frente a Gemini y OpenRouter. - [RGPD y API de LLM: 15 preguntas para su proveedor](https://broadrouter.com/es/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): Checklist del RGPD para API de LLM: DPA, subencargados, conservación, entrenamiento, transferencias internacionales y brechas de datos, con los artículos aplicables. - [GLM 5.3 vs. DeepSeek V4.1 Flash: cuál usar y cuándo](https://broadrouter.com/es/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 vs. DeepSeek V4.1 Flash en precio, contexto, herramientas, salida estructurada, imágenes y razonamiento, con costos calculados y una regla sencilla de uso. - [Cómo manejar el 429 de una API de LLM: backoff, jitter y limitadores](https://broadrouter.com/es/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 frente a 5xx, backoff exponencial con full jitter, limitadores de concurrencia y colas para API de LLM, con código en Python y Node listo para producción. - [Clave de API de LLM por hora: precio fijo por hora, explicado](https://broadrouter.com/es/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): La clave por hora de BroadNet da una hora de Gemma 4 VL 26B en un stream por un precio fijo. Qué incluye, sus límites, para quién es y cuándo compensa. - [Retención de datos en API de LLM: OpenAI, Anthropic, Gemini, Mistral](https://broadrouter.com/es/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): Qué dicen OpenAI, Anthropic, la API de Gemini, Mistral y BroadNet sobre guardar prompts de la API y entrenar con ellos, citado de sus páginas en noviembre de 2026. - [Claves de API de LLM seguras: almacenamiento, rotación y listas de IP](https://broadrouter.com/es/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): Mantenga las claves de API de LLM fuera del código y los logs, use una clave por carga de trabajo, rótelas, restrínjalas a sus IP y revoque una fuga en segundos. - [Precios de la API de LLM: entrada, caché, salida y razonamiento](https://broadrouter.com/es/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): Cómo se cobra una API de LLM por token: entrada, entrada en caché, salida, tokens de razonamiento, imágenes y audio por minuto, con ejemplos de una tarifa publicada. - [Datos sensibles en una API de LLM: ocultar, minimizar, registrar menos](https://broadrouter.com/es/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): Envíe datos de clientes, de salud o financieros a una API de LLM con seguridad: clasifique campos, use marcadores reversibles, retención cero y logs limpios. - [Control de gasto en APIs de LLM: presupuestos, límites y bucles](https://broadrouter.com/es/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): Evite que su factura de API de LLM se dispare: saldo prepago, clave por app, IPs permitidas, registro de uso, presupuestos en código, topes de bucles y precio fijo. - [Registro de uso de API de LLM: costo por usuario y función, sin PII](https://broadrouter.com/es/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): Registre cada llamada a la API de LLM con tokens, costo, latencia e ID de solicitud, atribuya el gasto a usuarios y funciones y oculte la PII antes de escribir. - [Costo de LLM por usuario: economía unitaria de una función de IA SaaS](https://broadrouter.com/es/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): Calcule el costo de una función de IA por usuario activo al mes, por qué los usuarios intensivos pesan más que el promedio y cómo fijar planes que cubran el modelo. - [Precios de modelos abiertos: cómo cobran 6 proveedores de inferencia](https://broadrouter.com/es/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): Cómo cobran OpenRouter, Together AI, Fireworks, Groq, DeepInfra y BroadNet los modelos abiertos: por token, en caché, por lotes, GPU por hora y planes fijos. - [Latencia de LLM: TTFT, tokens por segundo y latencia total](https://broadrouter.com/es/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): Qué miden el tiempo hasta el primer token, los tokens por segundo y la latencia total, cómo medirlos desde el cliente con código y qué ajuste mejora cada cifra. - [Streaming de LLM explicado: chunks SSE, deltas, usage y cancelación](https://broadrouter.com/es/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Cómo viaja el streaming de chat completions, cómo reconstruir la respuesta desde los deltas, leer usage en el último chunk y cancelar en Python, Node y curl. - [Contexto largo vs RAG: cuándo 1M de tokens reemplaza la base vectorial](https://broadrouter.com/es/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): Cuándo una ventana de contexto de 1M de tokens puede sustituir a una base de datos vectorial y cuándo gana RAG: costo por consulta, latencia, recall y mantenimiento. - [Gestión del contexto en agentes de larga duración con ventana de 1M](https://broadrouter.com/es/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): Sesiones largas de agentes precisas y asequibles con una ventana de 1M de tokens: historial de solo anexado, prompts aptos para caché, cuándo compactar y qué quitar. - [Agentes de larga duración: timeouts, reintentos y checkpoints sólidos](https://broadrouter.com/es/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): Mantenga un agente LLM funcionando durante horas: timeout por llamada, streaming para detectar bloqueos, reintentos seguros, herramientas idempotentes y checkpoints. - [Servidores MCP con cualquier modelo compatible con OpenAI](https://broadrouter.com/es/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP sirve con cualquier LLM: el cliente convierte sus herramientas en function calls. Cómo usar MCP con modelos abiertos en Cline, Continue, opencode, n8n o Python. - [Ventana de contexto de 1M de tokens: cuándo compensa y cuándo no](https://broadrouter.com/es/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): Qué cabe en 1,048,576 tokens, cuánto cuesta en DeepSeek V4.1 Flash, cómo cambian las cuentas el razonamiento y la caché, y cuándo RAG sigue siendo mejor opción. - [Cuánto cuesta un sistema multiagente: por token o sesiones planas](https://broadrouter.com/es/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): Un sistema multiagente gasta varias veces los tokens de un solo agente. Sesiones de tarifa plana vs pago por token, multiplicadores ocultos y cuándo gana cada uno. - [API de voz a texto multilingüe: 14 idiomas, un solo endpoint](https://broadrouter.com/es/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): Transcriba inglés, español, francés, chino, árabe y 9 idiomas más con un solo endpoint compatible con OpenAI, con precio por minuto, y cuándo usar el modelo árabe. - [AI Agent de n8n con un LLM barato: configuración, modelo y costo](https://broadrouter.com/es/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): El AI Agent de n8n sobre un modelo barato compatible con OpenAI: qué modelo elegir para herramientas, qué ajustes ahorran tokens y el costo por 10,000 ejecuciones. - [Modelos abiertos vs. cerrados para empresas: costo, control y riesgo](https://broadrouter.com/es/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): ¿LLM de pesos abiertos o cerrados para su empresa? Compare costo, control, privacidad, calidad y dependencia, y las tres formas de usar modelos abiertos. - [Cambiar a una API compatible con OpenAI: lista de verificación](https://broadrouter.com/es/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): Cambiar base_url es lo fácil. Lo que falla: IDs de modelo, tokens de razonamiento, uso en streaming, errores, límites y claves. Lista de verificación con código. - [OpenRouter o tarifa plana para agentes de IA: ¿qué cuesta menos?](https://broadrouter.com/es/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter cobra por token más una comisión sobre los créditos; un plan plano cobra por sesión. Costos de agentes siempre activos, comisiones, límites y cuál elegir. - [Cómo dejar agentes de código trabajando de noche, con seguridad](https://broadrouter.com/es/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): Configuración práctica para agentes de código desatendidos: sandbox, checkpoints en git, archivos de tarea, topes de pasos, revisión matutina y sesiones de plan. - [Cómo estructurar los prompts para que el prefijo quede en caché](https://broadrouter.com/es/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): Cómo organizar prompts de LLM para que el prefijo en caché no cambie: prompt del sistema, esquemas de tools, documentos, historial solo de adición y una prueba. - [Tokens de razonamiento: por qué paga más salida de la que ve](https://broadrouter.com/es/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): Los modelos razonadores cobran el pensamiento oculto como salida. Cómo medirlo, su costo por modelo y cómo bajarlo con reasoning_effort, max_tokens y enrutamiento. - [Cómo ejecutar agentes de programación en paralelo con tarifa plana](https://broadrouter.com/es/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Subagentes, orquestadores y workers por lotes en DeepSeek V4.1 Flash ilimitado: cuántas sesiones necesita, cómo repartirlas entre claves y limitar el solapamiento. - [Alojar un LLM propio o usar una API: costo real y punto de equilibrio](https://broadrouter.com/es/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): Precios oficiales de alquiler de GPU, la fórmula del costo por 1M de tokens en su GPU, la utilización necesaria para superar a una API y los costos que se olvidan. - [Precios de API de voz a texto comparados: precio por minuto en 2026](https://broadrouter.com/es/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): Precio por minuto y por 1,000 horas de OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq y BroadNet Transcribe, según sus páginas oficiales, y extras. - [Modo JSON o json_schema: salida estructurada que no falla](https://broadrouter.com/es/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): El modo JSON da JSON válido; json_schema da su estructura. Cómo pedir cada uno en modelos compatibles con OpenAI, validar con Pydantic y reparar errores en un bucle. - [Tokens por palabra en árabe vs inglés: por qué el árabe cuesta más](https://broadrouter.com/es/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): Por qué el árabe suele usar más tokens de LLM que el inglés para el mismo contenido, por qué los tokens por palabra engañan y un script para medir su proporción. - [Tool calling con modelos abiertos: esquemas, paralelismo, reintentos](https://broadrouter.com/es/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): Function calling en modelos abiertos compatibles con OpenAI: esquemas, llamadas en paralelo, los cinco fallos típicos y un bucle Python que valida y reintenta. - [Cómo transcribir grabaciones de call center: pipeline listo para árabe](https://broadrouter.com/es/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): Separe llamadas estéreo por canal, corte turnos, transcriba árabe e inglés, oculte datos personales y resuma con un LLM. Pipeline en Python con costos. - [Modelos abiertos en agentes de código: Aider, Cline, opencode, Codex](https://broadrouter.com/es/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Conecte Aider, Cline, Roo Code, opencode, Continue y Codex CLI a cualquier endpoint compatible con OpenAI. Los nombres exactos de cada campo y lo que falla. - [Notas de voz a texto para empresas: de WhatsApp y Telegram a tickets](https://broadrouter.com/es/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): Convierta notas de voz de clientes de WhatsApp Business y bots de Telegram en transcripciones y tickets: descargar, transcribir, extraer campos con un LLM, enrutar. - [Un plan LLM semanal para proyectos cortos: Órbita frente a hora y mes](https://broadrouter.com/es/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): Cuándo una semana ilimitada de DeepSeek V4.1 Flash supera a una sesión por hora, un plan mensual o pagar por token: cálculos para hackathons, sprints y migraciones. - [¿Qué significa realmente «ilimitado» en un plan de API de LLM?](https://broadrouter.com/es/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Medidor de tokens, ventanas de 5 horas, topes semanales, límites de frecuencia, concurrencia: siete preguntas para saber si un plan «ilimitado» lo es para su agente. - [Alternativa a la API de Whisper: cuándo cambiar y cómo probarla](https://broadrouter.com/es/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): Cuándo conviene una alternativa alojada a la API Whisper de OpenAI: error por palabra con su audio, idiomas, precio por hora, límites y privacidad, con código. - [API de LLM con retención cero de datos: lista de 10 preguntas](https://broadrouter.com/es/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): Qué debe significar retención cero de datos en una API de LLM: prompts, salidas, logs, entrenamiento, subencargados, auditorías. 10 preguntas antes de enviar datos. - [Guides in Spanish](https://broadrouter.com/es/guides/) - [Usar el SDK de OpenAI para Python con BroadNet Inference](https://broadrouter.com/es/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): Apunte el SDK oficial de OpenAI para Python a BroadNet Inference cambiando solo base_url: instalación, clave, primera solicitud y respuesta en streaming. - [Usar el SDK de OpenAI para Node.js con BroadNet Inference](https://broadrouter.com/es/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): Apunte el SDK oficial de OpenAI para Node.js y TypeScript a BroadNet Inference con baseURL: instalación, primera solicitud, streaming e IDs de solicitud. - [Pruebe BroadNet Inference en su navegador, sin cuenta](https://broadrouter.com/es/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): Pruebe voz a texto en árabe, lectura de documentos y chat con Gemma 4 en el playground de BroadNet, sin registro ni tarjeta, y pase luego la solicitud a la API. - [Primera solicitud a BroadNet Inference con curl](https://broadrouter.com/es/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): Sus primeras llamadas a BroadNet Inference con curl: un chat completion, una respuesta en streaming y una transcripción, y cómo leer respuesta, uso y errores. - [Streaming de chat completions y conexiones interrumpidas](https://broadrouter.com/es/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): Streaming en BroadNet Inference: Server-Sent Events, el marcador [DONE], el uso de tokens en el último fragmento y qué hacer si el stream termina antes. - [¿Por qué un modelo de razonamiento devuelve una respuesta vacía?](https://broadrouter.com/es/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3 y Qwen3.8 27B piensan antes de responder, y ese razonamiento cuenta para max_tokens. Cómo evitar las respuestas vacías. - [Códigos de error de BroadNet Inference y cuándo reintentar](https://broadrouter.com/es/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): Qué significa cada código HTTP de la API de BroadNet Inference, cuáles reintentar con backoff, cuáles corregir antes y qué enviar al soporte con el ID de solicitud. - [Limite sus claves de API a sus servidores con listas de IP permitidas](https://broadrouter.com/es/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): Proteja sus claves de BroadNet Inference con listas de IP permitidas, separe claves de inferencia y credenciales del portal y revoque una clave filtrada en segundos. - [Cómo funcionan los planes de API de LLM ilimitada para agentes](https://broadrouter.com/es/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): Los planes ilimitados de DeepSeek V4.1 Flash no miden tokens ni tienen límite semanal. Así funcionan las sesiones, los agentes en paralelo, las claves y el cobro. - [Aider con DeepSeek V4.1 Flash ilimitado](https://broadrouter.com/es/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Conecte el agente de programación Aider a BroadNet Inference con su opción compatible con OpenAI y úselo con un plan ilimitado de DeepSeek V4.1 Flash a precio fijo. - [Cómo usar LangChain con BroadNet Inference](https://broadrouter.com/es/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): Conecte ChatOpenAI de LangChain a BroadNet Inference con base_url y api_key, y sepa qué conserva y qué descarta LangChain de los modelos de razonamiento. - [Cline con BroadNet Inference (proveedor OpenAI Compatible)](https://broadrouter.com/es/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Configure la extensión Cline de VS Code con BroadNet Inference mediante su proveedor OpenAI Compatible: campos exactos, IDs de modelo, contexto y soluciones comunes. - [Roo Code con BroadNet Inference (proveedor OpenAI Compatible)](https://broadrouter.com/es/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code cerró en mayo de 2026. Configuración para instalaciones existentes y forks, y a dónde migrar (Cline o el fork Zoo Code) con BroadNet Inference. - [Continue con BroadNet Inference (config.yaml y apiBase)](https://broadrouter.com/es/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): Agregue modelos de BroadNet Inference a la extensión Continue con provider openai y apiBase en config.yaml, con la capacidad tool_use y la longitud de contexto. - [Agregue BroadNet Inference a opencode como proveedor personalizado](https://broadrouter.com/es/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): Agregue BroadNet Inference al agente de programación opencode como proveedor personalizado compatible con OpenAI en opencode.jsonc, con alias, capacidades y límites. - [Cómo conectar Open WebUI a BroadNet Inference](https://broadrouter.com/es/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Agregue BroadNet Inference a Open WebUI como conexión compatible con OpenAI, permita sus IDs de modelo y resuelva fallos de verificación, respuestas vacías y 429. - [Agregar BroadNet Inference a LibreChat como endpoint personalizado](https://broadrouter.com/es/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): Agregue BroadNet Inference a LibreChat como endpoint personalizado compatible con OpenAI en librechat.yaml, con la clave en .env, los modelos y el modelo de títulos. - [BroadNet Inference en n8n (credencial de OpenAI y Base URL)](https://broadrouter.com/es/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): Apunte la credencial de OpenAI de n8n a BroadNet Inference con el campo Base URL y use el nodo OpenAI Chat Model por ID con la opción Responses API desactivada. - [Cómo usar el Vercel AI SDK con BroadNet Inference](https://broadrouter.com/es/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): Llame a BroadNet Inference desde el Vercel AI SDK con createOpenAICompatible de @ai-sdk/openai-compatible: generateText, streamText, herramientas y reasoning effort. - [API de voz a texto en árabe, paso a paso](https://broadrouter.com/es/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): Transcriba audio en árabe con BroadNet Transcribe Arabic en un endpoint compatible con OpenAI: formatos, el campo language, ejemplos en Python y curl, y precio. - [Cómo enrutar LiteLLM a BroadNet Inference (SDK y proxy)](https://broadrouter.com/es/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): Llame a BroadNet Inference desde LiteLLM con el prefijo openai/ y api_base, en el SDK de Python o en el config.yaml del proxy, con límites de frecuencia acordes. - [OpenAI Agents SDK con BroadNet Inference](https://broadrouter.com/es/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Use el OpenAI Agents SDK para Python con BroadNet Inference mediante AsyncOpenAI y OpenAIChatCompletionsModel, y desactive el tracing que exige una clave de OpenAI. - [Transcribir notas de voz de WhatsApp a texto con una API](https://broadrouter.com/es/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): Convierta notas de voz de WhatsApp (.ogg) o grabaciones del teléfono (.m4a) a WAV con ffmpeg y transcríbalas en árabe u otros 13 idiomas con una sola llamada. - [Transcribir notas de voz de WhatsApp en n8n](https://broadrouter.com/es/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): Un workflow de n8n que recibe una nota de voz de WhatsApp, la descarga de Meta, la convierte con ffmpeg y la transcribe en árabe o en otros 13 idiomas. - [Lea facturas y recibos en árabe con un modelo de visión](https://broadrouter.com/es/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Envíe la foto de una factura o recibo a un modelo de visión de BroadNet Inference y obtenga total, fecha y proveedor en JSON, con preguntas en árabe o inglés. - [Usar Zed con BroadNet Inference como proveedor compatible con OpenAI](https://broadrouter.com/es/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Agregue BroadNet Inference a Zed como proveedor compatible con OpenAI en settings.json: IDs de modelo, límites de tokens, capacidades y variable de la clave. - [Kilo Code con BroadNet Inference (proveedor OpenAI Compatible)](https://broadrouter.com/es/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Conecte Kilo Code a BroadNet Inference como proveedor OpenAI Compatible, en el diálogo de VS Code o en kilo.json, con IDs de modelo, límites y herramientas. - [Agregar BroadNet Inference a goose como proveedor personalizado](https://broadrouter.com/es/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): Conecte el agente de IA goose a BroadNet Inference como proveedor personalizado OpenAI Compatible, desde goose Desktop, goose configure o un archivo JSON. - [Pydantic AI con BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/es/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Ejecute agentes de Pydantic AI en BroadNet Inference con OpenAIChatModel y OpenAICompatibleProvider: herramientas, salida estructurada y ajustes del perfil. - [CrewAI con BroadNet Inference como endpoint OpenAI personalizado](https://broadrouter.com/es/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): Conecte los agentes de CrewAI a BroadNet Inference con LLM(custom_openai=True, base_url=...) y elija modelos para herramientas, contexto largo y tarifa plana. - [Usar LlamaIndex con BroadNet Inference mediante OpenAILike](https://broadrouter.com/es/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): Conecte LlamaIndex a BroadNet Inference con la clase OpenAILike y configure bien api_base, context_window, is_chat_model e is_function_calling_model. - [BroadNet Inference en Dify con el plugin OpenAI-API-compatible](https://broadrouter.com/es/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): Conecte Dify a BroadNet Inference con el plugin de proveedor OpenAI-API-compatible: valores exactos de los campos, tamaños de contexto y llamadas a funciones. - [Jan con BroadNet Inference como proveedor compatible con OpenAI](https://broadrouter.com/es/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Agregue BroadNet Inference a la app Jan como proveedor personalizado compatible con OpenAI, añada los IDs de modelo a mano y active herramientas y visión. - [Conectar SillyTavern a BroadNet Inference (Custom OpenAI-compatible)](https://broadrouter.com/es/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): Conecte SillyTavern a BroadNet Inference con la fuente Custom (OpenAI-compatible) de Chat Completion, los IDs de modelo correctos, posprocesado y límites de tokens. - [Spring AI con BroadNet Inference (base-url de OpenAI)](https://broadrouter.com/es/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Apunte el starter de chat OpenAI de Spring AI a BroadNet Inference con spring.ai.openai.base-url, un ID de BroadNet y max-tokens, y llámelo desde ChatClient. - [Llame a BroadNet Inference desde Go con el SDK openai-go](https://broadrouter.com/es/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): Use el SDK oficial openai-go v3 con BroadNet Inference mediante option.WithBaseURL: chat completions, streaming con datos de uso y manejo de errores en Go. - [Agentes de LangGraph con BroadNet Inference](https://broadrouter.com/es/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): Ejecute un StateGraph o create_agent de LangGraph en BroadNet Inference con ChatOpenAI, base_url y un ID de BroadNet, con herramientas, streaming y memoria. - [Pipelines y agentes de Haystack con BroadNet Inference](https://broadrouter.com/es/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Apunte OpenAIChatGenerator de Haystack a BroadNet Inference con api_base_url, un ID de BroadNet y un Secret, para pipelines, herramientas, streaming y agentes. - [Semantic Kernel con BroadNet Inference](https://broadrouter.com/es/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Conecte Microsoft Semantic Kernel a BroadNet Inference en Python con un cliente AsyncOpenAI, o en C# con un endpoint personalizado, para chat, plugins y agentes. - [Usar agentes de Mastra con BroadNet Inference](https://broadrouter.com/es/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): Ejecute agentes de Mastra en BroadNet Inference con un objeto model que define id, url y apiKey, incluido el prefijo broadnet/ duplicado que necesita el ID. - [Salida estructurada con Instructor en BroadNet Inference](https://broadrouter.com/es/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Extraiga objetos Pydantic de BroadNet Inference con Instructor, mediante from_openai o from_provider con un base_url propio, y elija el Mode adecuado. - [DSPy con BroadNet Inference (defina model_type="chat")](https://broadrouter.com/es/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): Configure dspy.LM para BroadNet Inference con el prefijo openai/, api_base y model_type="chat", porque DSPy 3.5 envía los modelos de OpenAI a la Responses API. - [Evalúe los modelos de BroadNet con promptfoo](https://broadrouter.com/es/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): Ejecute evals de promptfoo y compare modelos lado a lado en BroadNet Inference con el proveedor openai:chat, apiBaseUrl y apiKeyEnvar, además de graders. - [smolagents en BroadNet Inference con OpenAIModel](https://broadrouter.com/es/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Conecte smolagents de Hugging Face a BroadNet Inference con OpenAIModel y api_base, para CodeAgent con cualquier modelo o ToolCallingAgent con GLM 5.3 y Qwen3.8 27B. - [Conectar Flowise a BroadNet Inference (OpenAI Custom Model)](https://broadrouter.com/es/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Use BroadNet Inference en chatflows de Flowise con el nodo OpenAI Custom Model, un Base Path, una credencial OpenAI API y un modelo de BroadNet escrito a mano. - [Conectar Langflow a BroadNet Inference (componente OpenAI)](https://broadrouter.com/es/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Use BroadNet Inference en Langflow con el campo OpenAI API Base del componente OpenAI y un nombre de modelo escrito a mano, y por qué OpenAI Compatible no lo lista. - [Glossary entries in Spanish](https://broadrouter.com/es/glossary/) - [Agente de IA (AI agent)](https://broadrouter.com/es/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): Un agente de IA es un sistema en el que un modelo de lenguaje dirige sus propios pasos hacia un objetivo: decide qué herramientas llamar, lee los resultados y repite hasta dar una respuesta final o alcanzar un límite. Un script fijo que llama a un modelo en puntos establecidos suele llamarse flujo de trabajo (workflow). - [Clave de API (API key)](https://broadrouter.com/es/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): Una clave de API es una cadena secreta que identifica y autoriza a un cliente que llama a una API. Las APIs de LLM suelen esperarla en un encabezado Authorization: Bearer, y cualquiera que tenga la clave puede hacer solicitudes que se facturan a la cuenta propietaria. - [Diacríticos árabes (tashkeel)](https://broadrouter.com/es/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): Los diacríticos árabes, o tashkeel, son pequeñas marcas que se escriben encima o debajo de las letras para indicar vocales breves, consonantes dobles y la ausencia de vocal. Son opcionales: la mayor parte del texto árabe, incluidas las noticias, las omite, así que lectores y software deducen las vocales por el contexto. - [Identificación de dialectos árabes](https://broadrouter.com/es/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): La identificación de dialectos árabes es la tarea de determinar en qué variedad del árabe está un texto o una grabación, como el egipcio, el del Golfo, el levantino, el magrebí o el árabe estándar moderno, a nivel de región, país o ciudad. Se usa para enrutar mensajes, elegir prompts y dividir conjuntos de evaluación. - [Normalización de texto árabe](https://broadrouter.com/es/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): La normalización de texto árabe es el paso que lleva las variantes ortográficas de una misma palabra a una sola forma antes de compararlas: quitar diacríticos y tatweel, unificar las formas de alef y, a menudo, convertir ى en ي y ة en ه. Se aplica a claves de búsqueda y evaluación, no al texto que se muestra. - [Arabizi (franco-árabe)](https://broadrouter.com/es/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): El arabizi es árabe escrito de manera informal con letras latinas, usando dígitos para los sonidos árabes sin equivalente latino, como 3 para ع. Es habitual en chats y redes sociales, es sobre todo dialectal, no tiene ortografía estándar y a menudo se mezcla con inglés o francés. - [AWQ (cuantización de pesos consciente de activaciones)](https://broadrouter.com/es/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (activation-aware weight quantization) es un método posterior al entrenamiento que comprime los pesos de un LLM a baja precisión, normalmente enteros de 4 bits. Usa estadísticas de activación de un pequeño conjunto de calibración para encontrar los canales de pesos más importantes y los escala para que pierdan menos precisión. No necesita retropropagación. - [URL base (base URL)](https://broadrouter.com/es/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): La URL base es la dirección raíz fija de una API, como https://api.example.com/v1, que un cliente antepone a cada ruta de endpoint, como /chat/completions. En los SDKs compatibles con OpenAI, cambiar la URL base, junto con la clave de API y el ID del modelo, dirige el mismo código a otro proveedor. - [Batch API (API de procesamiento por lotes)](https://broadrouter.com/es/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Una Batch API es un endpoint asíncrono de LLM que acepta muchas solicitudes a la vez, normalmente en un archivo, las procesa dentro de un plazo fijado en lugar de inmediatamente y devuelve los resultados juntos. Proveedores como OpenAI y Anthropic cobran las solicitudes por lotes con descuento respecto de sus APIs estándar. - [BM25 (ranking por palabras clave)](https://broadrouter.com/es/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) es una función de ranking para búsqueda por palabras clave que puntúa un documento según la frecuencia con que aparecen en él los términos de la consulta, lo raros que son esos términos en la colección y la longitud del documento. Es la similitud predeterminada en Lucene y Elasticsearch. - [Byte pair encoding (BPE)](https://broadrouter.com/es/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Byte pair encoding (BPE) es un algoritmo de tokenización en subpalabras que parte de caracteres o bytes individuales y fusiona repetidamente el par adyacente más frecuente en una nueva entrada del vocabulario, hasta alcanzar un tamaño de vocabulario objetivo. La mayoría de los tokenizadores de LLM actuales usan BPE o una variante a nivel de byte. - [Tokens de entrada en caché (cached input tokens)](https://broadrouter.com/es/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): Los tokens de entrada en caché son tokens del prompt que un proveedor de LLM procesó recientemente y puede reutilizar desde un caché en lugar de calcularlos de nuevo; normalmente son un prefijo repetido, como un prompt de sistema o un documento. Los proveedores los cobran a una tarifa menor que los tokens de entrada nuevos. - [Cadena de pensamiento (chain-of-thought, CoT)](https://broadrouter.com/es/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): La cadena de pensamiento (chain-of-thought, CoT) es una técnica de prompting en la que a un modelo de lenguaje se le muestra o se le pide que escriba pasos de razonamiento intermedios antes de su respuesta final. Mejora la precisión en problemas de varios pasos, como aritmética y lógica, y los modelos de razonamiento ya la generan automáticamente. - [Tasa de error por carácter (CER)](https://broadrouter.com/es/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): La tasa de error por carácter (CER) mide la precisión de una transcripción u OCR a nivel de carácter: las sustituciones, eliminaciones e inserciones necesarias para convertir la salida en el texto de referencia, divididas entre el número de caracteres de la referencia. Cuanto menor, mejor, y 0 es una coincidencia perfecta. - [API de Chat Completions](https://broadrouter.com/es/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): La API de Chat Completions es el endpoint HTTP de OpenAI, POST /v1/chat/completions, que recibe un nombre de modelo y una lista de mensajes etiquetados con su rol y devuelve el siguiente mensaje del modelo. Su forma de solicitud y respuesta se convirtió en el formato común que reutilizan los proveedores y SDKs compatibles con OpenAI. - [Plantilla de chat (chat template)](https://broadrouter.com/es/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): Una plantilla de chat es la regla de formato, normalmente escrita en Jinja e incluida con el tokenizador de un modelo, que convierte una lista de mensajes de chat con roles en la única secuencia de tokens que el modelo espera, incluidos sus tokens de control especiales. Usar la plantilla equivocada puede empeorar mucho las respuestas del modelo. - [Chunking (fragmentación de documentos en RAG)](https://broadrouter.com/es/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): El chunking es el paso de un pipeline de recuperación que divide los documentos en piezas más pequeñas, llamadas chunks o fragmentos, antes de convertirlas en embeddings e indexarlas. El tamaño del chunk fija cuánto texto contiene cada pieza, y el solapamiento repite parte del texto entre fragmentos vecinos para que no se pierdan las ideas cortadas en un límite. - [Code-switching (alternancia de códigos)](https://broadrouter.com/es/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): El code-switching, o alternancia de códigos, es el uso de dos o más idiomas, o de variedades de un mismo idioma, dentro de la misma conversación o frase, de forma oral o escrita. Por ejemplo, hablantes de árabe que mezclan palabras inglesas en su dialecto o que pasan del árabe estándar moderno a su dialecto. - [Agente de programación con IA (AI coding agent)](https://broadrouter.com/es/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): Un agente de programación con IA es una herramienta de software en la que un modelo de lenguaje trabaja sobre una base de código en un ciclo: lee archivos, los edita, ejecuta comandos como pruebas, lee la salida y decide el siguiente paso hasta que termina la tarea o alcanza un límite. - [Concurrencia (concurrency)](https://broadrouter.com/es/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): La concurrencia es el número de solicitudes que un cliente tiene en curso al mismo tiempo. Las APIs de LLM suelen limitarla por clave o por plan, aparte de las solicitudes por minuto, porque cada solicitud abierta ocupa capacidad de GPU hasta que su respuesta termina. - [Compactación de contexto (context compaction)](https://broadrouter.com/es/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): La compactación de contexto es la práctica de resumir una conversación que se acerca al límite de la ventana de contexto y continuar la sesión desde ese resumen, de modo que un agente conserve las decisiones y el estado clave mientras descarta el historial en bruto, como la salida antigua de herramientas. - [Ingeniería de contexto (context engineering)](https://broadrouter.com/es/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): La ingeniería de contexto es el trabajo de elegir y mantener el conjunto de tokens que un modelo de lenguaje ve en cada solicitud (prompt de sistema, definiciones de herramientas, historial, documentos recuperados y notas) para que el modelo reciba el conjunto más pequeño de información de alta señal que necesita para hacer bien la tarea. - [Context rot (deterioro del contexto)](https://broadrouter.com/es/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): El context rot (deterioro del contexto) es la pérdida gradual de precisión y fiabilidad que muestran los modelos de lenguaje cuando su entrada se alarga, aunque la tarea sea la misma y la entrada quepa holgadamente en la ventana de contexto. El término se popularizó con un informe técnico de Chroma de 2025 que evaluó 18 modelos. - [Ventana de contexto (context window)](https://broadrouter.com/es/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): La ventana de contexto es la cantidad máxima de tokens con la que un modelo de lenguaje puede trabajar en una sola solicitud. Incluye el prompt, el historial de la conversación, las definiciones de herramientas, las imágenes y la salida generada, incluidos los tokens de razonamiento. Lo que no cabe debe recortarse, resumirse o dividirse en varias solicitudes. - [Residencia de datos (data residency)](https://broadrouter.com/es/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): La residencia de datos es un requisito o compromiso de que los datos se almacenen y procesen en un país o región concretos. En una API de LLM abarca dónde se procesan los prompts durante la inferencia y dónde queda el contenido almacenado, y a menudo excluye los datos de cuenta y facturación. - [Árabe dialectal (dialectal Arabic)](https://broadrouter.com/es/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): El árabe dialectal es el árabe hablado a diario en una región o ciudad, como el árabe egipcio, del Golfo, levantino o magrebí. Los dialectos difieren del árabe estándar moderno en sonidos, palabras y gramática, no tienen ortografía estándar y se escriben cada vez más en internet, lo que los hace más difíciles para los modelos de lenguaje. - [Diarización de hablantes (speaker diarization)](https://broadrouter.com/es/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): La diarización de hablantes es la tarea de dividir una grabación de audio en segmentos por hablante para responder "quién habló y cuándo". Etiqueta cada tramo de habla con un ID de hablante anónimo, como SPEAKER_00, y suele combinarse con una transcripción para mostrar qué persona dijo cada línea. - [Destilación de conocimiento (distillation)](https://broadrouter.com/es/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): La destilación, o destilación de conocimiento, es un método de entrenamiento en el que un modelo alumno más pequeño aprende a reproducir las salidas de un modelo maestro más grande, como sus distribuciones de probabilidad o sus respuestas generadas, de modo que el alumno conserva buena parte de la calidad del maestro con un costo de ejecución menor. - [Acuerdo de tratamiento de datos (DPA)](https://broadrouter.com/es/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): Un acuerdo de tratamiento de datos (DPA) es el contrato entre el responsable y el encargado del tratamiento que regula cómo el encargado trata los datos personales. El artículo 28.3 del RGPD lo exige y fija su contenido: instrucciones documentadas, confidencialidad, seguridad, reglas sobre subencargados, supresión o devolución de los datos y auditorías. - [Embeddings (vectores de texto)](https://broadrouter.com/es/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): Los embeddings son listas de números de punto flotante (vectores) que un modelo produce para representar un texto, una imagen u otra entrada. Las entradas de significado parecido obtienen vectores cercanos, así que la distancia entre vectores sirve para ordenar resultados de búsqueda, agrupar documentos o encontrar elementos relacionados. - [Evaluación de LLM (eval)](https://broadrouter.com/es/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Una eval (abreviatura de evaluation, evaluación) es una prueba repetible de un modelo de lenguaje o de un prompt: un conjunto fijo de entradas, una descripción de cómo es una buena salida y un calificador (código, una persona u otro modelo) que puntúa cada salida para poder comparar versiones. - [Backoff exponencial (exponential backoff)](https://broadrouter.com/es/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): El backoff exponencial es una estrategia de reintento en la que el cliente espera más tiempo tras cada intento fallido, normalmente duplicando la espera hasta un tope, antes de volver a intentarlo. Añadir jitter aleatorio a cada espera dispersa los reintentos de muchos clientes para que no lleguen al servidor en el mismo instante. - [Few-shot prompting (prompting con pocos ejemplos)](https://broadrouter.com/es/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): El few-shot prompting es una técnica en la que el prompt incluye un pequeño número de ejemplos resueltos de entrada y salida antes de la entrada real, para que el modelo de lenguaje deduzca de ellos la tarea y el formato. Los pesos del modelo no cambian; zero-shot no usa ejemplos y one-shot usa uno. - [Fine-tuning (ajuste fino)](https://broadrouter.com/es/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): El fine-tuning (ajuste fino) es el proceso de seguir entrenando un modelo de lenguaje preentrenado con un conjunto de datos más pequeño y específico de una tarea, como conversaciones de ejemplo, para que sus pesos cambien y realice esa tarea o estilo con más fiabilidad sin necesitar instrucciones o ejemplos largos en cada prompt. - [finish_reason (motivo de finalización)](https://broadrouter.com/es/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason es un campo de cada opción en la respuesta de un chat completion que indica por qué el modelo dejó de generar. En el formato de OpenAI los valores son stop, length, tool_calls, content_filter y el obsoleto function_call. La API Messages de Anthropic expresa la misma idea con stop_reason. - [Plan de LLM con tarifa plana (flat-rate LLM plan)](https://broadrouter.com/es/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): Un plan de LLM con tarifa plana cobra un precio fijo por un periodo, como un mes, por acceder a la API de un modelo de lenguaje, en lugar de facturar por token. Estos planes suelen limitar otra cosa, como la concurrencia, la velocidad, los modelos o las ventanas de uso, así que los topes importan tanto como el precio. - [FP8](https://broadrouter.com/es/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 es un formato de coma flotante de 8 bits para deep learning, con dos codificaciones: E4M3, con 4 bits de exponente y 3 de mantisa, y E5M2, con 5 bits de exponente y 2 de mantisa. Reduce a la mitad la memoria frente a los formatos de 16 bits y se ejecuta de forma nativa en GPU recientes. - [Penalización de frecuencia y de presencia](https://broadrouter.com/es/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): Las penalizaciones de frecuencia y de presencia son ajustes de muestreo que bajan la puntuación de los tokens que ya están en la salida. La de presencia se aplica en cuanto un token ha aparecido; la de frecuencia crece con cada repetición. En la API de OpenAI ambas van de -2.0 a 2.0, con 0 por defecto. - [RGPD (GDPR)](https://broadrouter.com/es/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): El RGPD (Reglamento General de Protección de Datos, Reglamento (UE) 2016/679) es la ley de la UE sobre el tratamiento de datos personales. Se aplica desde el 25 de mayo de 2018 a organizaciones establecidas en la UE y a algunas de fuera. En las APIs de LLM regula contratos con encargados, transferencias internacionales y minimización de datos. - [GGUF](https://broadrouter.com/es/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF es un formato de archivo binario para almacenar modelos de lenguaje destinados a inferencia con runtimes basados en GGML, como llama.cpp. Un archivo GGUF contiene los pesos, a menudo cuantizados, y los metadatos necesarios para cargar el modelo, incluidos el tokenizador y la arquitectura, y puede mapearse en memoria para cargar rápido. - [Hora de GPU (GPU-hour)](https://broadrouter.com/es/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): Una hora de GPU es un procesador gráfico usado durante una hora. Los proveedores de nube facturan las instancias de GPU alquiladas por hora o por segundo, y los investigadores usan las horas de GPU para expresar cuánto cómputo requirió un trabajo. Se paga el tiempo de la máquina, esté ocupada o no. - [Grounding (anclaje en fuentes)](https://broadrouter.com/es/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): El grounding (anclaje en fuentes) es la práctica de conectar la salida de un modelo de lenguaje con fuentes de información concretas y verificables, como documentos aportados, resultados de búsqueda o registros de una base de datos, e indicarle que responda a partir de ellas. Las respuestas ancladas son rastreables hasta su fuente y menos propensas a datos inventados. - [Guardrails (barreras de protección)](https://broadrouter.com/es/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): Los guardrails son controles que una aplicación ejecuta alrededor de una llamada a un modelo de lenguaje para mantenerla dentro de límites definidos. Los de entrada filtran lo que entra al modelo, los de salida revisan lo que sale y los de herramientas validan las acciones antes de ejecutarlas. Un control fallido bloquea, reescribe o marca la solicitud. - [Alucinación (hallucination)](https://broadrouter.com/es/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): Una alucinación es una salida de un modelo de lenguaje que suena plausible pero es falsa, inventada o no está respaldada por la entrada que recibió, como una cita inventada, una cifra errónea o una API que no existe. Surge porque el modelo predice texto probable en lugar de verificar hechos. - [Hamza](https://broadrouter.com/es/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): La hamza (ء) es la letra árabe que representa la oclusiva glotal. Se escribe sola o sobre un soporte: alef (أ إ), waw (ؤ) o yeh (ئ), y آ combina alef con madda. Como el texto escrito es irregular al respecto, los buscadores y la evaluación de voz suelen plegar las formas de alef en ا. - [Búsqueda híbrida (hybrid search)](https://broadrouter.com/es/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): La búsqueda híbrida es un método de recuperación que ejecuta sobre la misma consulta una búsqueda por palabras clave, como BM25, y una búsqueda vectorial sobre embeddings, y luego fusiona las dos listas ordenadas en una sola, a menudo con reciprocal rank fusion (RRF) o con una suma ponderada de puntuaciones normalizadas. - [Clave de idempotencia (idempotency key)](https://broadrouter.com/es/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Una clave de idempotencia es un valor único, normalmente un ID aleatorio como un UUID, que el cliente adjunta a una solicitud que cambia el estado, como un pago o una compra. Si la misma solicitud se envía de nuevo con la misma clave, el servidor devuelve el resultado original en lugar de ejecutar la acción dos veces. - [Tokens de imagen (image tokens)](https://broadrouter.com/es/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): Los tokens de imagen son los tokens de entrada que un modelo de lenguaje con visión usa para representar una imagen en el prompt. Los proveedores los calculan a partir del tamaño de la imagen, normalmente dividiéndola en parches de tamaño fijo tras redimensionarla, y los cobran como tokens de entrada de texto, así que las imágenes grandes cuestan más. - [Proveedor de inferencia (inference provider)](https://broadrouter.com/es/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): Un proveedor de inferencia es una empresa que ejecuta modelos de aprendizaje automático, a menudo LLMs de pesos abiertos, en GPUs propias, arrendadas o alquiladas, y vende el acceso mediante una API, normalmente con precio por token, por solicitud o por unidad de tiempo. Se encarga del servicio, el escalado y el hardware. - [Tokens de entrada y de salida (input and output tokens)](https://broadrouter.com/es/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): Los tokens de entrada son los que un modelo de lenguaje lee en una solicitud, incluidos el prompt de sistema, el historial de la conversación, las definiciones de herramientas y las imágenes. Los tokens de salida son los que genera en la respuesta, incluido el razonamiento. Las APIs de LLM suelen cobrarlos por separado, y la salida suele costar más. - [INT4](https://broadrouter.com/es/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 es un formato de enteros de 4 bits que puede representar 16 niveles. En la inferencia de LLM suele referirse a la cuantización solo de pesos en 4 bits (W4A16): los pesos se guardan como INT4 con factores de escala y se expanden a 16 bits al calcular, con cerca de un cuarto de la memoria de 16 bits. - [Lista de IP permitidas (IP allow list)](https://broadrouter.com/es/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): Una lista de IP permitidas es un control de seguridad que solo acepta solicitudes de direcciones IP o rangos CIDR incluidos en ella y rechaza todas las demás. En una API, limita el daño de una clave filtrada, porque la clave no funciona desde redes que no estén en la lista. - [Jailbreak (LLM)](https://broadrouter.com/es/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): Un jailbreak es una entrada diseñada para que un modelo de lenguaje ignore su entrenamiento de seguridad o las reglas de la aplicación que lo despliega, de modo que produzca contenido o realice acciones que normalmente rechazaría. OWASP considera el jailbreaking una forma de inyección de prompts (prompt injection). - [JSON mode (modo JSON)](https://broadrouter.com/es/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): El modo JSON (JSON mode) es un ajuste de las APIs de LLM, response_format {"type": "json_object"} en las APIs al estilo de OpenAI, que hace que la respuesta del modelo sea JSON válido y procesable. A diferencia de la salida estructurada con un JSON Schema, no garantiza qué claves, tipos o valores contiene el objeto. - [JSON Schema (esquema JSON)](https://broadrouter.com/es/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema es un vocabulario, escrito en JSON, para describir y validar la estructura de datos JSON: tipos, campos obligatorios, valores permitidos y anidamiento. La versión actual es la 2020-12. Las APIs de LLM lo usan para definir los parámetros de las herramientas y la forma exacta que debe seguir una salida estructurada. - [Caché KV (KV cache)](https://broadrouter.com/es/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): La caché KV es la memoria donde un modelo de lenguaje transformer guarda las claves y valores de atención calculados para cada token ya procesado, de modo que cada token nuevo los reutiliza en lugar de recalcular toda la secuencia. Crece con la longitud del contexto y a menudo limita cuántas solicitudes puede atender una GPU. - [Latencia (latency)](https://broadrouter.com/es/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): La latencia, en una API de modelos de lenguaje, es el tiempo que tarda una solicitud desde que se envía hasta que se recibe el último token de salida. Es la suma del tiempo hasta el primer token y el tiempo de generación, y suele reportarse en percentiles como p50 y p95, no como promedio. - [LLM como juez (LLM-as-a-judge)](https://broadrouter.com/es/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM como juez (LLM-as-a-judge) es un método de evaluación en el que un modelo de lenguaje califica las salidas de otros modelos, ya sea puntuando una respuesta según una rúbrica o eligiendo la mejor de dos. Se aproxima al juicio humano a un costo mucho menor, pero tiene sesgos conocidos, como favorecer la primera respuesta o la más larga. - [Benchmark de LLM (LLM benchmark)](https://broadrouter.com/es/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): Un benchmark de LLM es un conjunto fijo, normalmente público, de tareas con un método de puntuación, que se usa para medir y comparar modelos de lenguaje en una habilidad, como preguntas de conocimiento, programación o reconocimiento de voz. Su puntuación muestra la capacidad relativa en esa prueba, no cómo rendirá un modelo con su propia carga de trabajo. - [Router de LLM (LLM router)](https://broadrouter.com/es/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): Un router de LLM es un software o servicio que se sitúa entre una aplicación y una o varias APIs de modelos de lenguaje y decide, para cada solicitud, qué modelo o proveedor la atiende, según reglas como el costo, la latencia, la disponibilidad o el tipo de tarea. - [Logprobs (log-probabilidades de tokens)](https://broadrouter.com/es/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Los logprobs son las probabilidades logarítmicas que un modelo de lenguaje asignó a los tokens que generó. Una API que los devuelve muestra, para cada token de salida, qué tan probable lo consideró el modelo y, de forma opcional, los tokens alternativos más probables en esa posición. - [LoRA (adaptación de bajo rango)](https://broadrouter.com/es/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation, adaptación de bajo rango) es un método de fine-tuning eficiente en parámetros que congela los pesos de un modelo preentrenado y entrena un par de pequeñas matrices de bajo rango para ciertas capas. La actualización aprendida puede guardarse como un pequeño adaptador o fusionarse con los pesos, sin añadir latencia de inferencia. - [Lost in the middle (perdido en el medio)](https://broadrouter.com/es/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle es la tendencia observada de los modelos de lenguaje a aprovechar mejor la información relevante cuando aparece cerca del principio o del final de una entrada larga, y peor cuando está en el medio. Recibió ese nombre en un estudio de Liu et al. de 2023. - [max_tokens (límite de tokens de salida)](https://broadrouter.com/es/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens es un parámetro de las APIs de LLM que fija el máximo de tokens que un modelo puede generar en una respuesta. Cuando la generación llega al límite, se detiene y la respuesta queda marcada como cortada, por ejemplo con finish_reason en "length". - [MCP (Model Context Protocol)](https://broadrouter.com/es/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol, protocolo de contexto de modelo) es un protocolo abierto, basado en JSON-RPC 2.0, que estandariza cómo las aplicaciones de LLM se conectan con herramientas y datos externos. Un servidor MCP expone herramientas, recursos y prompts; un cliente MCP dentro de la aplicación host los descubre y pasa al servidor las llamadas a herramientas del modelo. - [Mixture of experts (MoE, mezcla de expertos)](https://broadrouter.com/es/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Mixture of experts (MoE, mezcla de expertos) es un diseño de red neuronal en el que una capa contiene muchas subredes expertas y un router aprendido envía cada token solo a unas pocas. Un modelo de lenguaje MoE tiene un número total de parámetros grande, pero usa un pequeño subconjunto activo por token. - [Deprecación de modelos (model deprecation)](https://broadrouter.com/es/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): La deprecación de un modelo es el anuncio de un proveedor de API de que dejará de ofrecer un modelo, normalmente con una fecha de retirada o apagado y un reemplazo recomendado. Después de esa fecha, las solicitudes al modelo fallan, así que las aplicaciones deben migrar antes a otro ID de modelo. - [Fallback de modelos (model fallback)](https://broadrouter.com/es/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): El fallback de modelos es un patrón de fiabilidad: una solicitud fallida a un LLM se reintenta con otro modelo o proveedor de una lista ordenada. Puede dispararlo un error, un límite de frecuencia, un error de longitud de contexto o un rechazo por política de contenido. Se ejecuta en su cliente o en un gateway delante de las APIs. - [Licencia de modelo (model license)](https://broadrouter.com/es/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): La licencia de un modelo son los términos legales bajo los que se pueden usar, modificar y compartir sus pesos, su código y sus salidas. Los modelos de pesos abiertos usan licencias de software estándar como Apache 2.0 o MIT, o licencias propias que pueden añadir condiciones, como límites a grandes servicios comerciales. - [Árabe estándar moderno (MSA, fusha)](https://broadrouter.com/es/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): El árabe estándar moderno (MSA) es la forma estandarizada del árabe que se usa en la escritura, las noticias, la educación y el habla formal en todo el mundo árabe. No es la lengua materna cotidiana de nadie: la gente habla dialectos regionales, como el egipcio o el del Golfo, y cambia al MSA en contextos formales. - [Sistema multiagente (multi-agent system)](https://broadrouter.com/es/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): Un sistema multiagente es una aplicación de LLM en la que varios agentes, cada uno con su propio prompt, herramientas y contexto, trabajan en partes de una misma tarea. Un diseño común es orquestador-trabajador: un agente líder planifica, lanza agentes trabajadores en paralelo y combina sus resultados. - [IA multimodal (multimodal AI)](https://broadrouter.com/es/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): La IA multimodal es un modelo o sistema que trabaja con más de un tipo de datos, como texto, imágenes, vídeo y audio. Un modelo multimodal puede leer una foto y una pregunta a la vez, o recibir voz y responder en texto. El modelo de visión y lenguaje es el tipo más común. - [Reconocimiento de entidades nombradas (NER)](https://broadrouter.com/es/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): El reconocimiento de entidades nombradas (NER) es la tarea de encontrar nombres en un texto y etiquetar cada uno con un tipo, como persona, organización o lugar. Los sistemas clásicos etiquetan cada token; a los LLMs, en cambio, se les suele pedir una lista de entidades en JSON, que luego se contrasta con el texto original. - [Reconocimiento óptico de caracteres (OCR)](https://broadrouter.com/es/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): El reconocimiento óptico de caracteres (OCR) es la tecnología que detecta el texto impreso o manuscrito de una imagen, un escaneo o un PDF y lo convierte en caracteres legibles por máquina. El OCR clásico devuelve el texto y su posición en la página; los modelos de visión y lenguaje pueden leer el texto y responder preguntas sobre él directamente. - [Ogg Opus (formato de notas de voz de WhatsApp)](https://broadrouter.com/es/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus es audio codificado con el códec Opus (RFC 6716) y guardado en un contenedor Ogg (RFC 7845). Está pensado para voz y audio interactivo a bajas tasas de bits, y es el formato que WhatsApp usa para las notas de voz, normalmente como archivos .ogg u .opus. - [Modelo de pesos abiertos (open-weight model)](https://broadrouter.com/es/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): Un modelo de pesos abiertos es un modelo de aprendizaje automático cuyos pesos entrenados se publican para que cualquiera los descargue y ejecute, normalmente con una licencia con condiciones. Se diferencia de la IA totalmente de código abierto, que además publica el código de entrenamiento e información detallada sobre los datos, para que otros puedan estudiar y reproducir el modelo. - [API compatible con OpenAI](https://broadrouter.com/es/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): Una API compatible con OpenAI es una API HTTP que acepta los mismos formatos de solicitud y respuesta que la API de OpenAI, como POST /v1/chat/completions, de modo que los SDKs y herramientas de OpenAI pueden usarla cambiando solo la URL base, la clave de API y el nombre del modelo. - [Llamadas a herramientas en paralelo (parallel tool calls)](https://broadrouter.com/es/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Las llamadas a herramientas en paralelo se producen cuando un modelo devuelve más de una llamada a herramienta en una sola respuesta, por ejemplo tres lecturas de archivo a la vez. En el formato de OpenAI llegan como varias entradas en tool_calls, y parallel_tool_calls en false limita al modelo a cero o una llamada por turno. - [Pago por uso (pay-as-you-go)](https://broadrouter.com/es/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): El pago por uso es un modelo de precios en el que usted paga solo lo que usa, medido por unidades como tokens, solicitudes o minutos de audio, sin suscripción ni compromiso mínimo. La mayoría de las APIs de LLM cobran así, desde un saldo prepagado o con una factura periódica. - [PDPL (leyes de protección de datos de Arabia Saudí y EAU)](https://broadrouter.com/es/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) es el nombre común de la ley saudí de datos personales, promulgada por el Real Decreto M/19 y aplicada por la SDAIA, y del Decreto-ley Federal n.º 45 de 2021 de los EAU. Ambas regulan el tratamiento de datos personales y condicionan su transferencia al extranjero, incluidos los prompts a APIs de LLM extranjeras. - [PII (información de identificación personal)](https://broadrouter.com/es/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): La PII (información de identificación personal) es información que permite distinguir o rastrear la identidad de una persona, como su nombre, un número de identificación o un registro biométrico, más otra información vinculable a ella. El RGPD habla de datos personales, un término más amplio. Con LLMs, se anonimiza antes de que los prompts salgan de sus sistemas. - [Prefill y decode (fases de inferencia de un LLM)](https://broadrouter.com/es/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill y decode son las dos fases de la inferencia de un LLM. En el prefill, el modelo procesa todo el prompt en paralelo y llena su caché KV; en el decode, genera la respuesta token a token reutilizando esa caché. El prefill determina sobre todo el tiempo hasta el primer token; el decode, la velocidad de salida. - [Crédito prepago (prepaid credit)](https://broadrouter.com/es/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): El crédito prepago es dinero que se añade a una cuenta de API antes de usarla. Cada solicitud descuenta su coste del saldo y, cuando el saldo se agota, las solicitudes medidas se detienen hasta que usted recargue. Limita el gasto a lo que ha pagado. - [Ingeniería de prompts (prompt engineering)](https://broadrouter.com/es/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): La ingeniería de prompts es el proceso de escribir, estructurar y probar las instrucciones que se dan a un modelo de lenguaje para que produzca de forma constante una salida que cumpla sus requisitos. Abarca los mensajes de rol, instrucciones claras, ejemplos, formato y la medición de resultados con casos de prueba. - [Inyección de prompts (prompt injection)](https://broadrouter.com/es/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): La inyección de prompts es un ataque en el que un texto que recibe un modelo de lenguaje, escrito por el usuario u oculto en una web, un correo, un archivo o el resultado de una herramienta, cambia su comportamiento contra la intención del desarrollador. OWASP la sitúa primera (LLM01:2025) en su Top 10 para aplicaciones LLM. - [Cuantización (quantization)](https://broadrouter.com/es/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): La cuantización es el proceso de almacenar los números de una red neuronal, sobre todo sus pesos, en un formato de menor precisión, como 8 o 4 bits en lugar de 16, para que el modelo use menos memoria y a menudo funcione más rápido. Los buenos métodos mantienen pequeña la pérdida de precisión. - [RAG (generación aumentada por recuperación)](https://broadrouter.com/es/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG (generación aumentada por recuperación) es un patrón en el que una aplicación primero recupera, de un almacén externo como un índice de búsqueda o vectorial, fragmentos relevantes para una pregunta, y luego los agrega al prompt para que un modelo de lenguaje responda con ese texto y no solo de memoria. - [Límite de frecuencia (rate limit)](https://broadrouter.com/es/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): Un límite de frecuencia (rate limit) es un tope que una API impone a cuánto puede usarla un cliente en un periodo, como solicitudes por minuto o tokens por minuto. Las solicitudes que lo superan se rechazan, normalmente con HTTP 429 Too Many Requests, hasta que el uso vuelve a quedar por debajo. - [ReAct (razonar y actuar)](https://broadrouter.com/es/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (reason and act, razonar y actuar) es un patrón de prompting, presentado por Yao et al. en 2022, en el que un modelo de lenguaje intercala trazas de razonamiento con acciones como búsquedas o llamadas a herramientas. Lee cada observación, actualiza su plan y repite hasta poder dar una respuesta final. - [Tokens de razonamiento (reasoning tokens)](https://broadrouter.com/es/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): Los tokens de razonamiento son tokens que un modelo de razonamiento genera para resolver un problema antes de escribir su respuesta visible. Normalmente se facturan como tokens de salida, cuentan para el límite de salida y la ventana de contexto, y la API puede ocultarlos o resumirlos. - [Timeout de solicitud (tiempo de espera)](https://broadrouter.com/es/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): Un timeout de solicitud es el tiempo máximo que un cliente HTTP espera a un servidor antes de abandonar la solicitud. Los clientes suelen dividirlo en un timeout de conexión, para abrir la conexión, y uno de lectura, para esperar los datos. Las llamadas a LLMs necesitan timeouts de lectura generosos porque las respuestas largas tardan. - [Reranking (reordenamiento de resultados)](https://broadrouter.com/es/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): El reranking es una segunda etapa de recuperación en la que un modelo más preciso pero más lento, normalmente un cross-encoder, puntúa cada par de consulta y documento candidato de una primera búsqueda y los reordena, de modo que solo los pocos más relevantes pasan al modelo de lenguaje. - [Texto de derecha a izquierda (RTL)](https://broadrouter.com/es/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): El texto de derecha a izquierda (RTL) es escritura, como la árabe o la hebrea, que se muestra de derecha a izquierda. Las computadoras lo guardan en orden lógico de lectura, y el algoritmo bidireccional de Unicode decide el orden de visualización cuando se mezcla con texto de izquierda a derecha, dígitos, nombres de modelos o código. - [Frecuencia de muestreo (sample rate)](https://broadrouter.com/es/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): La frecuencia de muestreo es la cantidad de muestras de audio capturadas por segundo, medida en hercios (Hz). Los modelos de reconocimiento de voz suelen esperar audio mono a 16,000 Hz (16 kHz). Junto con la profundidad de bits y el número de canales, determina cuánto ocupa por minuto un archivo de audio sin comprimir. - [Parámetro seed (semilla)](https://broadrouter.com/es/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): El parámetro seed (semilla) es un entero que usted envía con la solicitud para que la API intente muestrear los mismos tokens cuando se repite la misma solicitud. OpenAI lo describe como un mejor esfuerzo: el determinismo no está garantizado, y el campo system_fingerprint muestra cuándo cambió la configuración del backend. - [Inferencia serverless (serverless inference)](https://broadrouter.com/es/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): La inferencia serverless es una forma de ejecutar modelos de machine learning en la que la plataforma inicia, escala y detiene automáticamente el cómputo detrás de un endpoint, a menudo hasta cero cuando está inactivo, y cobra por uso en lugar de por máquinas reservadas. La contrapartida es un posible retraso por arranque en frío. - [SLA (acuerdo de nivel de servicio)](https://broadrouter.com/es/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): Un SLA (acuerdo de nivel de servicio) es un compromiso contractual de un proveedor con un nivel de servicio medible, normalmente un porcentaje de disponibilidad mensual, con una compensación definida, normalmente créditos de servicio, si no se cumple. El SLA define qué cuenta como caída, qué queda excluido y cómo reclamar. - [Decodificación especulativa (speculative decoding)](https://broadrouter.com/es/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): La decodificación especulativa es una técnica de inferencia en la que un modelo borrador pequeño y rápido propone varios tokens siguientes y el modelo objetivo grande los verifica en una sola pasada hacia adelante, conservando los aceptados. Acelera la generación sin cambiar la distribución de la salida. - [Voz a texto (speech-to-text)](https://broadrouter.com/es/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): La conversión de voz a texto, o reconocimiento automático del habla (ASR), transforma audio hablado en texto escrito. Una API de voz a texto recibe un archivo o stream de audio y devuelve una transcripción, normalmente cobrada por minuto. La precisión se mide con la tasa de error por palabra y varía según el acento, el dialecto y la grabación. - [Límite de gasto (spend limit)](https://broadrouter.com/es/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): Un límite de gasto es un tope a cuánto puede gastar una cuenta, un proyecto o una clave de API en un periodo, normalmente un mes. Cuando el uso lo alcanza, el proveedor rechaza las solicitudes siguientes, a menudo con un error HTTP 429 o 400, hasta que el periodo se reinicia o se aumenta el límite. - [Subtítulos SRT (SRT subtitles)](https://broadrouter.com/es/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) es un formato de subtítulos en texto plano. Cada subtítulo consta de un número de secuencia, una hora de inicio y fin escrita como hh:mm:ss,mmm --> hh:mm:ss,mmm, una o más líneas de texto y una línea en blanco. WebVTT es el formato web similar que usa el video en HTML. - [Server-sent events (SSE)](https://broadrouter.com/es/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Server-sent events (SSE), o eventos enviados por el servidor, es un estándar web para que un servidor envíe un flujo de eventos de texto a un cliente a través de una única respuesta HTTP de larga duración con el tipo de contenido text/event-stream. Cada evento es una o más líneas de campo, como data:, terminadas por una línea en blanco. - [Secuencias de parada (stop sequences)](https://broadrouter.com/es/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): Las secuencias de parada (stop sequences) son cadenas que usted envía con la solicitud para que el modelo deje de generar en cuanto produzca una de ellas. OpenAI llama al campo stop y acepta hasta 4 secuencias; Anthropic lo llama stop_sequences. Normalmente el texto coincidente no se incluye en la salida devuelta. - [Streaming (respuestas en tiempo real)](https://broadrouter.com/es/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): El streaming es una forma de devolver la respuesta de un modelo de lenguaje en pequeños fragmentos mientras aún se genera, en lugar de una sola respuesta al final. Las APIs compatibles con OpenAI hacen streaming mediante server-sent events cuando la solicitud define stream como true. - [Salida estructurada (structured output)](https://broadrouter.com/es/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): La salida estructurada es una función de las APIs de LLM que restringe la respuesta del modelo a un formato que define el desarrollador, normalmente un JSON Schema, para que el código pueda procesarla de forma fiable. Es más estricta que JSON mode, que garantiza JSON válido pero no una forma concreta. - [Subagente (subagent)](https://broadrouter.com/es/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): Un subagente es un agente independiente que un agente principal inicia para resolver una tarea concreta. Se ejecuta con su propia ventana de contexto, prompt de sistema y acceso a herramientas, hace el trabajo y devuelve al agente principal solo un resultado o resumen, lo que mantiene pequeña la conversación principal. - [Subencargado del tratamiento (subprocessor)](https://broadrouter.com/es/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): Un subencargado del tratamiento es un tercero al que un encargado recurre para tratar datos personales por cuenta del responsable. En una API de LLM pueden ser hosts de GPU, proveedores de inferencia asociados o capacidad de respaldo. Según el artículo 28 del RGPD, requieren autorización del responsable y las mismas obligaciones de protección de datos. - [Prompt de sistema (system prompt)](https://broadrouter.com/es/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): Un prompt de sistema es el texto de instrucciones que una aplicación envía a un modelo de lenguaje antes de la conversación, normalmente como mensaje con el rol system o developer. Define el rol, las reglas, el tono y el formato de salida del modelo, que debe priorizarlo sobre mensajes del usuario que lo contradigan. - [Taa marbuta (ة)](https://broadrouter.com/es/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): La taa marbuta (ة, U+0629) es una letra árabe que solo aparece al final de una palabra y suele marcar sustantivos y adjetivos femeninos. Se parece a la heh (ه, U+0647) con dos puntos, por lo que el texto escrito mezcla ambas, y muchos normalizadores de búsqueda convierten ة en ه. - [Temperatura (temperature)](https://broadrouter.com/es/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): La temperatura es un parámetro de muestreo que controla qué tan aleatoria es la salida de un modelo de lenguaje. El modelo divide las puntuaciones de los tokens entre la temperatura antes de convertirlas en probabilidades: los valores bajos hacen dominar al token más probable y los altos aplanan la distribución, de modo que los menos probables salen más seguido. - [Texto a voz (TTS, text-to-speech)](https://broadrouter.com/es/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): Texto a voz (TTS), o síntesis de voz, convierte texto escrito en audio hablado. Una API de TTS recibe un texto, una voz y un formato de salida como MP3 o WAV, y devuelve voz generada. Es lo contrario de voz a texto, que convierte audio en texto. - [Throughput (rendimiento total)](https://broadrouter.com/es/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): El throughput, en el servicio de LLMs, es la cantidad total de trabajo que un sistema completa por unidad de tiempo en todas las solicitudes concurrentes, normalmente medida en tokens de salida por segundo o solicitudes por segundo. Es distinto de la velocidad por stream, que es qué tan rápido recibe sus tokens una sola solicitud. - [Tiempo hasta el primer token (TTFT)](https://broadrouter.com/es/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): El tiempo hasta el primer token (TTFT) es lo que pasa desde que se envía una solicitud a un modelo de lenguaje hasta que llega el primer token de salida. Incluye la red, la cola del servidor y el prefill, cuando el modelo lee todo el prompt, así que crece con el largo del prompt y la carga. - [Marcas de tiempo en voz a texto (timestamps)](https://broadrouter.com/es/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): Las marcas de tiempo en voz a texto son tiempos de inicio y fin, en segundos, asociados a partes de una transcripción. Las de segmento cubren una frase u oración; las de palabra cubren cada palabra. Se necesitan para subtítulos, para buscar dentro de grabaciones y para alinear el texto con los hablantes. - [Token](https://broadrouter.com/es/glossary/token/) (translation of https://broadrouter.com/glossary/token/): Un token es la unidad de texto que un modelo de lenguaje lee y escribe: una palabra, parte de una palabra, un signo de puntuación o una secuencia de bytes producida por el tokenizador del modelo. Las APIs de LLM cuentan, limitan y normalmente facturan el uso en tokens. - [Tokenizador (tokenizer)](https://broadrouter.com/es/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): Un tokenizador es el componente que convierte texto en la secuencia de IDs de token enteros que lee un modelo de lenguaje, y que vuelve a convertir en texto los IDs generados. Cada modelo trae su propio tokenizador, así que el mismo texto puede producir un número distinto de tokens en modelos distintos. - [Tokens por segundo (TPS)](https://broadrouter.com/es/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): Los tokens por segundo (TPS) son la velocidad a la que un modelo de lenguaje genera tokens de salida, normalmente medida en una solicitud tras llegar su primer token. Marcan qué tan rápido fluye la respuesta. Sumados en todas las solicitudes concurrentes son el throughput del sistema; la espera hasta el primer token se mide aparte. - [Llamadas a herramientas (tool calling)](https://broadrouter.com/es/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): Las llamadas a herramientas (tool calling), también llamadas function calling, son una función de las APIs de LLM en la que el modelo devuelve una solicitud estructurada para ejecutar una función definida por el desarrollador, con argumentos JSON. La aplicación ejecuta la función y devuelve el resultado; el modelo nunca ejecuta código. - [tool_choice (elección de herramienta)](https://broadrouter.com/es/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice es un parámetro de la solicitud que controla el uso de herramientas. En Chat Completions de OpenAI puede ser auto (el modelo decide), required (debe llamar a una o más herramientas), none (sin herramientas) o una función con nombre para forzarla. Las opciones equivalentes de Anthropic son auto, any, tool y none. - [Top-p (nucleus sampling o muestreo de núcleo)](https://broadrouter.com/es/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Top-p, también llamado muestreo de núcleo (nucleus sampling), es un método de decodificación en el que un modelo de lenguaje elige el siguiente token solo entre el conjunto más pequeño de tokens más probables cuya probabilidad combinada alcanza un umbral p, como 0.9, y descarta la cola improbable. - [Normalización Unicode (Unicode normalization)](https://broadrouter.com/es/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): La normalización Unicode convierte el texto a una forma estándar para que las cadenas que se ven iguales usen los mismos puntos de código. Las cuatro formas son NFD, NFC, NFKD y NFKC; las formas K también convierten caracteres de compatibilidad, como las formas de presentación árabes y las ligaduras, en letras normales. - [Detección de actividad de voz (VAD)](https://broadrouter.com/es/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): La detección de actividad de voz (VAD) es un paso de procesamiento de señal que clasifica tramas cortas de audio como habla o no habla. Los sistemas de voz la usan para saltar silencios, dividir grabaciones largas en las pausas, detectar cuándo un hablante terminó y evitar transcribir ruido de fondo. - [Base de datos vectorial (vector database)](https://broadrouter.com/es/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): Una base de datos vectorial es un almacén de datos que guarda vectores de embeddings, a menudo con su texto de origen y metadatos, y responde consultas de vecinos más cercanos: dado un vector de consulta, devuelve los vectores guardados más parecidos, normalmente por distancia coseno, distancia euclidiana o producto interno. - [Modelo de visión y lenguaje (VLM)](https://broadrouter.com/es/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): Un modelo de visión y lenguaje (VLM) es un modelo que recibe imágenes y texto a la vez y produce texto, por lo que puede describir imágenes, responder preguntas sobre ellas y leer documentos, gráficos o capturas de pantalla. También se les llama modelos image-text-to-text o multimodales. - [vLLM](https://broadrouter.com/es/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM es una biblioteca de código abierto para inferencia y servicio de LLM, desarrollada originalmente en el Sky Computing Lab de UC Berkeley. Gestiona la memoria de la caché KV con PagedAttention, agrupa las solicitudes entrantes de forma continua, admite muchos formatos de cuantización y ofrece un servidor HTTP compatible con OpenAI. - [Tamaño de vocabulario (vocabulary size)](https://broadrouter.com/es/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): El tamaño de vocabulario es el número de tokens distintos en el tokenizador de un modelo de lenguaje, cada uno con su propio ID y embedding. Un vocabulario más grande puede codificar texto común con menos tokens, pero agranda las capas de embedding y de salida, y solo ayuda con el texto con el que se entrenó el vocabulario. - [WAV (archivo de audio)](https://broadrouter.com/es/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format, o WAVE) es un contenedor de audio basado en el formato RIFF de Microsoft e IBM. Suele llevar muestras PCM lineales sin comprimir, así que su tamaño es predecible: frecuencia de muestreo por bytes por muestra por canales, por segundo. - [Tasa de error por palabra (WER)](https://broadrouter.com/es/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): La tasa de error por palabra (WER) es la medida estándar de precisión del reconocimiento de voz: el número de sustituciones, eliminaciones e inserciones de palabras necesarias para convertir una transcripción en la referencia correcta, dividido entre el número de palabras de la referencia. Cuanto más baja, mejor. - [Retención cero de datos (zero data retention, ZDR)](https://broadrouter.com/es/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): La retención cero de datos (ZDR) es una política de manejo de datos en la que un proveedor de IA procesa los prompts y las salidas solo para atender una solicitud y después no guarda ese contenido en logs, cachés ni datos de entrenamiento. Los registros sin contenido, como conteos de tokens y tiempos, normalmente se conservan. ### German (de, Deutsch): 294 translated pages - [Articles in German](https://broadrouter.com/de/blog/) - [LLM-API-Preisänderungen im Oktober 2026: 3 Erhöhungen, keine Senkung](https://broadrouter.com/de/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 Änderungen auf Preisseiten bei 5 von 7 LLM-API-Anbietern im Oktober 2026: 3 Erhöhungen, keine Senkung, 9 neue Modellpreise, 11 entfernt. Mit Archiv-Links. - [Kosten eines KI-Coding-Agent-Laufs: Schritte, Kontext, Caching](https://broadrouter.com/de/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): Warum ein Coding-Agent pro Schritt teurer wird, je länger er läuft, wie Caching die Summe ändert und wann eine Flatrate-Session gewinnt. Mit Token-Rechnung. - [Arabischen Text per LLM vokalisieren: Prompt, Prüfung, DER](https://broadrouter.com/de/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): Arabische Diakritika (Taschkil) per LLM-API ergänzen: ein Prompt, der jeden Buchstaben erhält, eine Prüfung gegen Änderungen, DER/WER und wann CATT passt. - [Arabische Grammatik- und Rechtschreibkorrektur per LLM: Anleitung](https://broadrouter.com/de/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Arabische Rechtschreibung und Grammatik per LLM-API korrigieren: Prompt mit minimalen Änderungen, Schutz vor Umformulierung, Edits nach Fehlerklasse, M2-Bewertung. - [Arabische Besprechungsprotokolle aus Aufnahmen: eine API-Pipeline](https://broadrouter.com/de/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Arabische Meeting-Aufnahmen werden zu Protokollen mit Beschlüssen und Aufgaben: transkribieren, bereinigen, per LLM als JSON zusammenfassen, Namen und Daten prüfen. - [Arabische Named Entity Extraction mit LLM: JSON, Daten, Prüfungen](https://broadrouter.com/de/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Arabische Namen, Organisationen, Orte und Hidschri-Daten per LLM als JSON extrahieren: Prompt, Span-Prüfung ohne Diakritika, Datumsumrechnung und Bewertungsskript. - [OCR-APIs für Arabisch: Google, Azure, Textract, Mistral, LLMs](https://broadrouter.com/de/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): Welche OCR-APIs Arabisch offiziell unterstützen, gedruckt und handschriftlich, was sie liefern und was 1.000 Seiten kosten, im Vergleich mit Vision-LLM-APIs. - [Arabische Untertitel (SRT) per Speech-to-Text-API: was Sie brauchen](https://broadrouter.com/de/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): Was eine SRT-Datei braucht, welche Speech-to-Text-APIs Zeitstempel oder SRT für Arabisch liefern, und eine getestete Pipeline für Untertitel aus einer Text-API. - [Arabische Textklassifikation mit LLM: Sentiment, Intent, Triage](https://broadrouter.com/de/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): Arabische Bewertungen, Nachrichten und Tickets per LLM-API klassifizieren: Labels, dialektbewusster Prompt, JSON-Ausgabe, Test mit Macro-F1, Kosten bei Volumen. - [Arabische Textnormalisierung für Suche und RAG: was ändern, was nicht](https://broadrouter.com/de/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Taschkil, Alif-Formen, Taa marbuta, Hamza-Träger, persische Buchstaben, Ziffern, Unicode: welche Regeln der Suche helfen, welche schaden, mit getestetem Python. - [Arabische Texte mit LLM zusammenfassen: Prompts, Länge, Bewertung](https://broadrouter.com/de/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): Arabische Artikel, Berichte und Dialektnachrichten per LLM-API zusammenfassen: Prompt-Muster, Wortzahl, Zahlenprüfung und warum ROUGE für Arabisch 0 liefert. - [Alternativen zu Artificial Analysis: LLMs vergleichen im Jahr 2026](https://broadrouter.com/de/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Alternativen zu Artificial Analysis im Vergleich: Arena, Epoch AI, Vals AI, OpenRouter-Rankings, Hugging Face und Preisdaten, nach Messgröße und Einsatzzweck. - [GPU-Hardware und Serving von BroadNet, wie BroadNet sie beschreibt](https://broadrouter.com/de/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): Was BroadNet über die Hardware hinter BroadNet Inference sagt: drei GPU-Stufen von B300 bis Kunden-Appliance, Durchsatz pro GPU und Erkenntnisse aus den Tests. - [LLM-Antworten cachen, Kosten senken: ein Exact-Match-Cache in Python](https://broadrouter.com/de/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Ein getesteter Exact-Match-Cache für LLM-API-Antworten in Python: was in den Schlüssel gehört, was nie, TTLs, Datenschutz und wie Sie die Trefferquote vorab messen. - [Claude Code mit offenen Modellen: Proxys, native Endpoints, Grenzen](https://broadrouter.com/de/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code spricht die Anthropic Messages API. So betreiben Sie es mit offenen Modellen über einen LiteLLM-Proxy oder einen nativen Endpoint, und was dabei hakt. - [Menschliche Prüfung als festen Teil Ihres Dokumenten-KI-Workflows](https://broadrouter.com/de/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): Den menschlichen Prüfschritt einer Dokumenten-KI-Pipeline planen: Validierungsregeln, was ein Mensch sieht, Zufallsstichproben, Audit-Trail und Prüfquote. - [Ägyptens Datenschutzgesetz (Gesetz 151 von 2020) und LLM-APIs](https://broadrouter.com/de/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Personenbezogene Daten aus Ägypten an eine LLM- oder Speech-to-Text-API im Ausland senden: Gesetz 151/2020 und Beschluss 816/2025 zu Transfers, Lizenzen, Pannen. - [Vor dem Fine-Tuning: das Problem klar benennen](https://broadrouter.com/de/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): Wann Prompting, RAG oder Evaluierung das Problem lösen und Fine-Tuning unnötig ist, welche Trainingsdaten Sie brauchen und wie Sie ein Fine-Tuning vorab testen. - [Ungültiges JSON vom LLM reparieren: Ursachen, Reparatur, Retry](https://broadrouter.com/de/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): Warum LLMs kaputtes JSON liefern (Code-Fences, Fließtext, Abbruch bei max_tokens, Kommas) und eine getestete Python-Funktion, die nur Eindeutiges repariert. - [Wie oft ändern sich LLM-API-Preise? 3 Jahre archivierte Preisseiten](https://broadrouter.com/de/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): 308 archivierte Preisseiten von 11 LLM-API-Anbietern, 2023 bis 2026: 80 % der Modelle blieben im Preis gleich; 75 % aller Änderungen kamen von DeepInfra. - [LLM-API-Kosten senken: eine Checkliste in 10 Schritten, in Reihenfolge](https://broadrouter.com/de/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): LLM-API-Kosten in der richtigen Reihenfolge senken: messen, Präfix cachen, Prompts kürzen, Output und Reasoning deckeln, Modell anpassen, Batch, Flatrate. - [LLM-API-Failover: im eigenen Code auf einen Zweitanbieter umschalten](https://broadrouter.com/de/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): Failover zwischen OpenAI-kompatiblen LLM-Anbietern: bei welchen Fehlern umschalten, Circuit Breaker, Modell-IDs, Streaming, Kosten und gleicher Datenschutz. - [Preissenkungen bei LLM-APIs 2023–2026: die größten, mit Quellen](https://broadrouter.com/de/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): Die größten Preissenkungen bei LLM-APIs 2023 bis 2026, jeweils mit Archiv-Links vorher und nachher: o3, Codestral, Gemini 1.5 und die kaum bemerkten Erhöhungen. - [Sinken LLM-API-Preise nur? 52 Preiserhöhungen und wie Sie planen](https://broadrouter.com/de/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): LLM-API-Preise steigen auch. In archivierten Preisseiten fanden wir 52 Erhöhungen für dasselbe Modell, 2024 bis 2026, im Median 40 %. Mit Quellen und Budgetregel. - [Rate Limits von LLM-APIs nach Anbieter: RPM, TPM und Parallelität](https://broadrouter.com/de/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): Rate Limits der Einstiegsstufen von 13 LLM-API-Anbietern laut eigener Doku: OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks u. a., plus was RPM und TPM heißt. - [LLM-API-SLAs: was Azure, Bedrock, Google und OpenAI zusagen](https://broadrouter.com/de/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): Was jeder LLM-API-Anbieter schriftlich zusagt: Verfügbarkeit in Prozent, was als Ausfall zählt, Service-Gutschriften und Latenzziele, direkt aus den SLA-Seiten. - [LLM-Batch-Verarbeitung ohne Batch API: ein fortsetzbarer Runner](https://broadrouter.com/de/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): 10.000 bis 1.000.000 Prompts über eine Chat-Completions-API ohne Batch API: Parallelitätslimits, Taktung, Retries, fortsetzbare Checkpoints und die Zeitrechnung. - [Die Safety Map: Wie 39 LLMs mit sensiblen Daten umgehen](https://broadrouter.com/de/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet hat 39 LLMs von 14 Labs zu Zugangsdaten, PII, Gesundheits- und Finanzdaten geprüft: was sich übertrug, was nicht, und Lecks über Tool Calls und Reasoning. - [LLM auf Sicherheit bei sensiblen Daten testen: BroadNets Methode](https://broadrouter.com/de/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): Die veröffentlichte Methode von BroadNet, um vor dem Deployment zu testen, ob ein LLM Zugangsdaten, PII, Gesundheits- oder Finanzdaten leakt: Kategorien, 3+ Läufe. - [Weg von der OpenAI API: Migration mit Code in Python, Node, cURL](https://broadrouter.com/de/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): Die Code-Änderungen für den Wechsel von OpenAI zu einer OpenAI-kompatiblen API: Client, Responses zu Chat Completions, Tools, Streaming, Embeddings, Smoke-Test. - [Welche Modelle BroadNet getestet und aussortiert hat, und warum](https://broadrouter.com/de/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet veröffentlicht getestete, abgelehnte Modelle mit Gründen: Chat-Modelle, die am Credential-Safety-Test scheiterten, plus Sprach-, OCR- und Embedding-Modelle. - [Mehrsprachige KI-Produkte gestalten: Arabisch, Englisch, Spanisch](https://broadrouter.com/de/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): KI-Produkt für Arabisch, Englisch und Spanisch gestalten: UI-Locale vs. Modellsprache, RTL und gemischter Text, Zahlen, Datumsangaben, Erkennung, Kosten. - [API-Kosten: Open-Weight- vs. geschlossene Modelle laut Preisdaten](https://broadrouter.com/de/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): Sind Open-Weight-Modelle per API günstiger als GPT, Claude oder Gemini? Preise pro Token von den Anbieterseiten, Spannen zwischen Hosts und Kosten pro Nutzer. - [Preisverlauf der APIs von OpenAI, Anthropic und Google, 2023–2026](https://broadrouter.com/de/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): API-Preise von GPT, Claude und Gemini im Vergleich, aus archivierten Preisseiten 2023 bis 2026: welche Reihen billiger, welche teurer wurden, mit Archiv-Links. - [Basis-URLs OpenAI-kompatibler APIs: 14 Anbieter geprüft](https://broadrouter.com/de/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): Die base_url von 14 OpenAI-kompatiblen LLM-APIs, jeweils aus der Doku des Anbieters, mit den gelisteten Endpoints und Eigenheiten, die einen Wechsel stören. - [OpenClaw und Hermes Agent mit günstigem Modell: Setup und Kosten](https://broadrouter.com/de/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): OpenClaw oder Hermes Agent an einen OpenAI-kompatiblen Endpunkt anbinden, verstehen, warum geplante Checks Tokens fressen, und mit weniger Kontext Kosten senken. - [Ein privater Endpoint beginnt mit einem klar definierten Workload](https://broadrouter.com/de/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): Was Sie vor der Anfrage nach einem privaten LLM-Endpoint festlegen sollten: Workload, Tokens pro Sekunde, Parallelität, Kontext, Region, Lizenz, Budget und Test. - [Prompt Injection bei Coding-Agenten: Schutzmaßnahmen, die halten](https://broadrouter.com/de/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): Wo Prompt Injection einen Coding-Agent erreicht und welche Schutzschichten Anbieter und Forschung dokumentieren: Sandbox, Egress-Sperre, Allowlists, Freigaben. - [RAG-Chunk-Größe und Overlap: Wirkung auf Retrieval und Kosten](https://broadrouter.com/de/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): Wie Chunk-Größe und Overlap Recall, Indexgröße, Embedding-Kosten und die Tokens pro LLM-Anfrage verändern, mit veröffentlichten Ergebnissen und Rechenbeispielen. - [max_tokens für Reasoning-Modelle festlegen: eine Budgetmethode](https://broadrouter.com/de/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Warum Reasoning-Modelle bei kleinem max_tokens leer antworten und wie Sie das Budget aus Perzentilen, Kontextfenster und eskalierendem Retry bemessen. - [Audio im Browser aufnehmen und an eine Speech-to-Text API senden](https://broadrouter.com/de/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): Mikrofon mit MediaRecorder aufnehmen, WebM oder Safari-MP4 in 16-kHz-WAV umwandeln und über einen kleinen Server, der den Schlüssel hält, transkribieren. - [Prompt-Tokens reduzieren: gemessene Effekte bei Prompts, Tools, Daten](https://broadrouter.com/de/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): Gemessene Token-Ersparnis durch kürzere System-Prompts, Few-Shot-Beispiele, Tool-Schemas, Datenformate, Chat-Verlauf und arabische Diakritika, mit echten Tokenizern. - [Gleiches Modell, anderer Anbieter: Wie stark LLM-API-Preise schwanken](https://broadrouter.com/de/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): Bei 119 Modellen mit mehreren Hosts liegt der teuerste Listenpreis im Median beim 1,76-Fachen des günstigsten. Daten von 41 Preisseiten der Anbieter, November 2026. - [PDPL in Saudi-Arabien und den VAE: Checkliste für LLM-API-Transfers](https://broadrouter.com/de/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Personenbezogene Daten aus Saudi-Arabien oder den VAE an eine LLM-API senden: Transferregeln der PDPL, Fragen an den Anbieter, wann Inferenz im Land zählt. - [Text, Sprache oder Vision? Beginnen Sie mit der Aufgabe](https://broadrouter.com/de/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): Text-Chat, Bild-Input oder Audiotranskription nach Aufgabe wählen, mit den Modellen und Limits, die BroadNet dokumentiert, und was fehlt, etwa Text-to-Speech. - [Audio über 25 MB transkribieren: aufteilen, senden, zusammenfügen](https://broadrouter.com/de/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): So transkribieren Sie Aufnahmen über dem Dateilimit einer Speech-to-Text API: Datei verkleinern, an Pausen mit ffmpeg teilen, parallel senden, Text zusammenfügen. - [App-Strings per LLM ins Arabische übersetzen: Platzhalter und Plurale](https://broadrouter.com/de/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): i18n-JSON-Dateien per LLM-API ins Arabische übersetzen: Platzhalter einfrieren, alle sechs arabischen Pluralformen erzeugen, ICU in CI prüfen, Namen isolieren. - [Arabisches Audio per API ins Englische übersetzen: getestete Pipeline](https://broadrouter.com/de/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): Arabische Aufnahmen mit zwei API-Aufrufen in englischen Text verwandeln: mit language=ar transkribieren, per LLM übersetzen, mit Zahlenprüfung, Kosten und Bewertung. - [Live, auf Anfrage, Forschung: einen Modellkatalog richtig lesen](https://broadrouter.com/de/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): Welche Katalogmodelle Sie heute aufrufen können, welche eine Anfrage brauchen und welche nur getestet wurden, am Beispiel des BroadNet-Katalogs und /v1/models. - [Uno in vLLM: was die Serving-Studie von BroadNet gemessen hat](https://broadrouter.com/de/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet hat den Uno-Self-Drafting-Adapter in vLLM eingebaut und auf einer H100 gemessen: 2,58x Durchsatz bei einem Stream, 1,74x bei 32. Methode, DOI, Grenzen. - [Uno auf Gemma 4 26B: die Mixture-of-Experts-Studie von BroadNet](https://broadrouter.com/de/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet hat den ersten Uno-Self-Drafting-Adapter für ein Mixture-of-Experts-Modell trainiert, Gemma 4 26B A4B. Messungen, Nutzen von Self-Drafting und DOI. - [Agenten-Token-Budget: wohin die Tokens einer Agenten-Schleife gehen](https://broadrouter.com/de/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): System-Prompt, Tool-Schemas, gelesene Dateien, Verlauf, Reasoning: wie eine Agenten-Schleife Tokens verbraucht, wie Sie jeden Teil messen und was Sie zuerst kürzen. - [Wortfehlerrate (WER) bei arabischer Spracherkennung erklärt](https://broadrouter.com/de/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): Was die Wortfehlerrate bedeutet, wie man sie berechnet und warum arabische WER so stark auf Diakritika, Schreibnormalisierung und Dialekte reagiert. Mit Python. - [Spracherkennung arabischer Dialekte: Golf, Ägypten, Levante, Maghreb](https://broadrouter.com/de/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Warum Speech-to-Text an Golf-, ägyptischem, levantinischem und maghrebinischem Arabisch scheitert und wie ein Dialekt-Testset Anbieter vergleichbar macht. - [Übersetzung Arabisch–Englisch mit LLMs: Prompts, Glossare, Bewertung](https://broadrouter.com/de/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): Arabisch ins Englische und zurück per LLM-API übersetzen: Prompts, die den Ton halten, Glossare für feste Begriffe, Zahlen und Namen, und wie Sie Qualität bewerten. - [Golf-, ägyptisches und levantinisches Arabisch: Dialekte in LLMs](https://broadrouter.com/de/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): Wie LLMs mit Golf-, ägyptischem und levantinischem Arabisch umgehen: wo sie scheitern, wie Sie für Dialekt prompten und testen, was BroadNet zu Qwen3.8 27B meldet. - [Arabische OCR vs. Vision-LLM: Dokumente und Rechnungen lesen](https://broadrouter.com/de/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): Wann klassische OCR, wann ein Vision-LLM für arabische Rechnungen, Belege und Formulare: Fehlerfallen, JSON-Ausgabe, Kosten pro Seite und eine Validierungspipeline. - [Prompt Engineering auf Arabisch: Register, RTL, Ziffern, Mischtext](https://broadrouter.com/de/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): Prompts, die gutes Arabisch aus einem LLM holen: Register und Dialekt wählen, Ziffern und Daten steuern, RTL sauber formatieren, Arabisch-Englisch-Mischung regeln. - [RAG auf Arabisch: Normalisierung, Chunking und Retrieval, die wirken](https://broadrouter.com/de/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): Retrieval-Augmented Generation mit arabischen Dokumenten: Alef, Ya und Ta marbuta normalisieren, sinnvoll chunken, mehrsprachige Embeddings, Stichwortsuche. - [Speech-to-Text-API für Arabisch: So wählen Sie 2026 die richtige](https://broadrouter.com/de/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): So wählen Sie eine Transkriptions-API für Arabisch: Wortfehlerrate, Dialekte, Dateilimits, Preis pro Audiostunde, Datenschutz und ein Testplan für eigene Aufnahmen. - [Arabischer WhatsApp-Chatbot: von der Sprachnachricht zur Antwort](https://broadrouter.com/de/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): Eine funktionierende Architektur für einen arabischen WhatsApp-Assistenten: Webhook, Queue, Transkription, LLM-Antwort, Richtlinien und Kosten, mit Python-Code. - [Arabizi und LLMs: Franco-Arabisch mit Ziffern wie 3 und 7 verarbeiten](https://broadrouter.com/de/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): Arabizi: Arabisch in lateinischen Buchstaben, mit 3 für ع und 7 für ح. So erkennen Sie es, wandeln es per LLM in arabische Schrift um und antworten richtig. - [Bestes LLM für Aider: Modell auswählen und verbinden](https://broadrouter.com/de/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Das beste LLM für Aider finden: Disziplin beim Edit-Format, Kontextgröße, Kosten pro Stunde, ein einstündiger Test im eigenen Repo und das OPENAI_API_BASE-Setup. - [Bestes offenes Modell für Coding-Agenten: Tools, Kontext und Kosten](https://broadrouter.com/de/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B oder Gemma 4 für Ihren Coding-Agenten? Wählen Sie nach Tool Calling, Kontext, Reasoning und Kosten pro Aufgabe. - [Gecachte Input-Tokens: So senken Sie die Rechnung Ihrer LLM-API](https://broadrouter.com/de/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): Gecachter Input kostet einen Bruchteil von normalem Input. Wie Prompt Caching abgerechnet wird, wie Sie cached_tokens lesen und Prompts dafür aufbauen. - [Günstigste Vision-LLM-API: Kosten pro Bild und pro Seite berechnet](https://broadrouter.com/de/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): Was ein Vision-LLM-Aufruf pro Bild und pro Dokumentseite kostet: Token-Rechnung für Gemma 4, Qwen3.8 27B und DeepSeek V4.1 Flash, und wo DeepInfra günstiger ist. - [LLM-API-Anbieter auswählen: Checkliste mit 20 Punkten](https://broadrouter.com/de/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): Zwanzig Fragen vor der Wahl einer LLM-API: Preise, Limits, OpenAI-Kompatibilität, Datenschutz, Modelle, Support und Verfügbarkeit, mit Bewertungsbogen und Testplan. - [LLM-API für Arabisch auswählen: Was Sie testen sollten](https://broadrouter.com/de/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): So wählen Sie ein LLM für arabische Apps: Dialektverständnis, Schreibqualität, Übersetzung, arabische Dokumente und Bilder, Kosten und ein Testset an einem Tag. - [Cline vs. Roo Code vs. Aider für Teams mit offenen Modellen (2026)](https://broadrouter.com/de/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (archiviert, fortgeführt als Zoo Code) und Aider im Vergleich: Tool Calls, Kontextverbrauch, MCP, Git und Freigaben mit offenen Modellen. - [Tokens zählen, bevor Sie eine LLM-Anfrage senden](https://broadrouter.com/de/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): Prompt-Tokens je Sprache und Modell vor dem API-Aufruf schätzen: warum Tokenizer abweichen, Messwerte in fünf Sprachen und sichere Puffer für max_tokens. - [Coding-Agent-Kosten senken: 10 Wege mit geschätzter Ersparnis](https://broadrouter.com/de/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Zehn konkrete Wege, wie Cline, Aider, opencode oder Ihr eigener Coding-Agent weniger Tokens verbraucht, jeweils mit Rechnung zur Ersparnis und zur Messung. - [DeepSeek-API-Alternativen: 5 Wege, DeepSeek-Modelle zu nutzen](https://broadrouter.com/de/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): DeepSeek über die offizielle API, einen Aggregator, einen Cloud-Katalog, eigene GPUs oder BroadNet Inference: Vor- und Nachteile bei Preis, Datenschutz, Limits. - [DeepSeek V4.1 Flash API Preise: pro Token, gecacht und Flatrate](https://broadrouter.com/de/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): DeepSeek V4.1 Flash API Preise pro 1 Mio. Input-, gecachter und Output-Tokens, Flatrate pro Monat, Woche oder Stunde, Vergleich mit DeepSeek, DeepInfra, Fireworks. - [LLM-API-Kosten vor dem Launch eines Features schätzen](https://broadrouter.com/de/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): Ein Kostenmodell für ein neues KI-Feature: Nutzer × Aktionen × Aufrufe × Tokens, mit Cache, Reasoning, Retries, Rate Limits und einer Beispieltabelle pro Monat. - [Wie Sie ein LLM vor dem Anbieterwechsel evaluieren](https://broadrouter.com/de/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): Bauen Sie aus echtem Traffic ein kleines Eval-Set, bewerten Sie es mit Code-Checks und einem entzerrten LLM-Judge und vergleichen Sie Qualität, Kosten und Latenz. - [Rechnungs-, Beleg- und Formulardaten mit einem Vision-LLM als JSON](https://broadrouter.com/de/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): Eine Produktions-Pipeline für Rechnungen, Belege und Formulare: Bildvorbereitung, JSON-Prompt, Pydantic-Validierung, Rechenprüfungen, Retries und manuelle Prüfung. - [LLM-API: Flatrate oder Abrechnung pro Token? Die Break-even-Rechnung](https://broadrouter.com/de/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): Wann lohnt sich eine LLM-Flatrate statt Abrechnung pro Token? Vier Monatsrechnungen für DeepSeek V4.1 Flash, das Break-even-Volumen und wie Sie selbst messen. - [Kostenloses LLM-API-Guthaben: Wie weit Startguthaben wirklich reicht](https://broadrouter.com/de/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): Was kostenloses LLM-API-Guthaben bringt: Tokens, Audiominuten und Stundenschlüssel mit dem Startguthaben von BroadNet, verglichen mit Gemini und OpenRouter. - [DSGVO und LLM-APIs: 15 Fragen an Ihren Anbieter](https://broadrouter.com/de/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): DSGVO-Checkliste für LLM-APIs: AVV, Unterauftragsverarbeiter, Speicherung, Training, Drittlandübermittlung und Meldung von Datenpannen, mit den Artikeln dazu. - [GLM 5.3 vs. DeepSeek V4.1 Flash: welches Modell wann?](https://broadrouter.com/de/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 vs. DeepSeek V4.1 Flash im Vergleich: Preis, Kontext, Tools, Structured Output, Bilder und Reasoning, mit Kostenbeispielen und einer einfachen Faustregel. - [LLM-API-Fehler 429 behandeln: Backoff, Jitter und Limiter](https://broadrouter.com/de/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 vs. 5xx, exponentielles Backoff mit Full Jitter, Parallelitätslimiter und Warteschlangen für LLM-APIs, mit Python- und Node-Code für die Produktion. - [LLM-API-Schlüssel pro Stunde: der Stunden-Festpreis erklärt](https://broadrouter.com/de/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): Der Stundenschlüssel von BroadNet: eine Stunde Gemma 4 VL 26B auf einem Stream zum Festpreis. Inhalt, Limits, Zielgruppe und wann er sich lohnt. - [Datenspeicherung bei LLM-APIs: OpenAI, Anthropic, Gemini, Mistral](https://broadrouter.com/de/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): Was OpenAI, Anthropic, Gemini API, Mistral und BroadNet offiziell zur Speicherung von API-Prompts und zum Training damit sagen, zitiert nach Stand November 2026. - [LLM-API-Schlüssel absichern: Speicherung, Rotation, IP-Allowlists](https://broadrouter.com/de/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): LLM-API-Schlüssel aus Code und Logs fernhalten, ein Schlüssel pro Workload, planmäßig rotieren, an Server-IPs binden und ein Leck in Sekunden widerrufen. - [LLM-API-Preise erklärt: Input, gecachter Input, Output und Reasoning](https://broadrouter.com/de/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): So funktionieren LLM-API-Preise pro Token: Input, gecachter Input, Output, Reasoning-Tokens, Bilder und Audio pro Minute, mit Rechenbeispielen aus einer Preisliste. - [Sensible Daten an LLM-APIs: schwärzen, minimieren, weniger loggen](https://broadrouter.com/de/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): So senden Sie Kunden-, Gesundheits- oder Finanzdaten sicher an eine LLM-API: Felder einstufen, umkehrbar pseudonymisieren, Zero Retention nutzen, Logs sauber halten. - [LLM-API-Kosten kontrollieren: Budgets, Key-Limits und Loop-Schutz](https://broadrouter.com/de/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): So läuft Ihre LLM-API-Rechnung nicht aus dem Ruder: Prepaid-Guthaben, ein Schlüssel pro App, IP-Allowlists, Logging, Budgets im Code, Loop-Limits und Festpreise. - [LLM-API-Nutzung loggen: Kosten pro Nutzer und Feature, ohne PII](https://broadrouter.com/de/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): Jeden LLM-API-Aufruf mit Tokens, Kosten, Latenz und Request-ID loggen, Ausgaben Nutzern und Features zuordnen und PII vor dem Schreiben schwärzen. Mit Python-Code. - [LLM-Kosten pro Nutzer: Unit Economics für ein KI-Feature in SaaS](https://broadrouter.com/de/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): Berechnen Sie, was ein KI-Feature pro aktivem Nutzer und Monat kostet, warum Vielnutzer mehr zählen als Durchschnitte und wie Sie Tarife kostendeckend kalkulieren. - [LLM-API-Preise für Open-Weight-Modelle: 6 Anbieter im Vergleich](https://broadrouter.com/de/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): Wie OpenRouter, Together AI, Fireworks, Groq, DeepInfra und BroadNet offene Modelle abrechnen: pro Token, gecacht, Batch, GPU-Stunden und Flatrate, mit Quellen. - [LLM-Latenz: TTFT vs. Tokens pro Sekunde vs. End-to-End-Zeit](https://broadrouter.com/de/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): Was Time to First Token, Tokens pro Sekunde und End-to-End-Latenz messen, wie Sie jede im Client mit Code messen und welche Maßnahme welche Zahl verbessert. - [LLM-Streaming erklärt: SSE-Chunks, Deltas, Usage und Abbruch](https://broadrouter.com/de/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Wie Streaming von Chat Completions auf der Leitung funktioniert, wie Sie die Antwort aus Deltas zusammensetzen, Usage lesen und sauber abbrechen: Python, Node, curl. - [Long Context vs. RAG: Wann 1M Tokens die Vektordatenbank ersetzen](https://broadrouter.com/de/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): Wann ein Kontextfenster mit 1M Tokens eine Vektordatenbank ersetzt und wann RAG gewinnt: Vergleich nach Kosten pro Anfrage, Latenz, Recall und Wartung. - [Kontextmanagement für lang laufende Agenten mit 1M-Kontextfenster](https://broadrouter.com/de/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): So bleiben lange Agenten-Sessions mit 1M-Token-Fenster genau und bezahlbar: Append-only-Verlauf, cachefreundliche Prompts, Kompaktierung und was zuerst wegfällt. - [Lang laufende Agenten: Timeouts, Retries und Checkpoints, die halten](https://broadrouter.com/de/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): LLM-Agenten stundenlang am Laufen halten: Timeout pro Aufruf, Streaming gegen Hänger, sichere Retries, idempotente Tool-Aktionen, Checkpoint nach jedem Schritt. - [MCP-Server mit OpenAI-kompatiblen Modellen nutzen (jeder Anbieter)](https://broadrouter.com/de/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP funktioniert mit jedem LLM: Der Client macht aus MCP-Tools Function Calls. So nutzen Sie MCP mit offenen Modellen in Cline, Continue, opencode, n8n oder Python. - [Wann sich ein Kontextfenster mit 1M Tokens lohnt (und wann nicht)](https://broadrouter.com/de/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): Was in 1.048.576 Tokens passt, was es bei DeepSeek V4.1 Flash kostet, wie Reasoning und Caching die Rechnung ändern und wann RAG weiterhin die bessere Wahl ist. - [Was Multi-Agenten-Setups wirklich kosten: pro Token vs. Flatrate](https://broadrouter.com/de/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): Multi-Agenten-Systeme verbrauchen ein Vielfaches der Tokens eines Agenten. Die Session-Rechnung für Flatrate vs. pro Token, versteckte Faktoren und wann was gewinnt. - [Mehrsprachige Speech-to-Text API: 14 Sprachen, ein Endpoint](https://broadrouter.com/de/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): Englisch, Spanisch, Französisch, Chinesisch, Arabisch und 9 weitere Sprachen über einen OpenAI-kompatiblen Endpoint transkribieren, pro Minute abgerechnet. - [n8n AI Agent mit günstigem LLM: Setup, Modellwahl, Kosten pro Lauf](https://broadrouter.com/de/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): n8n-AI-Agent-Workflows auf einem günstigen OpenAI-kompatiblen Modell: welches Modell für Tools, welche Einstellungen Tokens sparen, Kosten pro 10.000 Läufe. - [Open-Weight- vs. geschlossene KI-Modelle: Kosten, Kontrolle, Risiko](https://broadrouter.com/de/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): Open-Weight- oder geschlossene LLMs fürs Unternehmen? Kosten, Kontrolle, Datenschutz, Qualitätsabstand und Lock-in im Vergleich, dazu drei Wege zum Betrieb. - [Wechsel zu einer OpenAI-kompatiblen API: die Migrations-Checkliste](https://broadrouter.com/de/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): base_url zu ändern ist der leichte Teil. Modell-IDs, Reasoning-Tokens, Usage im Stream, Fehler, Rate Limits und Schlüssel gehen kaputt. Checkliste mit Code. - [OpenRouter oder Flatrate für KI-Agenten: Was kostet weniger?](https://broadrouter.com/de/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter rechnet pro Token plus Gebühr auf Guthaben ab, eine Flatrate pro Session. Die Kostenrechnung für Dauer-Agenten, Gebühren, Limits und was wozu passt. - [Coding-Agenten sicher und unbeaufsichtigt über Nacht laufen lassen](https://broadrouter.com/de/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): Ein praktisches Setup für unbeaufsichtigte Läufe von Coding-Agenten: Sandbox, Git-Checkpoints, Task-Dateien, Schrittlimits, Review am Morgen und Flatrate-Sessions. - [Prompts so strukturieren, dass das Präfix gecacht wird](https://broadrouter.com/de/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): So bauen Sie LLM-Prompts auf, damit das gecachte Präfix identisch bleibt: System-Prompt, Tool-Schemas, Dokumente, nur ergänzter Verlauf, plus ein Test als Beweis. - [Reasoning-Tokens: Warum Sie mehr Output zahlen, als Sie sehen](https://broadrouter.com/de/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): Reasoning-Modelle rechnen verstecktes Denken als Output ab. Wie Sie es messen, was es pro Modell kostet und wie reasoning_effort, max_tokens und Routing es senken. - [Parallele Coding-Agenten mit einer LLM-API-Flatrate betreiben](https://broadrouter.com/de/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Sub-Agenten, Orchestratoren und Batch-Worker auf unbegrenztem DeepSeek V4.1 Flash: Zahl der Sessions, Verteilung auf Schlüssel, Überlappung begrenzen. - [LLM selbst hosten oder API: echte Kosten und Break-even](https://broadrouter.com/de/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): GPU-Mietpreise von offiziellen Seiten, die Formel für Kosten pro 1 Mio. Tokens auf eigener GPU, die nötige Auslastung gegen eine API und vergessene Betriebskosten. - [Speech-to-Text-API-Preise im Vergleich: Preis pro Minute 2026](https://broadrouter.com/de/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): Preise pro Minute und pro 1.000 Stunden für OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq und BroadNet Transcribe, plus Zusatzfunktionen. - [JSON Mode vs. json_schema: strukturierte Ausgabe, die hält](https://broadrouter.com/de/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON Mode liefert gültiges JSON, json_schema Ihre Struktur. So fordern Sie beides bei OpenAI-kompatiblen Modellen an, validieren mit Pydantic und reparieren Fehler. - [Tokens pro Wort, Arabisch vs. Englisch: Warum Arabisch mehr kostet](https://broadrouter.com/de/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): Warum Arabisch für denselben Inhalt oft mehr LLM-Tokens braucht als Englisch, warum Tokens pro Wort täuschen, und ein Skript, das Ihr Verhältnis aus usage misst. - [Tool Calling mit offenen Modellen: Schemas, parallele Aufrufe, Retries](https://broadrouter.com/de/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): Function Calling mit OpenAI-kompatiblen offenen Modellen: Schema-Tipps, parallele Tool Calls, fünf typische Fehler, eine Python-Schleife mit Validierung und Retry. - [Callcenter-Aufzeichnungen transkribieren: Pipeline, auch für Arabisch](https://broadrouter.com/de/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): Stereo-Anrufe nach Kanal trennen, Sprechbeiträge schneiden, Arabisch und Englisch transkribieren, PII schwärzen, per LLM zusammenfassen. Python-Pipeline mit Kosten. - [Offene Modelle in Coding-Agenten: Aider, Cline, opencode, Codex](https://broadrouter.com/de/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): So verbinden Sie Aider, Cline, Roo Code, opencode, Continue und Codex CLI mit einem OpenAI-kompatiblen Endpoint: die genauen Feldnamen und die Fallstricke. - [Sprachnachrichten in Text für Firmen: WhatsApp und Telegram zu Tickets](https://broadrouter.com/de/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): Sprachnachrichten aus WhatsApp Business und Telegram-Bots in Transkripte und Support-Tickets verwandeln: herunterladen, transkribieren, Felder per LLM extrahieren. - [LLM-Wochenplan für kurze Projekte: Orbit vs. Stunde vs. Monat](https://broadrouter.com/de/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): Wann eine Woche unbegrenzt DeepSeek V4.1 Flash besser ist als Stunden-Session, Monatsplan oder Token-Abrechnung: Rechnung für Hackathons, Sprints, Migrationen. - [Was „unbegrenzt“ bei einem LLM-API-Plan wirklich bedeutet](https://broadrouter.com/de/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Token-Zähler, 5-Stunden-Fenster, Wochenlimits, Rate Limits, Parallelität: Sieben Fragen zeigen, ob ein „unbegrenzter“ KI-Plan für Ihren Agenten unbegrenzt ist. - [Alternative zur Whisper API: wann wechseln und wie testen](https://broadrouter.com/de/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): Wann sich eine gehostete Alternative zur OpenAI Whisper API lohnt: Wortfehlerrate auf Ihrem Audio, Sprachen, Preis pro Stunde, Limits, Datenschutz, mit Testcode. - [LLM-API mit Zero Data Retention: Checkliste mit 10 Fragen](https://broadrouter.com/de/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): Was „Zero Data Retention“ bei einer LLM-API heißen sollte: Prompts, Outputs, Logs, Training, Unterauftragnehmer, Audits. Zehn Fragen, bevor Sie Daten senden. - [Guides in German](https://broadrouter.com/de/guides/) - [OpenAI Python SDK mit BroadNet Inference nutzen](https://broadrouter.com/de/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): Das offizielle OpenAI Python SDK mit einer geänderten base_url auf BroadNet Inference umstellen: installieren, Schlüssel setzen, erste Anfrage, Antwort streamen. - [OpenAI Node.js SDK mit BroadNet Inference nutzen](https://broadrouter.com/de/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): Das offizielle OpenAI SDK für Node.js und TypeScript per baseURL auf BroadNet Inference umstellen: installieren, erste Anfrage, Tokens streamen, Request-IDs loggen. - [BroadNet Inference im Browser testen, ohne Konto](https://broadrouter.com/de/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): Arabisches Speech-to-Text, Dokumentenerkennung und Gemma-4-Chat im Browser-Playground von BroadNet testen, ohne Anmeldung oder Karte, danach per API. - [Schnellstart mit curl: erste Anfrage an BroadNet Inference](https://broadrouter.com/de/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): Erste Aufrufe an BroadNet Inference mit curl: Chat Completion, gestreamte Antwort und Audio-Transkription, dazu Antwort, Token-Nutzung und Fehler lesen. - [Chat Completions streamen und abgebrochene Verbindungen behandeln](https://broadrouter.com/de/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): So funktioniert Streaming bei BroadNet Inference: Server-Sent Events, der Marker [DONE], Token-Nutzung im letzten Chunk und was tun, wenn ein Stream zu früh endet. - [Warum ein Reasoning-Modell eine leere Antwort liefert](https://broadrouter.com/de/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3 und Qwen3.8 27B denken vor der Antwort nach, und dieses Reasoning zählt zu max_tokens. So vermeiden Sie leere Antworten. - [Fehlercodes von BroadNet Inference: Wann ist ein Retry sinnvoll?](https://broadrouter.com/de/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): Was jeder HTTP-Status der BroadNet Inference API bedeutet, welche Fehler Sie mit Backoff wiederholen, welche Sie zuerst beheben und was Sie dem Support senden. - [API-Schlüssel per IP-Allowlist an Ihre Server binden](https://broadrouter.com/de/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): Schützen Sie Schlüssel für BroadNet Inference mit IP-Allowlists, trennen Sie Inferenz-Schlüssel von Portal-Credentials und sperren Sie Leaks in Sekunden. Mit curl. - [So funktionieren die unbegrenzten LLM-API-Pläne für Agenten](https://broadrouter.com/de/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): Die unbegrenzten Pläne für DeepSeek V4.1 Flash haben keinen Token-Zähler und kein Wochenlimit. So funktionieren Sessions, parallele Agenten, Schlüssel, Abrechnung. - [Aider mit unbegrenztem DeepSeek V4.1 Flash betreiben](https://broadrouter.com/de/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Verbinden Sie den Coding-Agenten Aider über seine OpenAI-kompatible Einstellung mit BroadNet Inference und nutzen Sie DeepSeek V4.1 Flash zum Pauschalpreis. - [LangChain mit BroadNet Inference nutzen](https://broadrouter.com/de/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): So verbinden Sie ChatOpenAI aus LangChain über base_url und api_key mit BroadNet Inference, und was LangChain von Reasoning-Modellen übernimmt und was nicht. - [Cline mit BroadNet Inference nutzen (Provider OpenAI Compatible)](https://broadrouter.com/de/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Richten Sie die VS-Code-Erweiterung Cline über den Provider OpenAI Compatible mit BroadNet Inference ein: Felder, Modell-IDs, Kontextgrößen und typische Fehler. - [Roo Code mit BroadNet Inference nutzen (OpenAI Compatible)](https://broadrouter.com/de/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code wurde im Mai 2026 eingestellt: Einstellungen für bestehende Installationen und Forks und wohin Sie wechseln (Cline oder Zoo Code), mit BroadNet Inference. - [Continue mit BroadNet Inference nutzen (config.yaml und apiBase)](https://broadrouter.com/de/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): Fügen Sie der IDE-Erweiterung Continue BroadNet-Modelle per provider openai und apiBase in der config.yaml hinzu, inklusive tool_use und Kontextlänge. - [BroadNet Inference als eigenen Provider in opencode einrichten](https://broadrouter.com/de/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): BroadNet Inference im Coding-Agent opencode als eigenen OpenAI-kompatiblen Provider in opencode.jsonc einrichten, mit Modell-Aliasen, Capabilities und Limits. - [Open WebUI mit BroadNet Inference verbinden](https://broadrouter.com/de/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): BroadNet Inference in Open WebUI als OpenAI-kompatible Verbindung hinzufügen, Modell-IDs in die Allowlist eintragen, Verify-, Leer- und 429-Fehler beheben. - [BroadNet Inference als Custom Endpoint in LibreChat einbinden](https://broadrouter.com/de/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): BroadNet Inference als OpenAI-kompatiblen Custom Endpoint in librechat.yaml einrichten: Schlüssel in .env, Modellliste, Titelmodell und Fehlerbehebung. - [BroadNet Inference in n8n nutzen (OpenAI-Credential, Base URL)](https://broadrouter.com/de/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): Richten Sie das OpenAI-Credential von n8n per Base URL auf BroadNet Inference und nutzen Sie den OpenAI Chat Model Node per ID, mit deaktivierter Responses API. - [Das Vercel AI SDK mit BroadNet Inference nutzen](https://broadrouter.com/de/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): BroadNet Inference aus dem Vercel AI SDK aufrufen, mit createOpenAICompatible aus @ai-sdk/openai-compatible: generateText, streamText, Tools und Reasoning Effort. - [Speech-to-Text API für Arabisch: Schritt für Schritt](https://broadrouter.com/de/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): Arabisches Audio mit BroadNet Transcribe Arabic über einen OpenAI-kompatiblen Endpoint transkribieren: Formate, Feld language, Python- und curl-Beispiele, Preis. - [LiteLLM an BroadNet Inference anbinden (SDK und Proxy)](https://broadrouter.com/de/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): BroadNet Inference über LiteLLM aufrufen, mit dem Modellpräfix openai/ und api_base, im Python SDK oder in der config.yaml des Proxys, mit passenden Rate Limits. - [OpenAI Agents SDK mit BroadNet Inference betreiben](https://broadrouter.com/de/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Das OpenAI Agents SDK für Python per AsyncOpenAI und OpenAIChatCompletionsModel mit BroadNet Inference nutzen und das Tracing mit OpenAI-Schlüsselpflicht abschalten. - [WhatsApp-Sprachnachrichten per API in Text umwandeln](https://broadrouter.com/de/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): WhatsApp-Sprachnachrichten (.ogg) oder Handyaufnahmen (.m4a) mit ffmpeg in WAV umwandeln und per API-Aufruf auf Arabisch oder in 13 weiteren Sprachen transkribieren. - [WhatsApp-Sprachnachrichten in n8n transkribieren](https://broadrouter.com/de/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): n8n-Workflow, der eine WhatsApp-Sprachnachricht empfängt, bei Meta herunterlädt, mit ffmpeg umwandelt und auf Arabisch oder in 13 weiteren Sprachen transkribiert. - [Arabische Rechnungen und Belege mit einem Vision-Modell auslesen](https://broadrouter.com/de/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Rechnungs- oder Belegfoto an ein Vision-Modell auf BroadNet Inference senden und Summe, Datum und Händler als JSON erhalten, gefragt auf Arabisch oder Englisch. - [Zed mit BroadNet Inference als OpenAI-kompatiblen Provider nutzen](https://broadrouter.com/de/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Binden Sie BroadNet Inference in settings.json als OpenAI-kompatiblen Provider in den Editor Zed ein: Modell-IDs, Token-Limits, Capabilities und Schlüsselvariable. - [Kilo Code mit BroadNet Inference nutzen (Provider OpenAI Compatible)](https://broadrouter.com/de/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Verbinden Sie Kilo Code als OpenAI-kompatiblen Custom Provider mit BroadNet Inference, im VS-Code-Dialog oder in kilo.json, mit Modell-IDs, Limits und Tool-Flags. - [BroadNet Inference als Custom Provider in goose einbinden](https://broadrouter.com/de/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): Verbinden Sie den KI-Agenten goose als OpenAI-kompatiblen Custom Provider mit BroadNet Inference – über goose Desktop, goose configure oder eine JSON-Datei. - [Pydantic AI mit BroadNet Inference nutzen (OpenAICompatibleProvider)](https://broadrouter.com/de/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Pydantic-AI-Agents mit OpenAIChatModel und OpenAICompatibleProvider auf BroadNet Inference betreiben: Tool Calls, strukturierte Ausgabe, Model Profiles. - [CrewAI mit BroadNet Inference als eigenem OpenAI-Endpoint nutzen](https://broadrouter.com/de/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): Verbinden Sie CrewAI-Agenten per LLM(custom_openai=True, base_url=...) mit BroadNet Inference und wählen Sie Modelle für Tools, langen Kontext und die Flatrate. - [LlamaIndex über OpenAILike mit BroadNet Inference nutzen](https://broadrouter.com/de/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): LlamaIndex über die Klasse OpenAILike mit BroadNet Inference verbinden und api_base, context_window, is_chat_model und is_function_calling_model richtig setzen. - [BroadNet Inference per OpenAI-API-compatible-Plugin in Dify einbinden](https://broadrouter.com/de/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): Verbinden Sie Dify über das Model-Provider-Plugin OpenAI-API-compatible mit BroadNet Inference: exakte Feldwerte, Kontextgrößen und Function-Calling-Optionen. - [Jan mit BroadNet Inference als OpenAI-kompatiblem Custom Provider](https://broadrouter.com/de/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): BroadNet Inference als OpenAI-kompatibler Custom Provider in der Desktop-App Jan: Modell-IDs von Hand eintragen, Tools und Vision pro Modell aktivieren. - [SillyTavern an BroadNet Inference anbinden (Custom OpenAI-compatible)](https://broadrouter.com/de/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): SillyTavern über die Chat-Completion-Quelle Custom (OpenAI-compatible) an BroadNet Inference anbinden: Modell-IDs, Prompt Post-Processing und Token-Limits. - [Spring AI mit BroadNet Inference nutzen (OpenAI base-url)](https://broadrouter.com/de/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Den OpenAI-Chat-Starter von Spring AI per spring.ai.openai.base-url, BroadNet-Modell-ID und max-tokens auf BroadNet Inference richten und über ChatClient aufrufen. - [BroadNet Inference aus Go mit dem openai-go SDK aufrufen](https://broadrouter.com/de/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): Das offizielle SDK openai-go v3 per option.WithBaseURL mit BroadNet Inference nutzen: Chat Completions, Streaming mit Usage und Fehlerbehandlung in Go. - [LangGraph-Agenten auf BroadNet Inference bauen](https://broadrouter.com/de/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): Einen LangGraph-StateGraph oder create_agent auf BroadNet Inference betreiben: ChatOpenAI, base_url und BroadNet-Modell-ID, dazu Tools, Streaming und Memory. - [Haystack-Pipelines und -Agenten mit BroadNet Inference nutzen](https://broadrouter.com/de/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Richten Sie Haystacks OpenAIChatGenerator per api_base_url, BroadNet-Modell-ID und Secret auf BroadNet Inference aus: für Pipelines, Tools, Streaming und Agenten. - [Semantic Kernel mit BroadNet Inference nutzen](https://broadrouter.com/de/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Microsoft Semantic Kernel mit BroadNet Inference verbinden: in Python per AsyncOpenAI-Client, in C# per eigenem Endpunkt, für Chat, Plugins und Agenten. - [Mastra-Agenten mit BroadNet Inference nutzen](https://broadrouter.com/de/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): Mastra-Agenten auf BroadNet Inference betreiben: ein model-Objekt mit id, url und apiKey – inklusive des doppelten Präfixes broadnet/, das die Modell-ID braucht. - [Strukturierte Ausgabe mit Instructor auf BroadNet Inference](https://broadrouter.com/de/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Extrahieren Sie mit Instructor Pydantic-Objekte aus BroadNet Inference – per from_openai oder from_provider mit eigener base_url – und wählen Sie den passenden Mode. - [DSPy mit BroadNet Inference nutzen (model_type="chat" setzen)](https://broadrouter.com/de/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): Konfigurieren Sie dspy.LM für BroadNet Inference mit Präfix openai/, api_base und model_type="chat", denn DSPy 3.5 schickt OpenAI-Modelle an die Responses API. - [BroadNet-Modelle mit promptfoo evaluieren](https://broadrouter.com/de/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): promptfoo-Evals und Modellvergleiche nebeneinander auf BroadNet Inference ausführen: mit dem Provider openai:chat, apiBaseUrl und apiKeyEnvar, dazu Grader. - [smolagents mit OpenAIModel auf BroadNet Inference betreiben](https://broadrouter.com/de/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Hugging Face smolagents per OpenAIModel und api_base an BroadNet Inference anbinden: CodeAgent auf jedem Modell, ToolCallingAgent auf GLM 5.3 und Qwen3.8 27B. - [Flowise mit BroadNet Inference verbinden (OpenAI Custom Model)](https://broadrouter.com/de/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Nutzen Sie BroadNet Inference in Flowise-Chatflows: Node OpenAI Custom Model, Base Path, OpenAI-API-Credential und ein eingetippter BroadNet-Modellname. - [Langflow mit BroadNet Inference verbinden](https://broadrouter.com/de/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): BroadNet Inference in Langflow nutzen: OpenAI-Komponente mit dem Feld OpenAI API Base und eingetipptem Modellnamen – und warum OpenAI Compatible es nicht listet. - [Glossary entries in German](https://broadrouter.com/de/glossary/) - [KI-Agent (AI Agent)](https://broadrouter.com/de/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): Ein KI-Agent ist ein System, in dem ein Sprachmodell seine eigenen Schritte auf ein Ziel hin steuert: Es entscheidet, welche Tools es aufruft, liest die Ergebnisse und wiederholt das, bis es eine endgültige Antwort liefert oder an ein Limit stößt. Ein festes Skript, das ein Modell an vorgegebenen Stellen aufruft, heißt dagegen meist Workflow. - [API-Schlüssel (API Key)](https://broadrouter.com/de/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): Ein API-Schlüssel ist eine geheime Zeichenfolge, die einen Client beim Aufruf einer API identifiziert und autorisiert. LLM-APIs erwarten ihn meist im Header Authorization: Bearer, und wer den Schlüssel besitzt, kann Anfragen stellen, die dem Konto des Inhabers berechnet werden. - [Arabische Diakritika (Tashkeel)](https://broadrouter.com/de/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): Arabische Diakritika, auch Tashkeel genannt, sind kleine Zeichen über oder unter den Buchstaben, die kurze Vokale, verdoppelte Konsonanten und das Fehlen eines Vokals anzeigen. Sie sind optional: Die meisten arabischen Texte, auch Nachrichten, lassen sie weg, und Leser wie Software erschließen die Vokale aus dem Kontext. - [Arabische Dialekterkennung (dialect identification)](https://broadrouter.com/de/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): Arabische Dialekterkennung ist die Aufgabe, festzustellen, in welcher arabischen Varietät ein Text oder eine Aufnahme vorliegt, etwa Ägyptisch, Golf-Arabisch, Levantinisch, Maghrebinisch oder modernes Hocharabisch, auf Ebene von Region, Land oder Stadt. Man nutzt sie, um Nachrichten weiterzuleiten, Prompts auszuwählen und Testsets aufzuteilen. - [Arabische Textnormalisierung (text normalization)](https://broadrouter.com/de/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): Arabische Textnormalisierung ist der Schritt, der Schreibvarianten desselben Worts vor dem Abgleich auf eine Form abbildet: Diakritika und Tatweel entfernen, Alef-Formen vereinheitlichen und oft ى auf ي sowie ة auf ه abbilden. Sie wird auf Suchschlüssel und Evaluierung angewendet, nicht auf angezeigten Text. - [Arabizi (Franco-Arabisch)](https://broadrouter.com/de/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): Arabizi ist informell in lateinischen Buchstaben geschriebenes Arabisch, bei dem Ziffern für arabische Laute ohne lateinische Entsprechung stehen, etwa 3 für ع. Es ist in Chats und sozialen Medien verbreitet, meist dialektal, hat keine feste Rechtschreibung und wird oft mit Englisch oder Französisch gemischt. - [AWQ (Activation-aware Weight Quantization)](https://broadrouter.com/de/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (Activation-aware Weight Quantization) ist ein Post-Training-Verfahren, das die Gewichte eines LLM auf geringe Präzision komprimiert, meist auf 4-Bit-Ganzzahlen. Es nutzt Aktivierungsstatistiken aus einem kleinen Kalibrierungsset, um die wichtigsten Gewichtskanäle zu finden, und skaliert sie so, dass sie weniger Genauigkeit verlieren. Backpropagation ist nicht nötig. - [Basis-URL (Base URL)](https://broadrouter.com/de/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): Eine Basis-URL ist die feste Stammadresse einer API, etwa https://api.example.com/v1, die ein Client vor jeden Endpoint-Pfad wie /chat/completions setzt. In OpenAI-kompatiblen SDKs lenkt eine geänderte Basis-URL, zusammen mit API-Schlüssel und Modell-ID, denselben Code zu einem anderen Anbieter. - [Batch API](https://broadrouter.com/de/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Eine Batch API ist ein asynchroner LLM-Endpoint, der viele Anfragen auf einmal annimmt, meist als Datei, sie innerhalb eines festgelegten Zeitfensters statt sofort verarbeitet und die Ergebnisse gesammelt zurückgibt. Anbieter wie OpenAI und Anthropic berechnen Batch-Anfragen mit Rabatt gegenüber ihren Standard-APIs. - [BM25](https://broadrouter.com/de/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) ist eine Rankingfunktion für die Stichwortsuche, die ein Dokument danach bewertet, wie oft die Suchbegriffe darin vorkommen, wie selten diese Begriffe in der gesamten Sammlung sind und wie lang das Dokument ist. Sie ist die Standard-Ähnlichkeit in Lucene und Elasticsearch. - [Byte Pair Encoding (BPE)](https://broadrouter.com/de/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Byte Pair Encoding (BPE) ist ein Algorithmus zur Subwort-Tokenisierung, der bei einzelnen Zeichen oder Bytes beginnt und wiederholt das häufigste benachbarte Paar zu einem neuen Vokabulareintrag zusammenführt, bis eine Zielgröße des Vokabulars erreicht ist. Die meisten aktuellen LLM-Tokenizer nutzen BPE oder eine Byte-Level-Variante. - [Gecachte Input-Tokens (Cached Input Tokens)](https://broadrouter.com/de/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): Gecachte Input-Tokens sind Prompt-Tokens, die ein LLM-Anbieter kürzlich verarbeitet hat und aus einem Cache wiederverwenden kann, statt sie neu zu berechnen, meist ein wiederholtes Präfix wie ein System-Prompt oder ein Dokument. Anbieter berechnen sie günstiger als neue Input-Tokens. - [Chain-of-Thought (CoT)](https://broadrouter.com/de/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Chain-of-Thought (CoT) ist eine Prompting-Technik, bei der einem Sprachmodell Zwischenschritte des Denkens vorgemacht werden oder es gebeten wird, diese vor der endgültigen Antwort auszuschreiben. Sie verbessert die Genauigkeit bei mehrstufigen Aufgaben wie Rechnen und Logik, und Reasoning-Modelle erzeugen sie inzwischen automatisch. - [Zeichenfehlerrate (Character Error Rate, CER)](https://broadrouter.com/de/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Die Zeichenfehlerrate (CER) misst die Genauigkeit von Transkription oder OCR auf Zeichenebene: die Ersetzungen, Löschungen und Einfügungen, die nötig sind, um die Ausgabe in den Referenztext zu verwandeln, geteilt durch die Zahl der Zeichen in der Referenz. Niedriger ist besser, 0 bedeutet volle Übereinstimmung. - [Chat Completions API](https://broadrouter.com/de/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Die Chat Completions API ist der HTTP-Endpoint von OpenAI, POST /v1/chat/completions, der einen Modellnamen und eine Liste von Nachrichten mit Rollen entgegennimmt und die nächste Nachricht des Modells zurückgibt. Ihr Format für Anfrage und Antwort wurde zum gemeinsamen Standard, den OpenAI-kompatible Anbieter und SDKs übernehmen. - [Chat-Template](https://broadrouter.com/de/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): Ein Chat-Template ist die Formatierungsregel, meist in Jinja geschrieben und mit dem Tokenizer eines Modells ausgeliefert, die eine Liste von Chat-Nachrichten mit Rollen in die eine Tokenfolge umwandelt, die das Modell erwartet, einschließlich seiner speziellen Steuer-Tokens. Ein falsches Template kann die Antworten eines Modells deutlich verschlechtern. - [Chunking (RAG)](https://broadrouter.com/de/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Chunking ist der Schritt in einer Retrieval-Pipeline, der Dokumente in kleinere Stücke, sogenannte Chunks, zerlegt, bevor sie eingebettet und indexiert werden. Die Chunk-Größe legt fest, wie viel Text jedes Stück enthält, und der Chunk-Overlap wiederholt etwas Text zwischen benachbarten Chunks, damit an einer Grenze zerschnittene Gedanken nicht verloren gehen. - [Code-Switching](https://broadrouter.com/de/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Code-Switching ist die Verwendung von zwei oder mehr Sprachen oder Varietäten einer Sprache im selben Gespräch oder Satz, mündlich oder schriftlich. Beispiele sind Arabischsprachige, die englische Wörter in ihren Dialekt mischen oder zwischen modernem Hocharabisch und ihrem Dialekt wechseln. - [KI-Coding-Agent (AI coding agent)](https://broadrouter.com/de/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): Ein KI-Coding-Agent ist ein Software-Tool, in dem ein Sprachmodell in einer Schleife an einer Codebasis arbeitet: Es liest Dateien, bearbeitet sie, führt Befehle wie Tests aus, liest die Ausgabe und entscheidet über den nächsten Schritt, bis die Aufgabe erledigt oder ein Limit erreicht ist. - [Concurrency (Parallelität)](https://broadrouter.com/de/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): Concurrency (Parallelität) ist die Zahl der Anfragen, die ein Client gleichzeitig offen hat. LLM-APIs begrenzen sie oft pro Schlüssel oder Plan, getrennt von den Anfragen pro Minute, weil jede offene Anfrage GPU-Kapazität belegt, bis ihre Antwort fertig ist. - [Context Compaction (Kontextverdichtung)](https://broadrouter.com/de/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): Context Compaction bedeutet, eine Konversation, die sich dem Limit des Kontextfensters nähert, zusammenzufassen und die Session mit dieser Zusammenfassung fortzusetzen, sodass ein Agent wichtige Entscheidungen und den Zustand behält und rohen Verlauf wie alte Tool-Ausgaben verwirft. - [Context Engineering](https://broadrouter.com/de/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): Context Engineering ist die Arbeit, die Menge an Tokens auszuwählen und zu pflegen, die ein Sprachmodell bei jeder Anfrage sieht, also System-Prompt, Tool-Definitionen, Verlauf, abgerufene Dokumente und Notizen, damit das Modell die kleinste Menge aussagekräftiger Informationen erhält, die es für die Aufgabe braucht. - [Context Rot](https://broadrouter.com/de/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): Context Rot ist der schleichende Verlust an Genauigkeit und Zuverlässigkeit, den Sprachmodelle zeigen, wenn ihr Input länger wird, selbst wenn die Aufgabe gleich bleibt und der Input weit innerhalb des Kontextfensters liegt. Bekannt wurde der Begriff durch einen technischen Bericht von Chroma aus dem Jahr 2025, der 18 Modelle testete. - [Kontextfenster (Context Window)](https://broadrouter.com/de/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): Das Kontextfenster ist die maximale Anzahl an Tokens, mit der ein Sprachmodell in einer einzelnen Anfrage arbeiten kann. Es umfasst Prompt, Gesprächsverlauf, Tool-Definitionen, Bilder und den erzeugten Output einschließlich Reasoning-Tokens. Was nicht hineinpasst, muss gekürzt, zusammengefasst oder auf mehrere Anfragen verteilt werden. - [Datenresidenz (Data Residency)](https://broadrouter.com/de/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): Datenresidenz ist eine Anforderung oder Zusage, dass Daten in einem bestimmten Land oder einer bestimmten Region gespeichert und verarbeitet werden. Bei einer LLM-API betrifft sie, wo Prompts bei der Inferenz verarbeitet werden und wo gespeicherte Inhalte liegen; Konto- und Abrechnungsdaten sind oft ausgenommen. - [Arabische Dialekte (Dialectal Arabic)](https://broadrouter.com/de/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): Dialektales Arabisch ist das gesprochene Alltagsarabisch einer Region oder Stadt, etwa Ägyptisch, Golf-Arabisch, Levantinisch oder Maghrebinisch. Dialekte unterscheiden sich vom modernen Hocharabisch in Lauten, Wörtern und Grammatik, haben keine Standardschreibung und werden online zunehmend geschrieben, was sie für Sprachmodelle schwieriger macht. - [Sprecherdiarisierung (Speaker Diarization)](https://broadrouter.com/de/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): Sprecherdiarisierung ist die Aufgabe, eine Audioaufnahme nach Sprechern in Segmente zu teilen, also die Frage „Wer hat wann gesprochen?“ zu beantworten. Jeder Sprachabschnitt erhält eine anonyme Sprecher-ID wie SPEAKER_00; kombiniert mit einem Transkript zeigt das, welche Person welche Zeile gesagt hat. - [Destillation (Knowledge Distillation)](https://broadrouter.com/de/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): Destillation (Knowledge Distillation, Wissensdestillation) ist ein Trainingsverfahren, bei dem ein kleineres Schülermodell lernt, die Outputs eines größeren Lehrermodells nachzubilden, etwa seine Wahrscheinlichkeitsverteilungen oder erzeugten Antworten, sodass der Schüler viel von der Qualität des Lehrers behält und im Betrieb weniger kostet. - [Auftragsverarbeitungsvertrag (AVV, DPA)](https://broadrouter.com/de/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): Ein Auftragsverarbeitungsvertrag (AVV, englisch Data Processing Agreement, DPA) ist der Vertrag zwischen Verantwortlichem und Auftragsverarbeiter, der regelt, wie dieser mit personenbezogenen Daten umgeht. Artikel 28 Absatz 3 DSGVO schreibt ihn vor und nennt seine Inhalte: dokumentierte Weisungen, Vertraulichkeit, Sicherheit, Regeln für Unterauftragsverarbeiter, Löschung oder Rückgabe der Daten und Audits. - [Embeddings](https://broadrouter.com/de/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): Embeddings sind Listen von Gleitkommazahlen (Vektoren), die ein Modell erzeugt, um einen Text, ein Bild oder einen anderen Input darzustellen. Inputs mit ähnlicher Bedeutung erhalten nah beieinanderliegende Vektoren; über den Abstand lassen sich Suchergebnisse ranken, Dokumente clustern oder verwandte Elemente finden. - [Eval (LLM-Evaluierung)](https://broadrouter.com/de/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Ein Eval (kurz für Evaluation) ist ein wiederholbarer Test eines Sprachmodells oder Prompts: eine feste Menge an Inputs, eine Beschreibung, wie ein guter Output aussieht, und ein Bewerter (Code, ein Mensch oder ein anderes Modell), der jeden Output benotet, damit sich Versionen vergleichen lassen. - [Exponentieller Backoff (Exponential Backoff)](https://broadrouter.com/de/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): Exponentieller Backoff ist eine Retry-Strategie, bei der ein Client nach jedem fehlgeschlagenen Versuch länger wartet, meist mit Verdopplung der Wartezeit bis zu einer Obergrenze, bevor er es erneut versucht. Zufälliger Jitter auf jeder Wartezeit verteilt die Retries vieler Clients, damit sie den Server nicht im selben Moment treffen. - [Few-Shot-Prompting](https://broadrouter.com/de/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): Few-Shot-Prompting ist eine Technik, bei der ein Prompt vor dem eigentlichen Input einige ausgearbeitete Input-Output-Beispiele enthält, aus denen ein Sprachmodell Aufgabe und Format ableitet. Die Gewichte des Modells ändern sich nicht; Zero-Shot kommt ohne Beispiele aus, One-Shot nutzt eines. - [Fine-Tuning (Feinabstimmung)](https://broadrouter.com/de/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): Fine-Tuning ist das Weitertrainieren eines vortrainierten Sprachmodells mit einem kleineren, aufgabenspezifischen Datensatz, etwa Beispielgesprächen, sodass sich seine Gewichte ändern und es diese Aufgabe oder diesen Stil zuverlässiger beherrscht, ohne lange Anweisungen oder Beispiele in jedem Prompt. - [finish_reason](https://broadrouter.com/de/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason ist ein Feld an jeder Choice einer Chat-Completion-Antwort, das angibt, warum das Modell die Generierung beendet hat. Im OpenAI-Format lauten die Werte stop, length, tool_calls, content_filter und das veraltete function_call. Die Messages API von Anthropic meldet dasselbe als stop_reason. - [LLM-Flatrate (Flat-rate LLM plan)](https://broadrouter.com/de/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): Eine LLM-Flatrate ist ein Plan, der für den Zugang zu einer Sprachmodell-API einen Festpreis pro Zeitraum, etwa pro Monat, berechnet statt pro Token. Solche Pläne begrenzen meist etwas anderes, etwa Parallelität, Geschwindigkeit, Modelle oder Nutzungsfenster. Die Limits zählen also so viel wie der Preis. - [FP8](https://broadrouter.com/de/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 ist ein 8-Bit-Gleitkommaformat für Deep Learning, definiert in zwei Kodierungen: E4M3 mit 4 Exponenten- und 3 Mantissenbits und E5M2 mit 5 Exponenten- und 2 Mantissenbits. Es halbiert den Speicherbedarf gegenüber 16-Bit-Formaten und läuft auf aktuellen GPUs nativ. - [Frequency Penalty und Presence Penalty](https://broadrouter.com/de/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): Frequency Penalty und Presence Penalty sind Sampling-Einstellungen, die den Score von Tokens senken, die schon im Output stehen. Die Presence Penalty greift einmal, sobald ein Token überhaupt vorkam; die Frequency Penalty wächst mit jeder Wiederholung. In der API von OpenAI reichen beide von -2.0 bis 2.0, Standardwert 0. - [DSGVO (GDPR)](https://broadrouter.com/de/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): Die DSGVO (Datenschutz-Grundverordnung, Verordnung (EU) 2016/679, englisch GDPR) ist das EU-Gesetz zur Verarbeitung personenbezogener Daten. Sie gilt seit dem 25. Mai 2018 für Organisationen mit Sitz in der EU und für manche außerhalb. Bei LLM-APIs regelt sie Auftragsverarbeitungsverträge, internationale Übermittlungen und Datenminimierung. - [GGUF](https://broadrouter.com/de/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF ist ein binäres Dateiformat, in dem Sprachmodelle für die Inferenz mit GGML-basierten Runtimes wie llama.cpp gespeichert werden. Eine GGUF-Datei enthält die Gewichte, oft quantisiert, plus die zum Laden nötigen Metadaten wie Tokenizer- und Architekturangaben, und lässt sich für schnelles Laden per Memory-Mapping einbinden. - [GPU-Stunde (GPU-hour)](https://broadrouter.com/de/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): Eine GPU-Stunde ist ein Grafikprozessor, eine Stunde lang genutzt. Cloud-Anbieter rechnen gemietete GPU-Instanzen pro Stunde oder Sekunde ab, und Forschende geben in GPU-Stunden an, wie viel Rechenleistung ein Job brauchte. Sie zahlen für die Zeit der Maschine, ob sie ausgelastet ist oder nicht. - [Grounding (Quellenbindung)](https://broadrouter.com/de/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): Grounding bedeutet, den Output eines Sprachmodells an bestimmte, überprüfbare Informationsquellen zu binden, etwa mitgelieferte Dokumente, Suchergebnisse oder Datenbankeinträge, und es anzuweisen, daraus zu antworten. Gegroundete Antworten lassen sich auf eine Quelle zurückführen und enthalten seltener erfundene Fakten. - [Guardrails (Leitplanken für LLMs)](https://broadrouter.com/de/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): Guardrails sind Prüfungen, die eine Anwendung rund um den Aufruf eines Sprachmodells ausführt, um es innerhalb festgelegter Grenzen zu halten. Input-Guardrails prüfen, was ins Modell geht, Output-Guardrails, was herauskommt, und Tool-Guardrails validieren Aktionen vor der Ausführung. Eine fehlgeschlagene Prüfung blockiert, ändert oder markiert die Anfrage. - [Halluzination (KI)](https://broadrouter.com/de/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): Eine Halluzination ist Output eines Sprachmodells, der plausibel klingt, aber falsch, erfunden oder nicht durch den gegebenen Input gedeckt ist, etwa eine erfundene Quelle, eine falsche Zahl oder eine nicht existierende API. Sie entsteht, weil das Modell wahrscheinlichen Text vorhersagt, statt Fakten zu prüfen. - [Hamza](https://broadrouter.com/de/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): Hamza (ء) ist der arabische Buchstabe für den Knacklaut (Glottisschlag). Er steht allein oder auf einem Träger: Alif (أ إ), Waw (ؤ) oder Ya (ئ); آ verbindet Alif mit Madda. Weil getippter Text ihn uneinheitlich schreibt, fassen Suchmaschinen und Sprachbewertung die Alif-Formen oft zu ا zusammen. - [Hybride Suche (Hybrid Search)](https://broadrouter.com/de/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): Hybride Suche ist ein Retrieval-Verfahren, das für dieselbe Anfrage eine Keyword-Suche wie BM25 und eine Vektorsuche über Embeddings ausführt und die beiden Ergebnislisten zu einer zusammenführt, oft mit Reciprocal Rank Fusion (RRF) oder einer gewichteten Summe normalisierter Scores. - [Idempotenzschlüssel (Idempotency Key)](https://broadrouter.com/de/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Ein Idempotenzschlüssel ist ein eindeutiger Wert, meist eine zufällige ID wie eine UUID, den ein Client an eine zustandsändernde Anfrage wie eine Zahlung oder einen Kauf anhängt. Kommt dieselbe Anfrage mit demselben Schlüssel erneut, liefert der Server das ursprüngliche Ergebnis, statt die Aktion zweimal auszuführen. - [Bild-Tokens (Image Tokens)](https://broadrouter.com/de/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): Bild-Tokens sind die Input-Tokens, mit denen ein Vision-Language-Modell ein Bild im Prompt darstellt. Anbieter berechnen sie aus der Bildgröße, meist indem sie das Bild nach dem Skalieren in Patches oder Kacheln fester Größe zerlegen, und rechnen sie wie Text-Input-Tokens ab. Größere Bilder kosten also mehr. - [Inferenz-Anbieter (Inference Provider)](https://broadrouter.com/de/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): Ein Inferenz-Anbieter ist ein Unternehmen, das Machine-Learning-Modelle, oft Open-Weight-LLMs, auf eigenen, geleasten oder gemieteten GPUs betreibt und den Zugang über eine API verkauft, meist pro Token, pro Anfrage oder pro Zeiteinheit. Es kümmert sich um Serving, Skalierung und Hardware. - [Input- und Output-Tokens](https://broadrouter.com/de/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): Input-Tokens sind die Tokens, die ein Sprachmodell in einer Anfrage liest, einschließlich System-Prompt, Gesprächsverlauf, Tool-Definitionen und Bildern. Output-Tokens sind die Tokens, die es als Antwort erzeugt, einschließlich Reasoning. LLM-APIs berechnen beide meist getrennt, und Output ist in der Regel teurer. - [INT4 (4-Bit-Quantisierung)](https://broadrouter.com/de/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 ist ein 4-Bit-Ganzzahlformat mit 16 darstellbaren Stufen. Bei der LLM-Inferenz meint es meist 4-Bit-Weight-only-Quantisierung (W4A16): Gewichte werden als INT4 mit Skalierungsfaktoren gespeichert und bei der Berechnung auf 16 Bit erweitert. Sie belegen etwa ein Viertel des Speichers von 16-Bit-Gewichten. - [IP-Allowlist (IP allow list)](https://broadrouter.com/de/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): Eine IP-Allowlist ist eine Sicherheitskontrolle, die nur Anfragen von gelisteten IP-Adressen oder CIDR-Bereichen annimmt und alle anderen ablehnt. Bei einer API begrenzt sie den Schaden eines geleakten Schlüssels, weil der Schlüssel aus nicht gelisteten Netzen nicht funktioniert. - [Jailbreak (LLM)](https://broadrouter.com/de/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): Ein Jailbreak ist eine Eingabe, die ein Sprachmodell dazu bringen soll, sein Sicherheitstraining oder die Regeln der einsetzenden Anwendung zu ignorieren, sodass es Inhalte erzeugt oder Aktionen ausführt, die es normalerweise verweigern würde. OWASP behandelt Jailbreaking als eine Form von Prompt Injection. - [JSON-Modus (JSON mode)](https://broadrouter.com/de/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): Der JSON-Modus ist eine Einstellung von LLM-APIs, in APIs im OpenAI-Stil response_format {"type": "json_object"}, die dafür sorgt, dass die Antwort des Modells gültiges, parsebares JSON ist. Anders als Structured Outputs mit einem JSON Schema garantiert er nicht, welche Keys, Typen oder Werte das Objekt enthält. - [JSON Schema](https://broadrouter.com/de/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema ist ein in JSON geschriebenes Vokabular, mit dem Sie die Struktur von JSON-Daten beschreiben und validieren: Typen, Pflichtfelder, erlaubte Werte und Verschachtelung. Die aktuelle Version ist 2020-12. LLM-APIs definieren damit Tool-Parameter und die genaue Struktur, der eine strukturierte Ausgabe folgen muss. - [KV-Cache](https://broadrouter.com/de/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): Ein KV-Cache ist der Speicher, in dem ein Transformer-Sprachmodell die Attention-Keys und -Values aller bereits verarbeiteten Tokens ablegt, damit jedes neue Token sie wiederverwendet, statt die ganze Sequenz neu zu berechnen. Er wächst mit der Kontextlänge und begrenzt oft, wie viele Anfragen eine GPU bedienen kann. - [Latenz (LLM)](https://broadrouter.com/de/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): Latenz ist bei einer Sprachmodell-API die Zeit, die eine Anfrage vom Senden bis zum Empfang des letzten Output-Tokens braucht. Sie ist die Summe aus Time to First Token und Generierungszeit und wird meist als Perzentile wie p50 und p95 angegeben statt als Durchschnitt. - [LLM-as-a-Judge (LLM als Bewerter)](https://broadrouter.com/de/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-Judge ist eine Evaluierungsmethode, bei der ein Sprachmodell die Ausgaben anderer Modelle bewertet: Es benotet eine einzelne Antwort anhand einer Rubrik oder wählt die bessere von zwei Antworten. Das nähert menschliches Urteil zu viel geringeren Kosten an, hat aber bekannte Verzerrungen, etwa eine Vorliebe für die erste oder längere Antwort. - [LLM-Benchmark](https://broadrouter.com/de/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): Ein LLM-Benchmark ist eine feste, meist öffentliche Sammlung von Aufgaben mit einer Bewertungsmethode, mit der Sprachmodelle bei einer Fähigkeit gemessen und verglichen werden, etwa Wissensfragen, Programmieren oder Spracherkennung. Ein Benchmark-Score zeigt die relative Leistung in diesem Test, nicht wie ein Modell bei Ihrem eigenen Workload abschneidet. - [LLM-Router](https://broadrouter.com/de/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): Ein LLM-Router ist eine Software oder ein Dienst zwischen einer Anwendung und einer oder mehreren Sprachmodell-APIs, der für jede Anfrage entscheidet, welches Modell oder welcher Anbieter sie bearbeitet, nach Regeln wie Kosten, Latenz, Verfügbarkeit oder Art der Aufgabe. - [Logprobs (Log-Wahrscheinlichkeiten)](https://broadrouter.com/de/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Logprobs sind die logarithmierten Wahrscheinlichkeiten, die ein Sprachmodell den von ihm erzeugten Tokens zugewiesen hat. Eine API, die sie zurückgibt, zeigt für jedes Output-Token, wie wahrscheinlich das Modell es fand, und auf Wunsch die wahrscheinlichsten alternativen Tokens an dieser Stelle. - [LoRA (Low-Rank Adaptation)](https://broadrouter.com/de/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (Low-Rank Adaptation) ist eine parametereffiziente Fine-Tuning-Methode, die die Gewichte eines vortrainierten Modells einfriert und für ausgewählte Schichten ein Paar kleiner Matrizen mit niedrigem Rang trainiert. Das gelernte Update lässt sich als kleiner Adapter speichern oder in die Gewichte einrechnen und verursacht dann keine zusätzliche Inferenzlatenz. - [Lost in the Middle](https://broadrouter.com/de/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the Middle bezeichnet die beobachtete Neigung von Sprachmodellen, relevante Informationen am besten zu nutzen, wenn sie nahe am Anfang oder Ende einer langen Eingabe stehen, und am schlechtesten, wenn sie in der Mitte liegen. Benannt wurde der Effekt 2023 in einer Studie von Liu et al. - [max_tokens](https://broadrouter.com/de/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens ist ein Parameter von LLM-APIs, der festlegt, wie viele Tokens ein Modell in einer Antwort höchstens erzeugen darf. Erreicht die Generierung das Limit, bricht sie ab, und die Antwort wird als abgeschnitten markiert, zum Beispiel mit finish_reason „length“. - [MCP (Model Context Protocol)](https://broadrouter.com/de/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol) ist ein offenes Protokoll auf Basis von JSON-RPC 2.0, das standardisiert, wie LLM-Anwendungen sich mit externen Tools und Daten verbinden. Ein MCP-Server stellt Tools, Ressourcen und Prompts bereit; ein MCP-Client in der Host-Anwendung findet sie und leitet Tool Calls des Modells an den Server weiter. - [Mixture of Experts (MoE)](https://broadrouter.com/de/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Mixture of Experts (MoE) ist eine Architektur neuronaler Netze, bei der eine Schicht viele Experten-Teilnetze enthält und ein gelernter Router jedes Token nur an wenige davon schickt. Ein MoE-Sprachmodell hat eine große Gesamtzahl an Parametern, nutzt pro Token aber nur eine kleine aktive Teilmenge. - [Modell-Deprecation (Abkündigung)](https://broadrouter.com/de/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): Modell-Deprecation bedeutet, dass ein API-Anbieter ankündigt, ein Modell nicht mehr anzubieten, meist mit einem Abschalttermin und einem empfohlenen Nachfolger. Nach diesem Termin schlagen Anfragen an das Modell fehl, Anwendungen müssen also vorher auf eine andere Modell-ID umsteigen. - [Modell-Fallback (Model Fallback)](https://broadrouter.com/de/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): Modell-Fallback ist ein Zuverlässigkeitsmuster, bei dem eine fehlgeschlagene LLM-Anfrage auf einem anderen Modell oder bei einem anderen Anbieter aus einer geordneten Liste wiederholt wird. Auslöser kann ein Fehler, ein Rate Limit, ein Kontextlängen-Fehler oder eine Ablehnung wegen Inhaltsrichtlinien sein. Es läuft in Ihrem Client oder in einem Gateway vor den APIs. - [Modelllizenz (Model License)](https://broadrouter.com/de/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): Eine Modelllizenz sind die rechtlichen Bedingungen, unter denen Gewichte, Code und Ausgaben eines Modells genutzt, verändert und weitergegeben werden dürfen. Open-Weight-Modelle nutzen Standard-Softwarelizenzen wie Apache 2.0 oder MIT oder eigene Modelllizenzen, die Bedingungen hinzufügen können, etwa Grenzen für große kommerzielle Dienste. - [Modernes Hocharabisch (MSA)](https://broadrouter.com/de/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): Modernes Hocharabisch (MSA) ist die standardisierte Form des Arabischen, die in der gesamten arabischen Welt für Schrift, Nachrichten, Bildung und formelle Rede genutzt wird. Niemand spricht es als Alltagsmuttersprache: Gesprochen werden regionale Dialekte wie Ägyptisch oder Golfarabisch, in formellen Situationen wechselt man ins MSA. - [Multi-Agenten-System (Multi-Agent System)](https://broadrouter.com/de/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): Ein Multi-Agenten-System ist eine LLM-Anwendung, in der mehrere Agenten, jeder mit eigenem Prompt, eigenen Tools und eigenem Kontext, an Teilen einer Aufgabe arbeiten. Ein verbreitetes Design ist Orchestrator-Worker: Ein Lead-Agent plant, startet Worker-Agenten parallel und führt ihre Ergebnisse zusammen. - [Multimodale KI (Multimodal AI)](https://broadrouter.com/de/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): Multimodale KI ist ein Modell oder System, das mit mehr als einer Art von Daten arbeitet, etwa Text, Bildern, Video und Audio. Ein multimodales Modell kann zum Beispiel ein Foto zusammen mit einer Frage lesen oder Sprache entgegennehmen und als Text antworten. Ein Vision-Language-Modell ist die häufigste Art. - [Named Entity Recognition (NER, Eigennamenerkennung)](https://broadrouter.com/de/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): Named Entity Recognition (NER) ist die Aufgabe, Namen in einem Text zu finden und jedem einen Typ zuzuordnen, etwa Person, Organisation oder Ort. Klassische Systeme taggen jedes Token; LLMs sollen stattdessen oft eine Liste von Entitäten als JSON zurückgeben, die dann gegen den Ausgangstext geprüft wird. - [Texterkennung (OCR)](https://broadrouter.com/de/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): Optische Zeichenerkennung (OCR) ist eine Technik, die gedruckten oder handgeschriebenen Text in einem Bild, Scan oder PDF erkennt und in maschinenlesbare Zeichen umwandelt. Klassische OCR liefert Text und seine Position auf der Seite; Vision-Language-Modelle können Text lesen und direkt Fragen dazu beantworten. - [Ogg Opus](https://broadrouter.com/de/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus ist Audio, das mit dem Opus-Codec (RFC 6716) codiert und in einem Ogg-Container (RFC 7845) gespeichert ist. Es ist für Sprache und interaktives Audio bei niedrigen Bitraten gebaut und das Format, das WhatsApp für Sprachnachrichten nutzt, meist als .ogg- oder .opus-Datei. - [Open-Weight-Modell](https://broadrouter.com/de/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): Ein Open-Weight-Modell ist ein Machine-Learning-Modell, dessen trainierte Gewichte veröffentlicht sind, sodass jeder sie herunterladen und ausführen kann, meist unter einer Lizenz mit Bedingungen. Es unterscheidet sich von vollständig quelloffener KI, die auch Trainingscode und detaillierte Angaben zu den Trainingsdaten freigibt, damit andere das Modell untersuchen und nachbauen können. - [OpenAI-kompatible API](https://broadrouter.com/de/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): Eine OpenAI-kompatible API ist eine HTTP-API, die dieselben Anfrage- und Antwortformate wie die OpenAI API akzeptiert, etwa POST /v1/chat/completions. OpenAI-SDKs und Tools können sie deshalb nutzen, wenn nur Basis-URL, API-Schlüssel und Modellname geändert werden. - [Parallele Tool Calls (Parallel Tool Calls)](https://broadrouter.com/de/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Parallele Tool Calls liegen vor, wenn ein Modell in einer einzigen Antwort mehr als einen Tool Call zurückgibt, zum Beispiel drei Dateilesevorgänge auf einmal. Im OpenAI-Format kommen sie als mehrere Einträge in tool_calls, und parallel_tool_calls auf false begrenzt das Modell auf null oder einen Aufruf pro Runde. - [Pay-as-you-go](https://broadrouter.com/de/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): Pay-as-you-go ist ein Preismodell, bei dem Sie nur für tatsächliche Nutzung zahlen, gemessen pro Einheit wie Tokens, Anfragen oder Audiominuten, ohne Abonnement oder Mindestabnahme. Die meisten LLM-APIs rechnen so ab, aus einem vorausbezahlten Guthaben oder per regelmäßiger Rechnung. - [PDPL (Datenschutzgesetze Saudi-Arabiens und der VAE)](https://broadrouter.com/de/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) ist die gängige Bezeichnung für das saudische Gesetz über personenbezogene Daten, erlassen per Königlichem Dekret M/19 und durchgesetzt von der SDAIA, und für das Bundesgesetzesdekret Nr. 45 von 2021 der VAE. Beide regeln die Verarbeitung personenbezogener Daten und die Bedingungen für Übermittlungen ins Ausland, auch von Prompts an ausländische LLM-APIs. - [PII (personenbezogene Daten)](https://broadrouter.com/de/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): PII (personally identifiable information) sind Informationen, mit denen sich die Identität einer Person unterscheiden oder zurückverfolgen lässt, etwa Name, Ausweisnummer oder biometrische Daten, plus weitere Informationen, die mit dieser Person verknüpft oder verknüpfbar sind. Der weitere Begriff der DSGVO lautet personenbezogene Daten. Bei LLMs entfernt PII-Schwärzung sie aus Prompts, bevor diese Ihre Systeme verlassen. - [Prefill und Decode](https://broadrouter.com/de/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill und Decode sind die beiden Phasen der LLM-Inferenz. Im Prefill verarbeitet das Modell den ganzen Prompt parallel und füllt seinen KV-Cache; im Decode erzeugt es die Antwort Token für Token und nutzt diesen Cache wieder. Prefill bestimmt vor allem die Time to First Token, Decode die Ausgabegeschwindigkeit. - [Prepaid-Guthaben (Prepaid Credit)](https://broadrouter.com/de/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): Prepaid-Guthaben ist Geld, das vor der Nutzung auf ein API-Konto eingezahlt wird. Jede Anfrage zieht ihre Kosten vom Guthaben ab, und wenn es aufgebraucht ist, stoppen abgerechnete Anfragen, bis Sie aufladen. So sind die Ausgaben auf das eingezahlte Geld begrenzt. - [Prompt Engineering](https://broadrouter.com/de/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): Prompt Engineering ist das Schreiben, Strukturieren und Testen der Anweisungen an ein Sprachmodell, damit es zuverlässig Output liefert, der Ihre Anforderungen erfüllt. Dazu gehören Rollennachrichten, klare Anweisungen, Beispiele, Formatierung und das Messen der Ergebnisse an Testfällen. - [Prompt Injection](https://broadrouter.com/de/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): Prompt Injection ist ein Angriff, bei dem Text, den ein Nutzer einem Sprachmodell liefert oder der in einer Webseite, E-Mail, Datei oder einem Tool-Ergebnis versteckt ist, das Verhalten des Modells gegen die Absicht des Entwicklers ändert. OWASP führt sie in seinen Top 10 für LLM-Anwendungen auf Platz eins (LLM01:2025). - [Quantisierung (Quantization)](https://broadrouter.com/de/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): Quantisierung bedeutet, die Zahlen eines neuronalen Netzes, vor allem seine Gewichte, in einem Format mit geringerer Präzision wie 8 oder 4 Bit statt 16 Bit zu speichern, damit das Modell weniger Speicher braucht und oft schneller läuft. Gute Verfahren halten den Genauigkeitsverlust klein. - [RAG (Retrieval-Augmented Generation)](https://broadrouter.com/de/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG (Retrieval-Augmented Generation) ist ein Muster, bei dem eine Anwendung zuerst zur Frage passende Abschnitte aus einem externen Speicher abruft, etwa einem Such- oder Vektorindex, und sie dann in den Prompt einfügt, damit ein Sprachmodell aus diesem Text antwortet statt nur aus seinem Gedächtnis. - [Rate Limit (Ratenbegrenzung)](https://broadrouter.com/de/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): Ein Rate Limit ist eine Obergrenze, die eine API festlegt, wie stark ein Client sie in einem Zeitfenster nutzen darf, etwa Anfragen pro Minute oder Tokens pro Minute. Anfragen über dem Limit werden abgelehnt, meist mit HTTP 429 Too Many Requests, bis die Nutzung wieder darunter liegt. - [ReAct (Reason and Act)](https://broadrouter.com/de/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (Reason and Act) ist ein Prompting-Muster, 2022 von Yao et al. vorgestellt, bei dem ein Sprachmodell Denkschritte mit Aktionen wie Suchen oder Tool Calls verschränkt. Es liest jede Beobachtung, passt seinen Plan an und wiederholt das, bis es eine endgültige Antwort geben kann. - [Reasoning-Tokens](https://broadrouter.com/de/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): Reasoning-Tokens sind Tokens, die ein Reasoning-Modell erzeugt, um ein Problem durchzuarbeiten, bevor es seine sichtbare Antwort schreibt. Sie werden meist als Output-Tokens berechnet, zählen zum Output-Limit und zum Kontextfenster und können von der API verborgen oder zusammengefasst werden. - [Request-Timeout (Zeitüberschreitung)](https://broadrouter.com/de/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): Ein Request-Timeout ist die längste Zeit, die ein HTTP-Client auf einen Server wartet, bevor er die Anfrage aufgibt. Clients teilen ihn meist in einen Connect-Timeout für den Verbindungsaufbau und einen Read-Timeout für das Warten auf Daten. LLM-Aufrufe brauchen großzügige Read-Timeouts, weil lange Antworten Zeit kosten. - [Reranking](https://broadrouter.com/de/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Reranking ist eine zweite Retrieval-Stufe, in der ein genaueres, aber langsameres Modell, meist ein Cross-Encoder, jedes Paar aus Anfrage und Kandidatendokument aus einer ersten Suche bewertet und neu ordnet, sodass nur die wenigen relevantesten an das Sprachmodell gehen. - [Rechts-nach-links-Text (RTL)](https://broadrouter.com/de/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): Rechts-nach-links-Text (RTL) ist Schrift wie Arabisch oder Hebräisch, die von rechts nach links angezeigt wird. Computer speichern sie in logischer Lesereihenfolge; der Unicode Bidirectional Algorithm entscheidet über die Anzeigereihenfolge, wenn RTL-Text mit Links-nach-rechts-Text, Ziffern, Modellnamen oder Code gemischt wird. - [Abtastrate (Sample Rate)](https://broadrouter.com/de/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): Die Abtastrate (Sample Rate) ist die Zahl der Audio-Samples, die pro Sekunde erfasst werden, gemessen in Hertz (Hz). Spracherkennungsmodelle erwarten meist 16.000 Hz (16 kHz) mono. Zusammen mit Bittiefe und Kanalzahl bestimmt die Abtastrate, wie groß eine unkomprimierte Audiodatei pro Minute ist. - [Seed-Parameter](https://broadrouter.com/de/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): Der Seed-Parameter ist eine Ganzzahl, die Sie mit einer Anfrage senden, damit die API versucht, bei einer Wiederholung derselben Anfrage dieselben Tokens zu ziehen. OpenAI beschreibt das als Best Effort: Determinismus ist nicht garantiert, und das Feld system_fingerprint zeigt, wann sich die Backend-Konfiguration geändert hat. - [Serverless Inference](https://broadrouter.com/de/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): Serverless Inference ist eine Art, Machine-Learning-Modelle zu betreiben, bei der die Plattform die Rechenleistung hinter einem Endpoint automatisch startet, skaliert und stoppt, im Leerlauf oft bis auf null, und nach Nutzung statt nach reservierten Maschinen abrechnet. Der Preis dafür sind mögliche Verzögerungen durch Cold Starts. - [SLA (Service Level Agreement)](https://broadrouter.com/de/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): Ein SLA (Service Level Agreement) ist eine vertragliche Zusage eines Anbieters zu einem messbaren Leistungsniveau, meist einem monatlichen Verfügbarkeitsprozentsatz, mit einer festgelegten Abhilfe, meist Service-Gutschriften, wenn es verfehlt wird. Ein SLA legt fest, was als Ausfall zählt, was ausgeschlossen ist und wie man Ansprüche geltend macht. - [Speculative Decoding (spekulatives Dekodieren)](https://broadrouter.com/de/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): Speculative Decoding ist eine Inferenztechnik, bei der ein kleines, schnelles Draft-Modell mehrere nächste Tokens vorschlägt und das große Zielmodell sie in einem einzigen Forward Pass prüft und die akzeptierten behält. Das beschleunigt die Generierung, ohne die Verteilung der Ausgabe zu verändern. - [Speech-to-Text (Spracherkennung)](https://broadrouter.com/de/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): Speech-to-Text, auch automatische Spracherkennung (ASR), ist eine Technik, die gesprochenes Audio in geschriebenen Text umwandelt. Eine Speech-to-Text API nimmt eine Audiodatei oder einen Stream entgegen und liefert ein Transkript, meist abgerechnet pro Audiominute. Die Genauigkeit wird mit der Wortfehlerrate gemessen und hängt von Akzent, Dialekt und Aufnahmequalität ab. - [Ausgabenlimit (Spend Limit)](https://broadrouter.com/de/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): Ein Ausgabenlimit (Spend Limit) ist eine Obergrenze dafür, wie viel ein API-Konto, Projekt oder Schlüssel in einem Zeitraum ausgeben darf, meist pro Monat. Ist sie erreicht, lehnt der Anbieter weitere Anfragen ab, oft mit HTTP 429 oder 400, bis der Zeitraum neu beginnt oder das Limit erhöht wird. - [SRT-Untertitel](https://broadrouter.com/de/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) ist ein Untertitelformat in Klartext. Jeder Untertitel besteht aus einer laufenden Nummer, einer Start- und Endzeit im Format hh:mm:ss,mmm --> hh:mm:ss,mmm, einer oder mehreren Textzeilen und einer Leerzeile. WebVTT ist das ähnliche Webformat, das HTML-Video verwendet. - [Server-Sent Events (SSE)](https://broadrouter.com/de/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Server-Sent Events (SSE) sind ein Webstandard, mit dem ein Server einem Client über eine einzige, lang offene HTTP-Antwort mit dem Content-Type text/event-stream einen Strom von Text-Events schickt. Jedes Event besteht aus einer oder mehreren Feldzeilen, etwa data:, und endet mit einer Leerzeile. - [Stoppsequenzen (Stop Sequences)](https://broadrouter.com/de/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): Stop Sequences (Stoppsequenzen) sind Zeichenfolgen, die Sie mit einer Anfrage senden und die das Modell stoppen, sobald es eine davon erzeugt. OpenAI nennt das Feld stop und akzeptiert bis zu 4 Sequenzen; Anthropic nennt es stop_sequences. Der gefundene Text fehlt normalerweise im zurückgegebenen Output. - [Streaming](https://broadrouter.com/de/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): Streaming ist eine Art, die Antwort eines Sprachmodells in kleinen Chunks zurückzugeben, während sie noch erzeugt wird, statt als eine Antwort am Ende. OpenAI-kompatible APIs streamen über Server-Sent Events, wenn die Anfrage stream auf true setzt. - [Strukturierte Ausgabe (Structured Output)](https://broadrouter.com/de/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): Strukturierte Ausgabe (Structured Output) ist eine Funktion von LLM-APIs, die die Antwort eines Modells auf ein vom Entwickler vorgegebenes Format beschränkt, meist ein JSON Schema, damit Code sie zuverlässig parsen kann. Sie ist strenger als der JSON-Modus, der gültiges JSON garantiert, aber keine bestimmte Struktur. - [Subagent](https://broadrouter.com/de/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): Ein Subagent ist ein separater Agent, den ein Hauptagent für eine klar umrissene Aufgabe startet. Er läuft mit eigenem Kontextfenster, System-Prompt und Tool-Zugriff, erledigt die Arbeit und gibt nur ein Ergebnis oder eine Zusammenfassung an den Hauptagenten zurück, sodass die Hauptkonversation klein bleibt. - [Unterauftragsverarbeiter (Subprocessor)](https://broadrouter.com/de/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): Ein Unterauftragsverarbeiter (Subprocessor) ist ein Dritter, den ein Auftragsverarbeiter mit der Verarbeitung personenbezogener Daten im Auftrag des Verantwortlichen betraut. Bei einer LLM-API können das GPU-Hosts, Partner-Inferenzanbieter und Backup-Kapazität sein. Nach Artikel 28 DSGVO brauchen sie die Genehmigung des Verantwortlichen und dieselben Datenschutzpflichten. - [System-Prompt](https://broadrouter.com/de/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): Ein System-Prompt ist der Anweisungstext, den eine Anwendung einem Sprachmodell vor dem Gespräch sendet, meist als Nachricht mit der Rolle system oder developer. Er legt Rolle, Regeln, Ton und Ausgabeformat des Modells fest, und das Modell soll ihm auch bei widersprechenden Nutzernachrichten folgen. - [Taa marbuta (ة)](https://broadrouter.com/de/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): Taa marbuta (ة, U+0629) ist ein arabischer Buchstabe, der nur am Wortende vorkommt und oft feminine Substantive und Adjektive markiert. Er sieht aus wie Heh (ه, U+0647) mit zwei Punkten, daher werden beide in getipptem Text verwechselt, und viele Such-Normalisierer bilden ة auf ه ab. - [Temperatur (Temperature)](https://broadrouter.com/de/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): Die Temperatur ist ein Sampling-Parameter, der steuert, wie zufällig der Output eines Sprachmodells ist. Das Modell teilt seine Token-Scores durch die Temperatur, bevor es sie in Wahrscheinlichkeiten umrechnet: Niedrige Werte lassen das wahrscheinlichste Token dominieren, hohe Werte glätten die Verteilung, sodass seltenere Tokens öfter gewählt werden. - [Text-to-Speech (TTS, Sprachsynthese)](https://broadrouter.com/de/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): Text-to-Speech (TTS), auch Sprachsynthese, wandelt geschriebenen Text in gesprochenes Audio um. Eine TTS-API nimmt Text, eine Stimme und ein Ausgabeformat wie MP3 oder WAV entgegen und liefert erzeugte Sprache. Es ist das Gegenstück zu Speech-to-Text, das Audio in Text umwandelt. - [Durchsatz (Throughput)](https://broadrouter.com/de/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): Durchsatz (Throughput) ist beim Betrieb von LLMs die gesamte Arbeit, die ein System pro Zeiteinheit über alle gleichzeitigen Anfragen erledigt, meist gemessen in Output-Tokens pro Sekunde oder Anfragen pro Sekunde. Er unterscheidet sich von der Geschwindigkeit pro Stream, also davon, wie schnell eine einzelne Anfrage ihre Tokens erhält. - [Time to First Token (TTFT)](https://broadrouter.com/de/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): Time to First Token (TTFT) ist die Zeit vom Senden einer Anfrage an ein Sprachmodell bis zum Eintreffen des ersten Output-Tokens. Sie umfasst Netzwerkzeit, Warteschlange auf dem Server und Prefill, den Schritt, in dem das Modell den ganzen Prompt liest; sie wächst daher mit Prompt-Länge und Last. - [Zeitstempel (Speech-to-Text)](https://broadrouter.com/de/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): Zeitstempel bei Speech-to-Text sind Start- und Endzeiten in Sekunden, die Teilen eines Transkripts zugeordnet sind. Segment-Zeitstempel gelten für eine Wendung oder einen Satz, Wort-Zeitstempel für jedes Wort. Man braucht sie für Untertitel, die Suche in Aufnahmen und die Zuordnung von Text zu Sprechern. - [Token](https://broadrouter.com/de/glossary/token/) (translation of https://broadrouter.com/glossary/token/): Ein Token ist die Texteinheit, die ein Sprachmodell liest und schreibt: ein Wort, ein Wortteil, ein Satzzeichen oder eine Bytefolge, die der Tokenizer des Modells erzeugt. LLM-APIs zählen, begrenzen und berechnen die Nutzung meist in Tokens. - [Tokenizer](https://broadrouter.com/de/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): Ein Tokenizer ist die Komponente, die Text in die Folge ganzzahliger Token-IDs umwandelt, die ein Sprachmodell liest, und erzeugte IDs wieder in Text zurückverwandelt. Jedes Modell bringt seinen eigenen Tokenizer mit, daher kann derselbe Text bei verschiedenen Modellen unterschiedlich viele Tokens ergeben. - [Tokens pro Sekunde (TPS)](https://broadrouter.com/de/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): Tokens pro Sekunde (TPS) sind die Rate, mit der ein Sprachmodell Output-Tokens erzeugt, meist gemessen für eine Anfrage nach Ankunft ihres ersten Tokens. Sie bestimmen, wie schnell eine Antwort streamt, sobald sie begonnen hat. Über alle gleichzeitigen Anfragen addiert ergibt sich der Systemdurchsatz; die Wartezeit bis zum ersten Token wird getrennt gemessen. - [Tool Calling (Function Calling)](https://broadrouter.com/de/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): Tool Calling, auch Function Calling genannt, ist eine Funktion von LLM-APIs, bei der das Modell eine strukturierte Anfrage zurückgibt, eine vom Entwickler definierte Funktion mit JSON-Argumenten auszuführen. Die Anwendung führt die Funktion aus und sendet das Ergebnis zurück; das Modell selbst führt nie Code aus. - [tool_choice](https://broadrouter.com/de/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice ist ein Anfrageparameter, der die Tool-Nutzung steuert. In den Chat Completions von OpenAI kann er auto (das Modell entscheidet), required (es muss ein oder mehrere Tools aufrufen), none (keine Tools) oder eine benannte, erzwungene Funktion sein. Die entsprechenden Optionen bei Anthropic sind auto, any, tool und none. - [Top-p (Nucleus Sampling)](https://broadrouter.com/de/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Top-p, auch Nucleus Sampling genannt, ist ein Decoding-Verfahren, bei dem ein Sprachmodell das nächste Token nur aus der kleinsten Menge der wahrscheinlichsten Tokens zieht, deren gemeinsame Wahrscheinlichkeit einen Schwellenwert p erreicht, etwa 0,9. Der unwahrscheinliche Rest wird verworfen. - [Unicode-Normalisierung](https://broadrouter.com/de/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Unicode-Normalisierung wandelt Text in eine Standardform um, damit gleich aussehende Strings dieselben Codepoints verwenden. Die vier Formen sind NFD, NFC, NFKD und NFKC; die K-Formen führen zusätzlich Kompatibilitätszeichen wie arabische Präsentationsformen und Ligaturen auf gewöhnliche Buchstaben zurück. - [Sprachaktivitätserkennung (Voice Activity Detection, VAD)](https://broadrouter.com/de/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): Voice Activity Detection (VAD, Sprachaktivitätserkennung) ist ein Schritt der Signalverarbeitung, der kurze Audio-Frames als Sprache oder Nicht-Sprache einstuft. Sprachsysteme nutzen sie, um Stille zu überspringen, lange Aufnahmen an Pausen zu teilen, das Ende eines Redebeitrags zu erkennen und Hintergrundgeräusche nicht zu transkribieren. - [Vektordatenbank (vector database)](https://broadrouter.com/de/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): Eine Vektordatenbank ist ein Datenspeicher für Embedding-Vektoren, oft zusammen mit Quelltext und Metadaten, der Nearest-Neighbour-Abfragen beantwortet: Zu einem Anfragevektor liefert sie die ähnlichsten gespeicherten Vektoren, meist nach Kosinus-Distanz, euklidischer Distanz oder Skalarprodukt. - [Vision-Language-Modell (VLM)](https://broadrouter.com/de/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): Ein Vision-Language-Modell (VLM) ist ein Modell, das Bilder und Text zusammen als Eingabe nimmt und Text erzeugt. So kann es Bilder beschreiben, Fragen zu ihnen beantworten und Dokumente, Diagramme oder Screenshots lesen. VLMs heißen auch Image-Text-to-Text- oder multimodale Modelle. - [vLLM](https://broadrouter.com/de/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM ist eine Open-Source-Bibliothek für LLM-Inferenz und -Serving, ursprünglich entwickelt im Sky Computing Lab der UC Berkeley. Sie verwaltet den Speicher des KV-Cache mit PagedAttention, bündelt eingehende Requests per Continuous Batching, unterstützt viele Quantisierungsformate und stellt einen OpenAI-kompatiblen HTTP-Server bereit. - [Vokabulargröße (vocabulary size)](https://broadrouter.com/de/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): Die Vokabulargröße ist die Zahl der unterschiedlichen Tokens im Tokenizer eines Sprachmodells, jedes mit eigener ID und eigenem Embedding. Ein größeres Vokabular kann gängigen Text mit weniger Tokens codieren, vergrößert aber die Embedding- und Ausgabeschichten und hilft nur bei Text, auf dem das Vokabular trainiert wurde. - [WAV](https://broadrouter.com/de/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format, auch WAVE) ist ein Audiocontainer auf Basis des RIFF-Formats von Microsoft und IBM. Er enthält meist unkomprimierte lineare PCM-Samples, daher ist seine Größe vorhersehbar: Abtastrate mal Bytes pro Sample mal Kanäle, pro Sekunde. - [Wortfehlerrate (Word Error Rate, WER)](https://broadrouter.com/de/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): Die Wortfehlerrate (WER) ist das Standardmaß für die Genauigkeit der Spracherkennung: die Zahl der Wortersetzungen, -löschungen und -einfügungen, die nötig sind, um ein Transkript in die korrekte Referenz zu verwandeln, geteilt durch die Zahl der Wörter in der Referenz. Niedriger ist besser. - [Zero Data Retention (ZDR)](https://broadrouter.com/de/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): Zero Data Retention (ZDR) ist eine Datenrichtlinie, bei der ein KI-Anbieter Prompts und Outputs nur zur Bearbeitung einer Anfrage verarbeitet und diese Inhalte danach nicht speichert, weder in Logs noch in Caches oder Trainingsdaten. Inhaltsfreie Daten wie Token-Zahlen und Zeiten werden meist weiterhin aufbewahrt. ### Portuguese (pt, Português): 294 translated pages - [Articles in Portuguese](https://broadrouter.com/pt/blog/) - [Preços de API de LLM em outubro de 2026: 3 aumentos, nenhum corte](https://broadrouter.com/pt/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 mudanças em páginas de preços de 5 de 7 provedores de API de LLM em outubro de 2026: 3 aumentos, nenhum corte, 9 modelos com preço novo e 11 removidos. - [Custo de uma execução de agente de código: passos, contexto e cache](https://broadrouter.com/pt/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): Por que um agente de código custa mais por passo quanto mais tempo roda, como o cache muda o total e quando uma sessão fixa vence a cobrança por token, com contas. - [Diacríticos no árabe com um LLM: prompt, checagens e DER](https://broadrouter.com/pt/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): Restaure os diacríticos do árabe (tashkeel) com uma API de LLM: um prompt que preserva as letras, uma checagem que rejeita edições, DER e WER, e quando usar o CATT. - [Correção gramatical e ortográfica do árabe com um LLM: passo a passo](https://broadrouter.com/pt/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Corrija ortografia e gramática do árabe com uma API de LLM: prompt de edição mínima, proteção contra reescritas, edições por tipo de erro e pontuação no estilo M2. - [Atas de reuniões em árabe a partir de gravações: um pipeline de API](https://broadrouter.com/pt/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Transforme gravações de reuniões em árabe em atas com decisões e ações: transcreva, limpe, resuma com um LLM em JSON e confira nomes e datas. - [Extrair entidades nomeadas em árabe com LLM: JSON, datas e checagens](https://broadrouter.com/pt/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Extraia nomes, organizações, lugares e datas hégira de textos em árabe para JSON com um LLM: prompt, checagem sem diacríticos, conversão de datas e avaliação. - [OCR para árabe por API: Google, Azure, Textract, Mistral e LLMs](https://broadrouter.com/pt/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): Quais APIs de OCR suportam oficialmente árabe, impresso e manuscrito, o que cada uma retorna e quanto custa a cada 1.000 páginas, comparadas a LLMs com visão. - [Legendas em árabe (SRT) com uma API de fala para texto: o que precisa](https://broadrouter.com/pt/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): Do que um SRT precisa, quais APIs de fala para texto devolvem timestamps ou SRT para árabe e um pipeline testado que gera legendas com uma API só de texto. - [Classificação de texto em árabe com LLM: sentimento, intenção, triagem](https://broadrouter.com/pt/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): Classifique avaliações, mensagens e tickets em árabe com um LLM: rótulos, prompt atento ao dialeto, saída em JSON, teste com macro F1 e custo em volume. - [Normalizar texto em árabe para busca e RAG: o que mudar e o que não](https://broadrouter.com/pt/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Tashkeel, alef, ta marbuta, hamza, letras persas, dígitos e Unicode: quais regras de normalização do árabe ajudam a busca, quais atrapalham, com Python testado. - [Resumo de texto em árabe com LLM: prompts, tamanho e avaliação](https://broadrouter.com/pt/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): Resuma artigos, relatórios e mensagens em dialeto árabe com uma API de LLM: prompts, controle de palavras, checagem de números e por que o ROUGE padrão dá 0. - [Alternativas ao Artificial Analysis: onde comparar LLMs em 2026](https://broadrouter.com/pt/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Alternativas ao Artificial Analysis: Arena, Epoch AI, Vals AI, rankings do OpenRouter, Hugging Face e dados de preço, pelo que cada uma mede e quando usá-la. - [Hardware de GPU e serving da BroadNet, segundo a própria BroadNet](https://broadrouter.com/pt/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): O que a BroadNet diz sobre o hardware da BroadNet Inference: três camadas de GPU, das B300 aos appliances no cliente, throughput por GPU e o que aprendeu nos testes. - [Cache de respostas de LLM para cortar custo: um cache exato em Python](https://broadrouter.com/pt/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Um cache exato e testado para respostas de API de LLM em Python: o que vai na chave, o que nunca guardar, TTLs, privacidade e como medir a taxa de acerto antes. - [Claude Code com modelos abertos: proxies, endpoints nativos e limites](https://broadrouter.com/pt/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): O Claude Code fala a Messages API da Anthropic. Como rodá-lo com modelos abertos por um proxy LiteLLM ou um endpoint nativo, e o que quebra no caminho. - [Revisão humana no seu fluxo de IA para documentos](https://broadrouter.com/pt/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): Revisão humana em um pipeline de IA para documentos: regras de validação, o que mandar a uma pessoa, amostras aleatórias, trilha de auditoria e taxa de revisão. - [Lei de proteção de dados do Egito (Lei 151 de 2020) e APIs de LLM](https://broadrouter.com/pt/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Dados de egípcios em API de LLM ou speech-to-text no exterior: o que dizem a Lei 151/2020 e a Decisão 816/2025 sobre transferências, licenças e incidentes. - [Antes do fine-tuning: defina o problema a resolver](https://broadrouter.com/pt/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): Quando prompt, RAG ou avaliação resolvem o problema sem fine-tuning, de quais dados de treino você precisa e como testar um fine-tune antes de pedir um. - [Como corrigir JSON inválido de um LLM: causas, reparo e retry](https://broadrouter.com/pt/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): Por que LLMs retornam JSON quebrado (code fences, texto extra, corte em max_tokens, vírgulas finais) e uma função Python testada que só repara o que é seguro. - [Com que frequência o preço de API de LLM muda? 3 anos de arquivo](https://broadrouter.com/pt/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): Lemos 308 páginas de preços arquivadas de 11 provedores de API de LLM, 2023 a 2026. 80% dos modelos nunca mudaram de preço; a DeepInfra fez 75% das mudanças. - [Como reduzir o custo de API de LLM: checklist de 10 passos, em ordem](https://broadrouter.com/pt/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): Reduza a conta de API de LLM na ordem certa: meça, ponha o prefixo em cache, enxugue prompts, limite saída e raciocínio, ajuste o modelo e use batch ou preço fixo. - [Failover de API de LLM: passe para um segundo provedor no seu código](https://broadrouter.com/pt/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): Como fazer failover entre provedores de LLM compatíveis com a OpenAI: erros que acionam, circuit breaker, mapeamento de IDs, streaming, custo e privacidade. - [Cortes de preço de API de LLM 2023–2026: os maiores, com fontes](https://broadrouter.com/pt/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): Os maiores cortes de preço de API de LLM de 2023 a 2026, com links do arquivo antes e depois: o3, Codestral, Gemini 1.5, e os aumentos que quase ninguém notou. - [Preço de API de LLM só cai? 52 aumentos e como se planejar](https://broadrouter.com/pt/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): Preços de API de LLM sobem. Achamos 52 aumentos de preço de tabela no mesmo modelo em páginas arquivadas, 2024–2026, mediana de 40%. Fontes e regra de orçamento. - [Rate limits de API de LLM por provedor: RPM, TPM e concorrência](https://broadrouter.com/pt/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): Rate limits de entrada de 13 provedores de API de LLM, lidos na documentação: OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks e outros. E o que são RPM e TPM. - [SLAs de APIs de LLM: o que Azure, Bedrock, Google e OpenAI prometem](https://broadrouter.com/pt/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): Com o que cada provedor de API de LLM se compromete por escrito: porcentagem de uptime, o que conta como indisponibilidade, créditos de serviço e metas de latência. - [Processamento em lote com LLM sem Batch API: um runner retomável](https://broadrouter.com/pt/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Rode de 10.000 a 1.000.000 de prompts numa API de chat completions sem Batch API: limite de concorrência, ritmo, novas tentativas, checkpoints retomáveis e prazos. - [O Safety Map: como 39 LLMs lidam com dados sensíveis](https://broadrouter.com/pt/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): A BroadNet avaliou 39 LLMs de 14 laboratórios em credenciais, PII, saúde e finanças: o que se transferiu, o que não e os vazamentos em tool calls e no raciocínio. - [Como testar a segurança de um LLM com dados sensíveis: método BroadNet](https://broadrouter.com/pt/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): O método publicado pela BroadNet para testar, antes do deploy, se um LLM vaza credenciais, PII, dados de saúde ou financeiros: categorias, leak e miss, 3+ execuções. - [Como migrar da API da OpenAI: o código em Python, Node e cURL](https://broadrouter.com/pt/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): As mudanças de código para sair da OpenAI para uma API compatível: cliente, de Responses para Chat Completions, ferramentas, streaming, embeddings e um smoke test. - [Os modelos que a BroadNet avaliou e descartou, e por quê](https://broadrouter.com/pt/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): A BroadNet publica os modelos que testou e rejeitou, com os motivos: modelos de chat reprovados no teste de credenciais; voz, OCR e embeddings que perderam. - [Como projetar uma experiência de IA em vários idiomas](https://broadrouter.com/pt/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): Projete um produto de IA para usuários de árabe, inglês e espanhol: locale da interface vs. idioma do modelo, RTL, texto misto, números, datas, detecção e custo. - [Custo de API: modelos abertos vs fechados, segundo os preços](https://broadrouter.com/pt/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): Modelos open-weight saem mais baratos por API que GPT, Claude ou Gemini? Preços por token das páginas dos provedores, variação entre hosts e custo por usuário. - [Histórico de preços das APIs da OpenAI, Anthropic e Google, 2023–2026](https://broadrouter.com/pt/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): Preços das APIs do GPT, Claude e Gemini lado a lado em páginas arquivadas, 2023 a 2026: que linhas baratearam, quais encareceram, cada uma com link do arquivo. - [Base URL de APIs compatíveis com a OpenAI: 14 provedores conferidos](https://broadrouter.com/pt/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): A base_url de 14 APIs de LLM compatíveis com a OpenAI, lida na documentação de cada provedor, com os endpoints listados e as diferenças que quebram a troca direta. - [OpenClaw e Hermes Agent com um modelo barato: setup e custos](https://broadrouter.com/pt/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): Conecte o OpenClaw ou o Hermes Agent a um endpoint compatível com a OpenAI, entenda por que as verificações agendadas queimam tokens e como reduzir a conta. - [Um endpoint privado começa com uma carga de trabalho clara](https://broadrouter.com/pt/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): O que definir antes de pedir um endpoint privado de LLM: carga de trabalho, tokens por segundo, concorrência, contexto, região, licença e orçamento, e como testá-lo. - [Prompt injection em agentes de código: defesas que funcionam](https://broadrouter.com/pt/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): Por onde a prompt injection chega a agentes de código e as defesas em camadas que fornecedores e pesquisas documentam: sandbox, saída de rede, allow-list, aprovação. - [Tamanho de chunk e overlap no RAG: efeito na recuperação e no custo](https://broadrouter.com/pt/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): Como tamanho de chunk e overlap mudam o recall do RAG, o tamanho do índice, o custo de embeddings e os tokens enviados ao LLM por consulta, com resultados e contas. - [Como definir max_tokens em modelos de raciocínio: um método](https://broadrouter.com/pt/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Por que modelos de raciocínio devolvem respostas vazias com max_tokens baixo e como dimensionar o limite por percentis, janela de contexto e retentativas crescentes. - [Gravar áudio no navegador e enviar para uma API de speech-to-text](https://broadrouter.com/pt/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): Grave o microfone com MediaRecorder, converta WebM ou o MP4 do Safari em WAV de 16 kHz e envie para a API de transcrição por um pequeno servidor que guarda a chave. - [Reduzir tokens do prompt: cortes medidos em sistema, tools e dados](https://broadrouter.com/pt/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): Economia de tokens medida com tokenizers reais ao enxugar prompt de sistema, few-shot, schemas de ferramentas, formatos de dados, histórico e diacríticos árabes. - [Mesmo modelo, host diferente: quanto varia o preço de API de LLM](https://broadrouter.com/pt/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): Em 119 modelos vendidos por vários hosts, o preço de tabela mais caro é, na mediana, 1,76 vez o mais barato. Dados das páginas de 41 provedores, novembro de 2026. - [PDPL saudita e dos EAU e APIs de LLM: checklist de transferência](https://broadrouter.com/pt/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Enviar dados pessoais a uma API de LLM da Arábia Saudita ou dos EAU: regras de transferência da PDPL, o que perguntar e quando a inferência local importa. - [Texto, voz ou visão? Comece pela tarefa](https://broadrouter.com/pt/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): Escolha entre chat, entrada de imagem e transcrição de áudio pela tarefa, com os modelos e limites que a BroadNet documenta e o que ela não oferece, como TTS. - [Transcrever áudio longo acima de 25 MB: dividir, enviar e juntar](https://broadrouter.com/pt/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): Como transcrever gravações acima do limite de uma API de fala para texto: reduza o arquivo, divida nos silêncios com ffmpeg, envie em paralelo e junte o texto. - [Traduzir strings de app para árabe com LLM: placeholders e plurais](https://broadrouter.com/pt/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): Traduza arquivos JSON de i18n para árabe com uma API de LLM: congele placeholders, gere as seis formas de plural, valide o ICU na CI e isole os nomes inseridos. - [Traduzir áudio em árabe para texto em inglês com API: pipeline testado](https://broadrouter.com/pt/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): Transforme gravações em árabe em texto em inglês com duas chamadas de API: transcreva com language=ar e traduza com um LLM, checando números, custo e qualidade. - [Live, sob consulta, pesquisa: como ler um catálogo de modelos](https://broadrouter.com/pt/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): Como saber quais modelos de um catálogo você pode chamar hoje, quais exigem consulta e quais só foram testados, usando o catálogo da BroadNet e /v1/models. - [Uno no vLLM: o que o estudo de serving da BroadNet mediu](https://broadrouter.com/pt/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): A BroadNet integrou ao vLLM o adapter de self-drafting Uno e o mediu em uma H100: 2,58x o throughput com um stream, 1,74x com 32. Método, DOI e limites. - [Uno no Gemma 4 26B: o estudo da BroadNet com mixture of experts](https://broadrouter.com/pt/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): A BroadNet treinou o primeiro adapter Uno de self-drafting para um modelo mixture of experts, o Gemma 4 26B A4B. O que mediu, quando compensa e o DOI. - [Orçamento de tokens de agentes: para onde vão os tokens de um loop](https://broadrouter.com/pt/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): System prompt, schemas de ferramentas, leituras de arquivos, histórico e raciocínio: como um agente gasta tokens, como medir cada parte e o que cortar primeiro. - [Word error rate (WER) no ASR em árabe: o que é e como calcular](https://broadrouter.com/pt/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): O que é word error rate, como calcular e por que o WER em árabe é tão sensível a diacríticos, normalização da grafia e dialetos. Com Python e números reais. - [Dialetos árabes e speech-to-text: Golfo, Egito, Levante, Magreb](https://broadrouter.com/pt/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Por que o speech-to-text tem dificuldade com o árabe do Golfo, egípcio, levantino e magrebino, e como montar um conjunto de teste por dialeto com seu áudio real. - [Tradução árabe–inglês com LLMs: prompts, glossários e avaliação](https://broadrouter.com/pt/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): Traduza do árabe para o inglês e vice-versa com uma API de LLM: prompts que mantêm o tom, glossários de termos fixos, números e nomes, e como avaliar a qualidade. - [Árabe do Golfo, egípcio e levantino: como os LLMs lidam com dialetos](https://broadrouter.com/pt/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): Como os LLMs lidam com o árabe do Golfo, egípcio e levantino: onde erram, como escrever o prompt, como testar e o que a BroadNet informa sobre o Qwen3.8 27B. - [OCR em árabe ou LLM de visão: lendo documentos e notas fiscais](https://broadrouter.com/pt/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): Quando usar OCR clássico e quando um LLM de visão para notas fiscais, recibos e formulários em árabe: armadilhas, saída em JSON, custo por página e validação. - [Engenharia de prompt em árabe: registro, RTL, números e texto misto](https://broadrouter.com/pt/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): Prompts para tirar um bom árabe de um LLM: registro e dialeto, números e datas, formatação RTL limpa e a mistura de árabe e inglês sob controle. - [RAG em árabe: normalização, chunking e retrieval que funcionam](https://broadrouter.com/pt/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): RAG sobre documentos em árabe: normalize alef, ya e ta marbuta, divida em chunks com critério, escolha embeddings multilíngues e adicione busca por palavra-chave. - [API de fala para texto em árabe: como escolher em 2026](https://broadrouter.com/pt/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): Como escolher uma API de transcrição de árabe: WER, dialetos, limites de arquivo, preço por hora de áudio e privacidade, e um plano de teste com as suas gravações. - [Arquitetura de chatbot de WhatsApp em árabe: de áudios a respostas](https://broadrouter.com/pt/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): Uma arquitetura que funciona para um assistente de WhatsApp em árabe: webhook, fila, transcrição de áudios, resposta do LLM, regras de política e custos, com Python. - [Arabizi e LLMs: como lidar com o franco-árabe e números como 3 e 7](https://broadrouter.com/pt/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): O Arabizi escreve árabe com letras latinas e dígitos como 3 para ع e 7 para ح. Como detectá-lo, convertê-lo para o alfabeto árabe com um LLM e responder bem. - [Melhor LLM para o Aider: como escolher o modelo e conectar](https://broadrouter.com/pt/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Como escolher o melhor LLM para o Aider: disciplina no formato de edição, contexto e custo por hora, um teste de uma hora no seu repo e o OPENAI_API_BASE. - [Melhor modelo aberto para agentes de código: ferramentas e custo](https://broadrouter.com/pt/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B ou Gemma 4 no seu agente de código? Escolha por tool calling, contexto, raciocínio e custo por tarefa, com as contas. - [Tokens de entrada em cache: como reduzir a conta de API de LLM](https://broadrouter.com/pt/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): A entrada em cache custa uma fração da entrada normal. Como o prompt caching é cobrado, como ler cached_tokens em usage e como estruturar os prompts. - [API de LLM com visão mais barata: custo por imagem e por página](https://broadrouter.com/pt/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): Quanto custa um LLM com visão por imagem e por página: a conta de tokens para Gemma 4, Qwen3.8 27B e DeepSeek V4.1 Flash, e onde a DeepInfra sai mais barata. - [Como escolher um provedor de API de LLM: checklist de 20 pontos](https://broadrouter.com/pt/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): Vinte perguntas antes de escolher uma API de LLM: preço, limites, compatibilidade OpenAI, privacidade, modelos, suporte e uptime, com pontuação e plano de teste. - [Como escolher uma API de LLM para árabe: o que testar](https://broadrouter.com/pt/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): Como escolher um LLM para apps em árabe: dialetos, qualidade de escrita, tradução, documentos e imagens, custo, e um conjunto de teste que você monta em um dia. - [Cline vs Roo Code vs Aider para equipes com modelos abertos (2026)](https://broadrouter.com/pt/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (agora arquivado, continuado como Zoo Code) e Aider comparados no que importa com modelos abertos: ferramentas, contexto, MCP, git e aprovações. - [Como contar tokens antes de enviar uma requisição a um LLM](https://broadrouter.com/pt/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): Estime os tokens do prompt por idioma e modelo antes de chamar uma API de LLM: por que tokenizers diferem, exemplos em cinco idiomas e margens para max_tokens. - [10 formas de reduzir o custo de LLM de agentes de código, com economia](https://broadrouter.com/pt/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Dez formas concretas de baixar o gasto em tokens do Cline, Aider, opencode ou do seu próprio agente de código, cada uma com a conta da economia e como medir. - [Alternativas à API da DeepSeek: 5 formas de usar os modelos DeepSeek](https://broadrouter.com/pt/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): Modelos DeepSeek pela API oficial, um agregador, o catálogo de uma nuvem, GPUs próprias ou a BroadNet Inference: prós e contras de preço, privacidade e limites. - [Preço da API do DeepSeek V4.1 Flash: por token, em cache e planos](https://broadrouter.com/pt/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): Preço da API do DeepSeek V4.1 Flash por 1M de tokens (entrada, cache, saída), planos fixos por mês, semana e hora e comparação com DeepSeek, DeepInfra e Fireworks. - [Como estimar o custo de API de LLM antes de lançar uma funcionalidade](https://broadrouter.com/pt/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): Modelo de custo passo a passo para uma nova funcionalidade de IA: usuários × ações × chamadas × tokens, com cache, raciocínio, retries, rate limits e exemplo mensal. - [Como avaliar um LLM antes de trocar de provedor](https://broadrouter.com/pt/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): Monte um pequeno eval set com tráfego real, avalie com checagens em código e um LLM juiz sem viés, e compare qualidade, custo e latência antes de migrar. - [Extrair dados de faturas, recibos e formulários com LLM de visão](https://broadrouter.com/pt/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): Pipeline de produção para faturas, recibos e formulários: preparo de imagens, prompt com esquema JSON, validação Pydantic, checagens de soma, retentativas e revisão. - [API de LLM com preço fixo ou por token: a conta do ponto de equilíbrio](https://broadrouter.com/pt/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): Quando um plano de LLM de preço fixo vence o pagamento por token? Quatro contas mensais para o DeepSeek V4.1 Flash, o volume de equilíbrio e como medir o seu. - [Crédito grátis de API de LLM: quanto o crédito inicial rende](https://broadrouter.com/pt/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): O que um crédito grátis de API de LLM compra de verdade: tokens, minutos de áudio e chaves por hora com o crédito inicial da BroadNet, vs. Gemini e OpenRouter. - [RGPD (GDPR) e APIs de LLM: 15 perguntas para o seu provedor](https://broadrouter.com/pt/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): Checklist do RGPD (GDPR) para APIs de LLM: DPA, suboperadores, retenção, treinamento, transferências internacionais e incidentes, com o artigo de cada pergunta. - [GLM 5.3 vs DeepSeek V4.1 Flash: qual usar e quando](https://broadrouter.com/pt/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 vs DeepSeek V4.1 Flash em preço, contexto, ferramentas, saída estruturada, imagens e raciocínio, com custos de exemplo e uma regra simples para escolher. - [Erro 429 em API de LLM: backoff, jitter e limitadores](https://broadrouter.com/pt/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 vs 5xx, backoff exponencial com full jitter, limitadores de concorrência e filas para APIs de LLM, com código Python e Node pronto para produção. - [Chave de API de LLM por hora: preço fixo por hora, explicado](https://broadrouter.com/pt/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): A chave por hora da BroadNet compra uma hora de Gemma 4 VL 26B num único stream a preço fixo. O que inclui, os limites, para quem serve e quando compensa. - [Retenção de dados em APIs de LLM: OpenAI, Anthropic, Gemini e Mistral](https://broadrouter.com/pt/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): O que OpenAI, Anthropic, Gemini API, Mistral e BroadNet dizem oficialmente sobre guardar prompts da API e treinar com eles, citado das páginas em novembro de 2026. - [Segurança de chaves de API de LLM: armazenamento, rotação e IPs](https://broadrouter.com/pt/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): Chaves de API de LLM fora do código e dos logs, uma chave por carga de trabalho, rotação periódica, só os IPs dos servidores e revogação de vazamento em segundos. - [Preço de API de LLM explicado: entrada, cache, saída e raciocínio](https://broadrouter.com/pt/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): Como funciona o preço por token de APIs de LLM: entrada, entrada em cache, saída, tokens de raciocínio, imagens e áudio por minuto, com exemplos de uma tabela real. - [Dados sensíveis numa API de LLM: mascare, minimize, registre menos](https://broadrouter.com/pt/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): Como enviar dados de clientes, de saúde ou financeiros a uma API de LLM com segurança: classifique campos, use placeholders reversíveis, retenção zero e logs limpos. - [Controle de gastos com API de LLM: orçamentos, limites e loops](https://broadrouter.com/pt/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): Evite que a conta da API de LLM dispare: saldo pré-pago, uma chave por app, IPs permitidos, registro de uso, orçamento no código, limite de loops e chaves fixas. - [Log de uso de API de LLM: custo por usuário e recurso, sem PII](https://broadrouter.com/pt/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): Registre cada chamada de API de LLM com tokens, custo, latência e request ID, atribua o gasto a usuários e recursos e mascare PII antes de gravar. Código em Python. - [Custo de LLM por usuário: unit economics de um recurso de IA no SaaS](https://broadrouter.com/pt/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): Quanto um recurso de IA custa por usuário ativo por mês, por que usuários intensos importam mais que a média e como precificar planos SaaS para cobrir o modelo. - [Preço de modelos open-weight: 6 provedores de inferência comparados](https://broadrouter.com/pt/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): Como OpenRouter, Together AI, Fireworks, Groq, DeepInfra e BroadNet cobram por modelos abertos: por token, cache, batch, horas de GPU e planos fixos, com fontes. - [Latência de LLM: TTFT, tokens por segundo e tempo de ponta a ponta](https://broadrouter.com/pt/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): O que medem o time to first token, os tokens por segundo e a latência de ponta a ponta, como cronometrar cada um no cliente e qual ajuste move qual número. - [Streaming de LLM explicado: chunks SSE, deltas, usage e cancelamento](https://broadrouter.com/pt/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Como o streaming de chat completions funciona na conexão, como montar a resposta a partir dos deltas, ler o usage do último chunk e cancelar em Python, Node e curl. - [Contexto longo vs RAG: quando 1M de tokens substitui um banco vetorial](https://broadrouter.com/pt/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): Quando uma janela de contexto de 1M de tokens pode substituir um banco vetorial e quando o RAG ainda vence: custo por consulta, latência, recall e manutenção. - [Gestão de contexto para agentes de longa duração com janela de 1M](https://broadrouter.com/pt/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): Como manter sessões longas de agentes precisas e baratas com uma janela de 1M de tokens: histórico append-only, cache, quando compactar e o que cortar primeiro. - [Agentes de longa duração: timeouts, retentativas e checkpoints](https://broadrouter.com/pt/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): Mantenha um agente LLM vivo por horas: timeout por chamada, streaming para detectar travamentos, retentativas seguras, efeitos colaterais idempotentes e checkpoints. - [Servidores MCP com modelos compatíveis com OpenAI (qualquer provedor)](https://broadrouter.com/pt/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP funciona com qualquer LLM: o cliente transforma ferramentas MCP em function calls. Como usar MCP com modelos abertos no Cline, Continue, opencode, n8n ou Python. - [Quando vale a pena uma janela de contexto de 1M de tokens (ou não)](https://broadrouter.com/pt/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): O que cabe em 1.048.576 tokens, quanto custa no DeepSeek V4.1 Flash, como raciocínio e cache mudam a conta e quando o RAG ainda é a melhor escolha. - [Quanto custam de verdade os sistemas multiagente: token ou sessão fixa](https://broadrouter.com/pt/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): Sistemas multiagente gastam várias vezes os tokens de um agente só. A conta de sessões fixas vs cobrança por token, os multiplicadores ocultos e quem vence. - [API de fala para texto multilíngue: 14 idiomas, um endpoint](https://broadrouter.com/pt/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): Transcreva inglês, espanhol, francês, chinês, árabe e mais 9 idiomas em um endpoint compatível com a OpenAI, cobrado por minuto, e saiba quando usar o modelo árabe. - [AI Agent do n8n num LLM barato: setup, modelo e custo por execução](https://broadrouter.com/pt/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): Workflows do AI Agent do n8n num modelo barato compatível com a OpenAI: qual modelo usar com ferramentas, ajustes que cortam tokens e a conta por 10.000 execuções. - [IA open-weight vs fechada na empresa: custo, controle e risco](https://broadrouter.com/pt/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): LLMs open-weight ou fechados na empresa? Custo, controle, privacidade, qualidade e lock-in, e as três formas de rodar modelos abertos e quando cada uma vence. - [Migrar para uma API compatível com a OpenAI: checklist de migração](https://broadrouter.com/pt/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): Trocar o base_url é a parte fácil. O que quebra são IDs de modelo, tokens de raciocínio, usage no streaming, erros, rate limits e chaves. Checklist com código. - [OpenRouter ou preço fixo para agentes de IA: qual sai mais barato?](https://broadrouter.com/pt/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): O OpenRouter cobra por token mais uma taxa sobre créditos; um plano de preço fixo cobra por sessão. A conta para agentes sempre ligados, taxas, limites e a escolha. - [Como rodar agentes de código à noite sem supervisão, com segurança](https://broadrouter.com/pt/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): Configuração prática para agentes de código sem supervisão: sandbox, checkpoints no git, arquivo de tarefa, limite de passos, revisão de manhã e plano fixo. - [Como estruturar prompts para o prefixo entrar em cache](https://broadrouter.com/pt/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): Como organizar prompts de LLM para o prefixo em cache ficar idêntico: prompt de sistema, schemas de ferramentas, documentos, histórico só com acréscimos e um teste. - [Tokens de raciocínio: por que você paga mais saída do que vê](https://broadrouter.com/pt/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): Modelos de raciocínio cobram o pensamento oculto como saída. Como medir, quanto custa em cada modelo e como reduzir com reasoning_effort, max_tokens e roteamento. - [Como rodar agentes de código em paralelo numa API de LLM de preço fixo](https://broadrouter.com/pt/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Subagentes, orquestradores e workers em lote no DeepSeek V4.1 Flash ilimitado: quantas sessões você precisa, como dividir entre chaves e como limitar a concorrência. - [Hospedar um LLM próprio ou usar API: custo real e ponto de equilíbrio](https://broadrouter.com/pt/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): Preços de aluguel de GPU das páginas oficiais, a fórmula do custo por 1M de tokens na sua GPU, a utilização para vencer uma API e os custos de operação esquecidos. - [Preço de API speech-to-text comparado: preço por minuto em 2026](https://broadrouter.com/pt/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): Preços por minuto e por 1.000 horas de OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq e BroadNet Transcribe, das páginas oficiais, mais adicionais. - [JSON mode vs json_schema: saída estruturada que não quebra](https://broadrouter.com/pt/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): O JSON mode garante JSON válido; o json_schema garante o seu formato. Como pedir cada um em modelos compatíveis com a OpenAI, validar com Pydantic e corrigir. - [Tokens por palavra em árabe vs inglês: por que o árabe custa mais](https://broadrouter.com/pt/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): Por que o árabe costuma usar mais tokens de LLM que o inglês para o mesmo sentido, por que tokens por palavra enganam, e um script para medir a sua proporção real. - [Tool calling com modelos abertos: schemas, chamadas paralelas, retry](https://broadrouter.com/pt/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): Function calling em modelos abertos compatíveis com a OpenAI: schemas, chamadas paralelas, as cinco formas de falha e um loop em Python que valida e tenta de novo. - [Como transcrever gravações de call center: pipeline pronto para árabe](https://broadrouter.com/pt/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): Separe chamadas estéreo por canal, corte os turnos, transcreva árabe e inglês, oculte dados pessoais e resuma cada chamada com um LLM. Código Python e custos. - [Modelos abertos em agentes de código: Aider, Cline, opencode, Codex](https://broadrouter.com/pt/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Conecte Aider, Cline, Roo Code, opencode, Continue e Codex CLI a qualquer endpoint compatível com a OpenAI: nomes exatos dos campos e as configurações que quebram. - [Transcrever áudios para empresas: do WhatsApp e Telegram a tickets](https://broadrouter.com/pt/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): Transforme áudios de clientes do WhatsApp Business e de bots do Telegram em transcrições e tickets de suporte: baixar, transcrever, extrair campos com LLM e rotear. - [Plano semanal de LLM para projetos curtos: Orbit, hora ou mês](https://broadrouter.com/pt/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): Quando uma semana ilimitada de DeepSeek V4.1 Flash vale mais que a sessão por hora, o plano mensal ou o pagamento por token: contas para hackathons e migrações. - [O que "ilimitado" significa de verdade em um plano de API de LLM](https://broadrouter.com/pt/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Medidores de tokens, janelas móveis de 5 horas, limites semanais, rate limits e concorrência: sete perguntas para ver se um plano "ilimitado" serve ao seu agente. - [Alternativa à API do Whisper: quando trocar e como testar](https://broadrouter.com/pt/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): Quando vale uma alternativa hospedada à API Whisper da OpenAI: taxa de erro de palavras no seu áudio, idiomas, preço por hora, limites e privacidade, com código. - [API de LLM com retenção zero de dados: checklist de 10 perguntas](https://broadrouter.com/pt/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): O que "retenção zero de dados" deveria significar numa API de LLM: prompts, saídas, logs, treinamento, subprocessadores e auditorias. 10 perguntas para fazer antes. - [Guides in Portuguese](https://broadrouter.com/pt/guides/) - [Como usar o SDK Python da OpenAI com a BroadNet Inference](https://broadrouter.com/pt/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): Aponte o SDK Python oficial da OpenAI para a BroadNet Inference mudando só a base_url. Instale, defina a chave, faça a primeira requisição e use streaming. - [Como usar o SDK Node.js da OpenAI com a BroadNet Inference](https://broadrouter.com/pt/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): Aponte o SDK oficial da OpenAI para Node.js e TypeScript para a BroadNet Inference com baseURL. Instale, faça a primeira requisição, use streaming e registre IDs. - [Teste a BroadNet Inference no navegador, sem conta](https://broadrouter.com/pt/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): Teste speech-to-text em árabe, leitura de documentos e chat com o Gemma 4 no playground da BroadNet, sem cadastro nem cartão, e depois leve a requisição para a API. - [Primeira requisição à BroadNet Inference com curl](https://broadrouter.com/pt/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): Primeiras chamadas à BroadNet Inference com curl: um chat completion, uma resposta em streaming e uma transcrição de áudio, e como ler resposta, usage e erros. - [Streaming de chat completions e conexões que caem](https://broadrouter.com/pt/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): Como o streaming funciona na BroadNet Inference: Server-Sent Events, o marcador [DONE], o uso de tokens no último chunk e o que fazer se o stream acabar antes. - [Por que um modelo de raciocínio devolve uma resposta vazia](https://broadrouter.com/pt/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3 e Qwen3.8 27B raciocinam antes de responder, e o raciocínio conta no max_tokens. Veja como parar de receber respostas vazias. - [Códigos de erro da BroadNet Inference e quando tentar de novo](https://broadrouter.com/pt/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): O que cada status HTTP da API da BroadNet Inference significa, quais repetir com backoff, quais corrigir antes e o que enviar ao suporte com o ID da requisição. - [Restrinja chaves de API aos seus servidores com lista de IPs](https://broadrouter.com/pt/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): Proteja as chaves da BroadNet Inference com listas de IPs permitidos, separe chaves de inferência de credenciais do portal e revogue uma chave vazada em segundos. - [Como funcionam os planos de API de LLM ilimitada para agentes](https://broadrouter.com/pt/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): Os planos ilimitados do DeepSeek V4.1 Flash não têm medidor de tokens nem limite semanal. Veja como funcionam sessões, agentes em paralelo, chaves e cobrança. - [Rode o Aider no DeepSeek V4.1 Flash ilimitado](https://broadrouter.com/pt/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Conecte o agente de programação Aider à BroadNet Inference pela configuração compatível com a OpenAI e use um plano ilimitado de preço fixo do DeepSeek V4.1 Flash. - [Como usar o LangChain com a BroadNet Inference](https://broadrouter.com/pt/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): Conecte o ChatOpenAI do LangChain à BroadNet Inference com base_url e api_key, e veja o que o LangChain mantém e o que descarta dos modelos de raciocínio. - [Use o Cline com a BroadNet Inference (provedor OpenAI Compatible)](https://broadrouter.com/pt/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Configure a extensão Cline do VS Code com a BroadNet Inference pelo provedor OpenAI Compatible: campos exatos, IDs de modelo, contexto e soluções comuns. - [Usar o Roo Code com a BroadNet Inference (OpenAI Compatible)](https://broadrouter.com/pt/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): O Roo Code foi encerrado em maio de 2026. Configurações para instalações e forks existentes, e para onde migrar (Cline ou o fork Zoo Code) com a BroadNet Inference. - [Use o Continue com a BroadNet Inference (config.yaml e apiBase)](https://broadrouter.com/pt/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): Adicione os modelos da BroadNet Inference à extensão Continue com provider openai e apiBase no config.yaml, incluindo a capability tool_use e o tamanho do contexto. - [Adicionar a BroadNet Inference ao opencode como provedor personalizado](https://broadrouter.com/pt/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): Adicione a BroadNet Inference ao agente de código opencode como provedor personalizado compatível com a OpenAI no opencode.jsonc, com aliases, capacidades e limites. - [Como conectar o Open WebUI à BroadNet Inference](https://broadrouter.com/pt/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Adicione a BroadNet Inference ao Open WebUI como conexão compatível com a OpenAI, coloque os IDs na allowlist e resolva verificação, respostas vazias e 429. - [Adicionar a BroadNet Inference ao LibreChat como custom endpoint](https://broadrouter.com/pt/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): Adicione a BroadNet Inference ao LibreChat como custom endpoint compatível com a OpenAI no librechat.yaml, com a chave no .env, lista de modelos e title model. - [Como usar a BroadNet Inference no n8n (credencial OpenAI, Base URL)](https://broadrouter.com/pt/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): Aponte a credencial OpenAI do n8n para a BroadNet Inference no campo Base URL e use o nó OpenAI Chat Model por ID, com Use Responses API desligado. - [Como usar o Vercel AI SDK com a BroadNet Inference](https://broadrouter.com/pt/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): Chame a BroadNet Inference pelo Vercel AI SDK com createOpenAICompatible do @ai-sdk/openai-compatible: generateText, streamText, ferramentas e reasoning effort. - [API de transcrição de áudio em árabe, passo a passo](https://broadrouter.com/pt/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): Transcreva áudio em árabe com o BroadNet Transcribe Arabic por um endpoint compatível com a OpenAI. Formatos, o campo language, exemplos em Python e curl, e o preço. - [Como rotear o LiteLLM para a BroadNet Inference (SDK e proxy)](https://broadrouter.com/pt/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): Chame a BroadNet Inference pelo LiteLLM com o prefixo openai/ e api_base, no SDK Python ou no config.yaml do proxy, com rate limits compatíveis com a sua chave. - [Rodar o OpenAI Agents SDK na BroadNet Inference](https://broadrouter.com/pt/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Use o OpenAI Agents SDK para Python com a BroadNet Inference via AsyncOpenAI e OpenAIChatCompletionsModel, e desative o tracing que exige uma chave da OpenAI. - [Transcrever áudios do WhatsApp para texto com uma API](https://broadrouter.com/pt/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): Converta áudios .ogg do WhatsApp ou gravações .m4a do celular para WAV com ffmpeg e transcreva em árabe ou em outros 13 idiomas com uma única chamada de API. - [Transcrever áudios do WhatsApp no n8n](https://broadrouter.com/pt/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): Um workflow do n8n que recebe um áudio do WhatsApp, baixa o arquivo da Meta, converte com ffmpeg e transcreve em árabe ou em outros 13 idiomas. - [Ler notas fiscais e recibos em árabe com um modelo de visão](https://broadrouter.com/pt/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Envie a foto de uma nota ou recibo a um modelo de visão na BroadNet Inference e receba total, data e fornecedor em JSON, com perguntas em árabe ou inglês. - [Use o Zed com a BroadNet Inference como provedor compatível com OpenAI](https://broadrouter.com/pt/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Adicione a BroadNet Inference ao Zed como provedor compatível com a OpenAI no settings.json, com IDs de modelo, limites de tokens, capabilities e variável da chave. - [Usar o Kilo Code com a BroadNet Inference (provedor OpenAI Compatible)](https://broadrouter.com/pt/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Conecte o Kilo Code à BroadNet Inference como provedor OpenAI Compatible personalizado, pela janela do VS Code ou pelo kilo.json, com IDs, limites e tools. - [Adicionar a BroadNet Inference ao goose como provedor personalizado](https://broadrouter.com/pt/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): Conecte o agente de IA goose à BroadNet Inference como custom provider OpenAI Compatible, pelo goose Desktop, pelo goose configure ou por um arquivo JSON. - [Usar o Pydantic AI com a BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/pt/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Rode agentes do Pydantic AI na BroadNet Inference com OpenAIChatModel e OpenAICompatibleProvider, com tool calls, saída estruturada e ajustes de perfil. - [CrewAI com a BroadNet Inference como endpoint OpenAI personalizado](https://broadrouter.com/pt/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): Aponte os agentes do CrewAI para a BroadNet Inference com LLM(custom_openai=True, base_url=...) e escolha modelos para ferramentas, contexto longo e preço fixo. - [Usar o LlamaIndex com a BroadNet Inference via OpenAILike](https://broadrouter.com/pt/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): Conecte o LlamaIndex à BroadNet Inference com a classe OpenAILike, definindo corretamente api_base, context_window, is_chat_model e is_function_calling_model. - [BroadNet Inference no Dify com o plugin OpenAI-API-compatible](https://broadrouter.com/pt/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): Conecte o Dify à BroadNet Inference pelo plugin OpenAI-API-compatible, com os valores exatos dos campos, tamanhos de contexto e ajustes de function calling. - [Jan com a BroadNet Inference: provedor compatível com a OpenAI](https://broadrouter.com/pt/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Adicione a BroadNet Inference ao app desktop Jan como provedor personalizado compatível com a OpenAI, cadastre os IDs dos modelos e ative ferramentas e visão. - [Conectar o SillyTavern à BroadNet Inference (Custom OpenAI-compatible)](https://broadrouter.com/pt/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): Conecte o SillyTavern à BroadNet Inference com a fonte Custom (OpenAI-compatible): IDs de modelo corretos, pós-processamento do prompt e limites de tokens. - [Usar o Spring AI com a BroadNet Inference (base-url da OpenAI)](https://broadrouter.com/pt/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Aponte o starter de chat da OpenAI do Spring AI para a BroadNet Inference com spring.ai.openai.base-url, um ID da BroadNet e max-tokens, e chame pelo ChatClient. - [Chamar a BroadNet Inference em Go com o SDK openai-go](https://broadrouter.com/pt/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): Use o SDK oficial openai-go v3 com a BroadNet Inference via option.WithBaseURL: chat completions, streaming com usage e tratamento de erros em Go. - [Criar agentes LangGraph na BroadNet Inference](https://broadrouter.com/pt/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): Rode um StateGraph ou create_agent do LangGraph na BroadNet Inference com ChatOpenAI, base_url e um ID de modelo da BroadNet, com ferramentas, streaming e memória. - [Usar pipelines e agentes do Haystack com a BroadNet Inference](https://broadrouter.com/pt/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Aponte o OpenAIChatGenerator do Haystack para a BroadNet Inference com api_base_url, um ID de modelo da BroadNet e uma chave Secret, para pipelines, tools e agentes. - [Usar o Semantic Kernel com a BroadNet Inference](https://broadrouter.com/pt/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Conecte o Microsoft Semantic Kernel à BroadNet Inference: em Python com um cliente AsyncOpenAI, em C# com um endpoint personalizado, para chat, plugins e agentes. - [Usar agentes Mastra com a BroadNet Inference](https://broadrouter.com/pt/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): Rode agentes Mastra na BroadNet Inference com um objeto model que define id, url e apiKey, incluindo o prefixo broadnet/ duplicado que o ID do modelo exige. - [Saída estruturada com o Instructor na BroadNet Inference](https://broadrouter.com/pt/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Extraia objetos Pydantic da BroadNet Inference com o Instructor, usando from_openai ou from_provider com um base_url personalizado, e escolha o Mode certo. - [Use o DSPy com a BroadNet Inference (defina model_type="chat")](https://broadrouter.com/pt/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): Configure o dspy.LM para a BroadNet Inference com o prefixo openai/, api_base e model_type="chat", porque o DSPy 3.5 manda modelos OpenAI para a Responses API. - [Avaliar os modelos da BroadNet com o promptfoo](https://broadrouter.com/pt/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): Rode evals do promptfoo e comparações lado a lado de modelos na BroadNet Inference com o provider openai:chat, apiBaseUrl e apiKeyEnvar, além de graders. - [Rodar o smolagents na BroadNet Inference com OpenAIModel](https://broadrouter.com/pt/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Conecte o smolagents da Hugging Face à BroadNet Inference com OpenAIModel e api_base: CodeAgent em qualquer modelo, ToolCallingAgent no GLM 5.3 e no Qwen3.8 27B. - [Conectar o Flowise à BroadNet Inference (OpenAI Custom Model)](https://broadrouter.com/pt/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Use a BroadNet Inference em chatflows do Flowise com o nó OpenAI Custom Model, um Base Path, uma credencial OpenAI API e o nome do modelo da BroadNet. - [Conectar o Langflow à BroadNet Inference](https://broadrouter.com/pt/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Use a BroadNet Inference no Langflow com o campo OpenAI API Base do componente OpenAI e um nome de modelo digitado, e veja por que o OpenAI Compatible não a lista. - [Glossary entries in Portuguese](https://broadrouter.com/pt/glossary/) - [Agente de IA (AI agent)](https://broadrouter.com/pt/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): Um agente de IA é um sistema em que um modelo de linguagem conduz os próprios passos rumo a um objetivo: decide quais ferramentas chamar, lê os resultados e repete até produzir uma resposta final ou atingir um limite. Um script fixo que chama um modelo em pontos definidos costuma ser chamado de workflow. - [Chave de API (API key)](https://broadrouter.com/pt/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): Uma chave de API é uma string secreta que identifica e autoriza um cliente que chama uma API. As APIs de LLM normalmente a esperam em um cabeçalho Authorization: Bearer, e qualquer pessoa que tenha a chave pode fazer requisições cobradas da conta dona dela. - [Diacríticos árabes (tashkeel)](https://broadrouter.com/pt/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): Diacríticos árabes, ou tashkeel, são pequenas marcas escritas acima ou abaixo das letras para indicar vogais curtas, consoantes dobradas e a ausência de vogal. São opcionais: a maior parte do texto árabe, incluindo notícias, não as usa, então leitores e software deduzem as vogais pelo contexto. - [Identificação de dialeto árabe (dialect identification)](https://broadrouter.com/pt/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): Identificação de dialeto árabe é a tarefa de decidir em qual variedade de árabe está um texto ou uma gravação, como egípcio, do Golfo, levantino, magrebino ou árabe padrão moderno, no nível de região, país ou cidade. Serve para rotear mensagens, escolher prompts e dividir conjuntos de avaliação. - [Normalização de texto árabe (Arabic text normalization)](https://broadrouter.com/pt/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): Normalização de texto árabe é a etapa que converte as variantes de grafia de uma mesma palavra em uma única forma antes da comparação: remove diacríticos e tatweel, unifica as formas de alef e muitas vezes troca ى por ي e ة por ه. Ela se aplica a chaves de busca e à avaliação, não ao texto exibido. - [Arabizi](https://broadrouter.com/pt/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): Arabizi é árabe escrito de modo informal com letras latinas, usando dígitos para sons árabes sem equivalente latino, como 3 para ع. É comum em chats e redes sociais, é quase sempre dialetal, não tem grafia padrão e muitas vezes vem misturado com inglês ou francês. - [AWQ (activation-aware weight quantization)](https://broadrouter.com/pt/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (activation-aware weight quantization) é um método pós-treinamento que comprime os pesos de um LLM para baixa precisão, geralmente inteiros de 4 bits. Ele usa estatísticas de ativação de um pequeno conjunto de calibração para achar os canais de pesos mais importantes e os escala para que percam menos precisão. Não precisa de backpropagation. - [Base URL (URL base)](https://broadrouter.com/pt/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): A base URL é o endereço raiz fixo de uma API, como https://api.example.com/v1, que um cliente coloca antes de cada caminho de endpoint, como /chat/completions. Nos SDKs compatíveis com a OpenAI, trocar a base URL, junto com a chave de API e o ID do modelo, aponta o mesmo código para outro provedor. - [Batch API (API de lote)](https://broadrouter.com/pt/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Batch API é um endpoint assíncrono de LLM que aceita muitas requisições de uma vez, normalmente em um arquivo, processa tudo dentro de uma janela definida em vez de na hora e devolve os resultados juntos. Provedores como OpenAI e Anthropic cobram as requisições em lote com desconto em relação às APIs padrão. - [BM25](https://broadrouter.com/pt/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) é uma função de ranking para busca por palavra-chave que pontua um documento conforme a frequência dos termos da consulta nele, a raridade desses termos na coleção e o tamanho do documento. É a similaridade padrão no Lucene e no Elasticsearch. - [Byte pair encoding (BPE)](https://broadrouter.com/pt/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Byte pair encoding (BPE) é um algoritmo de tokenização em subpalavras que parte de caracteres ou bytes individuais e junta repetidamente o par adjacente mais frequente em uma nova entrada do vocabulário, até chegar ao tamanho de vocabulário desejado. A maioria dos tokenizers de LLM atuais usa BPE ou uma variante em nível de byte. - [Tokens de entrada em cache (cached input)](https://broadrouter.com/pt/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): Tokens de entrada em cache são tokens do prompt que um provedor de LLM processou há pouco e pode reaproveitar de um cache em vez de calcular de novo, normalmente um prefixo repetido, como um system prompt ou um documento. Os provedores os cobram com uma tarifa menor que a dos tokens de entrada novos. - [Chain-of-thought (CoT, cadeia de pensamento)](https://broadrouter.com/pt/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Chain-of-thought (CoT, cadeia de pensamento) é uma técnica de prompting em que se mostra ou se pede a um modelo de linguagem que escreva passos intermediários de raciocínio antes da resposta final. Ela melhora a precisão em problemas de várias etapas, como aritmética e lógica, e os modelos de raciocínio agora a geram automaticamente. - [Character error rate (CER)](https://broadrouter.com/pt/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Character error rate (CER, taxa de erro de caracteres) mede a precisão de uma transcrição ou OCR no nível de caractere: substituições, exclusões e inserções necessárias para transformar a saída no texto de referência, divididas pelo número de caracteres da referência. Quanto menor, melhor, e 0 é uma correspondência perfeita. - [Chat Completions API](https://broadrouter.com/pt/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): A Chat Completions API é o endpoint HTTP da OpenAI, POST /v1/chat/completions, que recebe o nome de um modelo e uma lista de mensagens marcadas com papéis e devolve a próxima mensagem do modelo. O formato da requisição e da resposta virou o padrão comum que provedores e SDKs compatíveis com a OpenAI reutilizam. - [Chat template (template de chat)](https://broadrouter.com/pt/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): Chat template é a regra de formatação, geralmente escrita em Jinja e distribuída com o tokenizer do modelo, que converte uma lista de mensagens de chat com papéis na sequência única de tokens que o modelo espera, incluindo os tokens especiais de controle. Usar o template errado pode piorar muito as respostas do modelo. - [Chunking (RAG)](https://broadrouter.com/pt/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Chunking é a etapa de um pipeline de recuperação que divide os documentos em pedaços menores, chamados chunks, antes de gerar embeddings e indexá-los. O chunk size define quanto texto cada pedaço tem, e o chunk overlap repete parte do texto entre chunks vizinhos para que ideias cortadas numa fronteira não se percam. - [Code-switching (alternância de código)](https://broadrouter.com/pt/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Code-switching é o uso de dois ou mais idiomas, ou variedades de um mesmo idioma, dentro da mesma conversa ou frase, na fala ou na escrita. Exemplos são falantes de árabe que misturam palavras em inglês ao dialeto, ou que alternam entre o árabe padrão moderno e o próprio dialeto. - [Agente de programação com IA (AI coding agent)](https://broadrouter.com/pt/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): Um agente de programação com IA é uma ferramenta de software em que um modelo de linguagem trabalha numa codebase em ciclo: lê arquivos, edita, roda comandos como os testes, lê a saída e decide o próximo passo até a tarefa terminar ou um limite ser atingido. - [Concorrência (concurrency)](https://broadrouter.com/pt/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): Concorrência é o número de requisições que um cliente tem em andamento ao mesmo tempo. APIs de LLM muitas vezes a limitam por chave ou por plano, separadamente das requisições por minuto, porque cada requisição aberta ocupa capacidade de GPU até a resposta terminar. - [Context compaction (compactação de contexto)](https://broadrouter.com/pt/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): Context compaction é a prática de resumir uma conversa que está chegando ao limite da janela de contexto e continuar a sessão a partir desse resumo, para que o agente mantenha decisões importantes e o estado, descartando histórico bruto como saídas antigas de ferramentas. - [Context engineering (engenharia de contexto)](https://broadrouter.com/pt/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): Engenharia de contexto é o trabalho de escolher e manter o conjunto de tokens que um modelo de linguagem vê a cada requisição, incluindo system prompt, definições de ferramentas, histórico, documentos recuperados e anotações, para que o modelo receba o menor conjunto de informações de alto sinal de que precisa para fazer bem a tarefa. - [Context rot (degradação de contexto)](https://broadrouter.com/pt/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): Context rot é a perda gradual de precisão e confiabilidade que os modelos de linguagem mostram conforme a entrada fica mais longa, mesmo quando a tarefa é a mesma e a entrada está bem dentro da janela de contexto. O termo foi popularizado por um relatório técnico da Chroma de 2025 que testou 18 modelos. - [Janela de contexto (context window)](https://broadrouter.com/pt/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): A janela de contexto é o número máximo de tokens com que um modelo de linguagem consegue trabalhar em uma única requisição. Ela abrange o prompt, o histórico da conversa, as definições de ferramentas, as imagens e a saída gerada, incluindo os tokens de raciocínio. O que não couber precisa ser cortado, resumido ou dividido em várias requisições. - [Residência de dados (data residency)](https://broadrouter.com/pt/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): Residência de dados é um requisito ou compromisso de que os dados sejam armazenados e processados em um país ou região específicos. Em uma API de LLM, cobre onde os prompts são processados durante a inferência e onde fica qualquer conteúdo armazenado, e muitas vezes exclui dados de conta e de faturamento. - [Árabe dialetal (dialectal Arabic)](https://broadrouter.com/pt/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): Árabe dialetal é o árabe falado no dia a dia de uma região ou cidade, como o árabe egípcio, do Golfo, levantino ou magrebino. Os dialetos diferem do árabe padrão moderno em sons, palavras e gramática, não têm ortografia padrão e são cada vez mais escritos na internet, o que os torna mais difíceis para modelos de linguagem. - [Diarização de locutores (speaker diarization)](https://broadrouter.com/pt/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): Diarização de locutores é a tarefa de dividir uma gravação de áudio em segmentos por falante, respondendo "quem falou e quando". Ela marca cada trecho de fala com um ID anônimo, como SPEAKER_00, e costuma ser combinada com uma transcrição para mostrar quem disse cada fala. - [Destilação (knowledge distillation)](https://broadrouter.com/pt/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): Destilação, ou knowledge distillation, é um método de treinamento em que um modelo aluno menor aprende a reproduzir as saídas de um modelo professor maior, como suas distribuições de probabilidade ou respostas geradas, para manter boa parte da qualidade do professor com um custo de execução menor. - [Acordo de tratamento de dados (DPA)](https://broadrouter.com/pt/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): Um acordo de tratamento de dados (DPA) é o contrato entre o controlador e o operador que define como o operador trata dados pessoais. O artigo 28(3) do GDPR o exige e lista o conteúdo: instruções documentadas, confidencialidade, segurança, regras para suboperadores, exclusão ou devolução dos dados e auditorias. - [Embeddings](https://broadrouter.com/pt/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): Embeddings são listas de números de ponto flutuante (vetores) que um modelo produz para representar um texto, uma imagem ou outra entrada. Entradas com significado parecido recebem vetores próximos, então a distância entre vetores serve para ordenar resultados de busca, agrupar documentos ou encontrar itens relacionados. - [Eval (avaliação de LLM)](https://broadrouter.com/pt/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Uma eval (abreviação de evaluation, avaliação) é um teste repetível de um modelo de linguagem ou de um prompt: um conjunto fixo de entradas, uma descrição de como é uma boa saída e um avaliador (código, uma pessoa ou outro modelo) que pontua cada saída, para que versões possam ser comparadas. - [Backoff exponencial (exponential backoff)](https://broadrouter.com/pt/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): Backoff exponencial é uma estratégia de retry em que o cliente espera mais tempo depois de cada tentativa que falha, normalmente dobrando o atraso até um teto, antes de tentar de novo. Adicionar um jitter aleatório a cada espera espalha os retries de muitos clientes, para que não atinjam o servidor no mesmo instante. - [Few-shot prompting (prompt com exemplos)](https://broadrouter.com/pt/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): Few-shot prompting é uma técnica em que o prompt inclui um pequeno número de exemplos resolvidos de entrada e saída antes da entrada real, para que o modelo de linguagem deduza deles a tarefa e o formato. Os pesos do modelo não mudam; zero-shot não usa exemplos e one-shot usa um. - [Fine-tuning (ajuste fino de LLM)](https://broadrouter.com/pt/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): Fine-tuning é o processo de continuar treinando um modelo de linguagem pré-treinado com um conjunto de dados menor e específico, como conversas de exemplo, para que seus pesos mudem e ele execute aquela tarefa ou estilo de forma mais confiável, sem instruções longas ou exemplos em todo prompt. - [finish_reason](https://broadrouter.com/pt/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason é um campo de cada choice na resposta de uma chat completion que informa por que o modelo parou de gerar. No formato da OpenAI, os valores são stop, length, tool_calls, content_filter e o obsoleto function_call. A Messages API da Anthropic informa a mesma coisa como stop_reason. - [Plano de LLM com preço fixo (flat-rate)](https://broadrouter.com/pt/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): Um plano de LLM com preço fixo cobra um valor fixo por um período, como um mês, pelo acesso à API de um modelo de linguagem, em vez de cobrar por token. Esses planos costumam limitar outra coisa, como simultaneidade, velocidade, modelos ou janelas de uso, então os limites importam tanto quanto o preço. - [FP8 (ponto flutuante de 8 bits)](https://broadrouter.com/pt/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 é um formato numérico de ponto flutuante de 8 bits usado em deep learning, definido em duas codificações: E4M3, com 4 bits de expoente e 3 de mantissa, e E5M2, com 5 bits de expoente e 2 de mantissa. Ele reduz pela metade a memória em relação aos formatos de 16 bits e roda nativamente em GPUs recentes. - [Frequency e presence penalty](https://broadrouter.com/pt/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): Frequency e presence penalty são configurações de amostragem que reduzem a pontuação de tokens que já estão na saída. A presence penalty se aplica assim que o token aparece; a frequency penalty cresce a cada repetição. Na API da OpenAI, ambas vão de -2.0 a 2.0, com padrão 0. - [GDPR (Regulamento Geral de Proteção de Dados)](https://broadrouter.com/pt/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): O GDPR (General Data Protection Regulation, Regulamento (UE) 2016/679) é a lei da UE sobre o tratamento de dados pessoais. Vale desde 25 de maio de 2018 para organizações estabelecidas na UE e algumas de fora. Em APIs de LLM, regula os contratos com operadores, as transferências internacionais e a minimização de dados. - [GGUF (formato de modelo do llama.cpp)](https://broadrouter.com/pt/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF é um formato de arquivo binário para armazenar modelos de linguagem para inferência com runtimes baseados em GGML, como o llama.cpp. Um único arquivo GGUF guarda os pesos, muitas vezes quantizados, e os metadados necessários para carregar o modelo, incluindo detalhes do tokenizer e da arquitetura, e pode ser mapeado em memória para carregar rápido. - [Hora de GPU (GPU-hour)](https://broadrouter.com/pt/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): Uma hora de GPU (GPU-hour) é um processador gráfico usado durante uma hora. Provedores de nuvem cobram instâncias de GPU alugadas por hora ou por segundo, e pesquisadores usam horas de GPU para dizer quanta computação um trabalho exigiu. Você paga pelo tempo da máquina, ocupada ou não. - [Grounding (ancoragem em fontes)](https://broadrouter.com/pt/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): Grounding é a prática de ligar a saída de um modelo de linguagem a fontes de informação específicas e verificáveis, como documentos fornecidos, resultados de busca ou registros de banco de dados, e instruí-lo a responder com base nelas. Respostas ancoradas podem ser rastreadas até uma fonte e têm menos chance de trazer fatos inventados. - [Guardrails (barreiras de proteção para LLM)](https://broadrouter.com/pt/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): Guardrails são checagens que uma aplicação executa em volta de uma chamada a um modelo de linguagem para mantê-la dentro de limites definidos. Guardrails de entrada filtram o que entra no modelo, os de saída verificam o que sai e os de ferramenta validam ações antes que rodem. Uma checagem que falha bloqueia, reescreve ou sinaliza a requisição. - [Alucinação (hallucination)](https://broadrouter.com/pt/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): Alucinação é uma saída de um modelo de linguagem que parece plausível, mas é falsa, inventada ou não tem apoio na entrada recebida, como uma citação inventada, um número errado ou uma API que não existe. Ela surge porque o modelo prevê o texto mais provável em vez de checar fatos. - [Hamza (letra árabe)](https://broadrouter.com/pt/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): Hamza (ء) é a letra árabe que representa a oclusiva glotal. Ela é escrita sozinha ou sobre um suporte: alef (أ إ), waw (ؤ) ou yeh (ئ), e آ junta alef e madda. Como o texto digitado é inconsistente com ela, buscadores e a avaliação de fala costumam reduzir as formas com alef a ا. - [Busca híbrida (hybrid search)](https://broadrouter.com/pt/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): Busca híbrida é um método de recuperação que executa, na mesma consulta, uma busca por palavras-chave, como BM25, e uma busca vetorial sobre embeddings, e depois funde as duas listas ranqueadas em uma só, muitas vezes com reciprocal rank fusion (RRF) ou com uma soma ponderada de scores normalizados. - [Chave de idempotência (idempotency key)](https://broadrouter.com/pt/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Uma chave de idempotência é um valor único, normalmente um ID aleatório como um UUID, que o cliente anexa a uma requisição que altera estado, como um pagamento ou uma compra. Se a mesma requisição for enviada de novo com a mesma chave, o servidor retorna o resultado original em vez de executar a ação duas vezes. - [Tokens de imagem (image tokens)](https://broadrouter.com/pt/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): Tokens de imagem são os tokens de entrada que um modelo de visão e linguagem usa para representar uma imagem no prompt. Os provedores os calculam a partir do tamanho da imagem, normalmente dividindo-a em patches ou tiles de tamanho fixo depois de redimensioná-la, e os cobram como tokens de entrada de texto, então imagens maiores custam mais. - [Provedor de inferência (inference provider)](https://broadrouter.com/pt/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): Um provedor de inferência é uma empresa que roda modelos de machine learning, muitas vezes LLMs open-weight, em GPUs próprias, arrendadas ou alugadas, e vende o acesso a eles por uma API, normalmente cobrando por token, por requisição ou por unidade de tempo. Ele cuida do serving, da escala e do hardware. - [Tokens de entrada e de saída](https://broadrouter.com/pt/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): Tokens de entrada são os tokens que um modelo de linguagem lê em uma requisição, incluindo o system prompt, o histórico da conversa, as definições de ferramentas e as imagens. Tokens de saída são os que ele gera na resposta, incluindo o raciocínio. As APIs de LLM costumam cobrá-los separadamente, e a saída costuma custar mais. - [INT4 (quantização de 4 bits)](https://broadrouter.com/pt/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 é um formato numérico inteiro de 4 bits que representa 16 níveis. Na inferência de LLMs, costuma significar quantização de 4 bits só dos pesos (W4A16): os pesos são guardados em INT4 com fatores de escala e expandidos para 16 bits durante o cálculo, usando cerca de um quarto da memória de pesos em 16 bits. - [Lista de IPs permitidos (IP allow list)](https://broadrouter.com/pt/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): Uma lista de IPs permitidos é um controle de segurança que aceita requisições apenas de endereços IP ou faixas CIDR listados e recusa todos os outros. Em uma API, ela limita o estrago de uma chave vazada, porque a chave não funciona a partir de redes que não estão na lista. - [Jailbreak (LLM)](https://broadrouter.com/pt/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): Jailbreak é uma entrada criada para fazer um modelo de linguagem ignorar seu treinamento de segurança ou as regras definidas pela aplicação que o usa, de modo que ele produza conteúdo ou execute ações que normalmente recusaria. A OWASP trata o jailbreaking como uma forma de prompt injection. - [JSON mode (modo JSON)](https://broadrouter.com/pt/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): JSON mode é uma configuração de APIs de LLM, response_format {"type": "json_object"} em APIs no estilo da OpenAI, que faz a resposta do modelo ser JSON válido e parseável. Ao contrário de structured outputs com JSON Schema, ele não garante quais chaves, tipos ou valores o objeto contém. - [JSON Schema](https://broadrouter.com/pt/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema é um vocabulário, escrito em JSON, para descrever e validar a estrutura de dados JSON: tipos, campos obrigatórios, valores permitidos e aninhamento. A versão atual é a 2020-12. APIs de LLM o usam para definir parâmetros de tools e o formato exato que uma saída estruturada deve seguir. - [KV cache (cache de chave-valor)](https://broadrouter.com/pt/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): KV cache é a memória em que um modelo de linguagem transformer guarda as chaves e os valores de atenção calculados para cada token já processado, de modo que cada novo token os reutiliza em vez de recalcular a sequência inteira. Ele cresce com o tamanho do contexto e muitas vezes limita quantas requisições uma GPU consegue atender. - [Latência (latency)](https://broadrouter.com/pt/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): Latência, em uma API de modelos de linguagem, é o tempo entre o envio de uma requisição e o recebimento do último token de saída. É a soma do time to first token com o tempo de geração e costuma ser informada em percentis, como p50 e p95, e não como média. - [LLM-as-a-judge (LLM como juiz)](https://broadrouter.com/pt/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge é um método de avaliação em que um modelo de linguagem avalia saídas de outros modelos, dando nota a uma única resposta com base em uma rubrica ou escolhendo a melhor de duas. Aproxima o julgamento humano a um custo bem menor, mas tem vieses conhecidos, como favorecer a primeira resposta ou a mais longa. - [Benchmark de LLM](https://broadrouter.com/pt/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): Um benchmark de LLM é um conjunto fixo de tarefas, normalmente público, com um método de pontuação, usado para medir e comparar modelos de linguagem em uma habilidade, como perguntas de conhecimento, programação ou reconhecimento de fala. A nota mostra a capacidade relativa naquele teste, não como um modelo vai se sair na sua carga de trabalho. - [LLM router (roteador de LLMs)](https://broadrouter.com/pt/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): Um LLM router é um software ou serviço que fica entre uma aplicação e uma ou mais APIs de modelos de linguagem e decide, a cada requisição, qual modelo ou provedor vai atendê-la, com base em regras como custo, latência, disponibilidade ou tipo de tarefa. - [Logprobs (log-probabilidades)](https://broadrouter.com/pt/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Logprobs são as log-probabilidades que um modelo de linguagem atribuiu aos tokens que gerou. Uma API que os retorna mostra, para cada token de saída, quão provável o modelo o considerou e, opcionalmente, os tokens alternativos mais prováveis naquela posição. - [LoRA (low-rank adaptation)](https://broadrouter.com/pt/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation, adaptação de baixo posto) é um método de fine-tuning eficiente em parâmetros que congela os pesos de um modelo pré-treinado e treina um par de pequenas matrizes de baixo posto para camadas escolhidas. A atualização aprendida pode ser guardada como um pequeno adapter ou mesclada aos pesos, sem adicionar latência na inferência. - [Lost in the middle (perdido no meio)](https://broadrouter.com/pt/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle é a tendência observada de modelos de linguagem usarem melhor as informações relevantes quando elas aparecem perto do início ou do fim de uma entrada longa, e pior quando ficam no meio. O nome vem de um estudo de 2023 de Liu et al. - [max_tokens](https://broadrouter.com/pt/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens é um parâmetro de APIs de LLM que define o máximo de tokens que um modelo pode gerar em uma resposta. Quando a geração atinge o limite, ela para e a resposta é marcada como cortada, por exemplo com finish_reason igual a "length". - [MCP (Model Context Protocol)](https://broadrouter.com/pt/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol) é um protocolo aberto, baseado em JSON-RPC 2.0, que padroniza como aplicações de LLM se conectam a ferramentas e dados externos. Um servidor MCP expõe ferramentas, recursos e prompts; um client MCP dentro da aplicação host os descobre e repassa ao servidor as chamadas de ferramenta feitas pelo modelo. - [Mixture of experts (MoE)](https://broadrouter.com/pt/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Mixture of experts (MoE) é uma arquitetura de rede neural em que uma camada contém muitas sub-redes especialistas e um roteador aprendido envia cada token para apenas algumas delas. Um modelo de linguagem MoE tem um número total de parâmetros grande, mas usa um pequeno subconjunto ativo para cada token. - [Descontinuação de modelos (model deprecation)](https://broadrouter.com/pt/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): A descontinuação (deprecation) de um modelo acontece quando um provedor de API anuncia que deixará de oferecê-lo, normalmente informando uma data de aposentadoria ou desligamento e um substituto recomendado. Depois dessa data, as requisições ao modelo falham, então as aplicações precisam migrar antes para outro ID de modelo. - [Fallback de modelo (model fallback)](https://broadrouter.com/pt/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): Fallback de modelo é um padrão de confiabilidade em que uma requisição de LLM que falhou é repetida em outro modelo ou provedor de uma lista ordenada. O gatilho pode ser um erro, um rate limit, um erro de tamanho de contexto ou uma recusa por política de conteúdo. Roda no cliente ou em um gateway na frente das APIs. - [Licença do modelo (model license)](https://broadrouter.com/pt/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): A licença de um modelo são os termos legais sob os quais os pesos, o código e as saídas de um modelo podem ser usados, alterados e compartilhados. Modelos open-weight usam licenças de software padrão, como Apache 2.0 ou MIT, ou licenças próprias que podem acrescentar condições, como limites para grandes serviços comerciais. - [Árabe padrão moderno (MSA)](https://broadrouter.com/pt/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): O árabe padrão moderno (MSA) é a forma padronizada do árabe usada na escrita, no noticiário, na educação e na fala formal em todo o mundo árabe. Ele não é a língua materna do dia a dia de ninguém: as pessoas falam dialetos regionais, como o egípcio ou o árabe do Golfo, e passam para o MSA em contextos formais. - [Sistema multiagente (multi-agent system)](https://broadrouter.com/pt/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): Um sistema multiagente é uma aplicação de LLM em que vários agentes, cada um com seu próprio prompt, ferramentas e contexto, trabalham em partes de uma mesma tarefa. Um desenho comum é o orchestrator-worker: um agente líder planeja, inicia agentes workers em paralelo e combina os resultados deles. - [IA multimodal (multimodal AI)](https://broadrouter.com/pt/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): IA multimodal é um modelo ou sistema que trabalha com mais de um tipo de dado, como texto, imagens, vídeo e áudio. Um modelo multimodal pode ler uma foto e uma pergunta juntas, ou receber fala e responder em texto. O tipo mais comum é o modelo de visão e linguagem (vision-language model). - [Reconhecimento de entidades nomeadas (NER)](https://broadrouter.com/pt/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): Reconhecimento de entidades nomeadas (NER) é a tarefa de encontrar nomes em um texto e rotular cada um com um tipo, como pessoa, organização ou local. Sistemas clássicos rotulam cada token; já aos LLMs costuma-se pedir uma lista de entidades em JSON, que depois é conferida contra o texto original. - [Reconhecimento óptico de caracteres (OCR)](https://broadrouter.com/pt/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): OCR (reconhecimento óptico de caracteres) é a tecnologia que detecta texto impresso ou manuscrito em uma imagem, digitalização ou PDF e o converte em caracteres legíveis por máquina. O OCR clássico devolve o texto e a posição dele na página; modelos de visão e linguagem conseguem ler o texto e responder perguntas sobre ele diretamente. - [Ogg Opus](https://broadrouter.com/pt/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus é áudio codificado com o codec Opus (RFC 6716) e guardado em um contêiner Ogg (RFC 7845). Ele foi feito para fala e áudio interativo com bitrates baixos, e é o formato que o WhatsApp usa para mensagens de voz, normalmente como arquivos .ogg ou .opus. - [Modelo de pesos abertos (open-weight)](https://broadrouter.com/pt/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): Um modelo de pesos abertos (open-weight) é um modelo de machine learning cujos pesos treinados são publicados para qualquer pessoa baixar e rodar, em geral sob uma licença com condições. Ele difere da IA totalmente open source, que também libera o código de treinamento e informações detalhadas sobre os dados, para que outros possam estudar e reproduzir o modelo. - [API compatível com a OpenAI](https://broadrouter.com/pt/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): Uma API compatível com a OpenAI é uma API HTTP que aceita os mesmos formatos de requisição e de resposta da API da OpenAI, como POST /v1/chat/completions, de modo que SDKs e ferramentas da OpenAI podem usá-la trocando só a base URL, a chave de API e o nome do modelo. - [Chamadas de ferramentas em paralelo (parallel tool calls)](https://broadrouter.com/pt/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Chamadas de ferramentas em paralelo acontecem quando um modelo devolve mais de uma chamada de ferramenta em uma única resposta, por exemplo três leituras de arquivo de uma vez. No formato da OpenAI elas chegam como várias entradas em tool_calls, e parallel_tool_calls como false limita o modelo a zero ou uma chamada por turno. - [Pagamento por uso (pay-as-you-go)](https://broadrouter.com/pt/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): Pagamento por uso é um modelo de preço em que você paga só pelo que usa, medido por unidade, como tokens, requisições ou minutos de áudio, sem assinatura nem compromisso mínimo. A maioria das APIs de LLM cobra assim, a partir de um saldo pré-pago ou de uma fatura periódica. - [PDPL (leis de proteção de dados da Arábia Saudita e EAU)](https://broadrouter.com/pt/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) é o nome comum da lei saudita de dados pessoais, editada pelo Decreto Real M/19 e aplicada pela SDAIA, e do Decreto-Lei Federal nº 45 de 2021 dos EAU. As duas regulam o tratamento de dados pessoais e impõem condições a transferências ao exterior, incluindo prompts enviados a APIs de LLM estrangeiras. - [PII (informações pessoais identificáveis)](https://broadrouter.com/pt/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): PII (personally identifiable information) é a informação que permite distinguir ou rastrear a identidade de uma pessoa, como nome, número de documento ou registro biométrico, mais outras informações ligadas ou vinculáveis a ela. O termo mais amplo do GDPR é dados pessoais. Com LLMs, a anonimização de PII remove esses dados dos prompts antes que saiam dos seus sistemas. - [Prefill e decode](https://broadrouter.com/pt/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill e decode são as duas fases da inferência de um LLM. No prefill, o modelo processa o prompt inteiro em paralelo e preenche o KV cache; no decode, gera a resposta um token por vez, reaproveitando esse cache. O prefill define principalmente o tempo até o primeiro token; o decode, a velocidade de saída. - [Crédito pré-pago (prepaid credit)](https://broadrouter.com/pt/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): Crédito pré-pago é dinheiro adicionado a uma conta de API antes do uso. Cada requisição desconta o próprio custo do saldo e, quando o saldo acaba, as requisições cobradas por uso param até você fazer uma recarga. Ele limita o gasto ao que você já pagou. - [Engenharia de prompt (prompt engineering)](https://broadrouter.com/pt/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): Engenharia de prompt é o processo de escrever, estruturar e testar as instruções dadas a um modelo de linguagem para que ele produza, de forma consistente, uma saída que atenda aos seus requisitos. Inclui mensagens de papel, instruções claras, exemplos, formatação e a medição dos resultados com casos de teste. - [Prompt injection (injeção de prompt)](https://broadrouter.com/pt/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): Prompt injection é um ataque em que um texto entregue a um modelo de linguagem, por um usuário ou escondido em uma página web, e-mail, arquivo ou resultado de ferramenta, muda o comportamento do modelo contra a intenção do desenvolvedor. A OWASP o coloca em primeiro lugar (LLM01:2025) no seu Top 10 para aplicações de LLM. - [Quantização (quantization)](https://broadrouter.com/pt/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): Quantização é o processo de armazenar os números de uma rede neural, principalmente os pesos, em um formato de menor precisão, como 8 ou 4 bits em vez de 16 bits, para que o modelo use menos memória e muitas vezes rode mais rápido. Bons métodos mantêm pequena a perda de precisão. - [RAG (geração aumentada por recuperação)](https://broadrouter.com/pt/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG (retrieval-augmented generation, ou geração aumentada por recuperação) é um padrão em que a aplicação primeiro busca, em um repositório externo como um índice de busca ou vetorial, trechos relevantes para uma pergunta e depois os adiciona ao prompt, para que um modelo de linguagem responda a partir desse texto e não só da memória. - [Rate limit (limite de taxa)](https://broadrouter.com/pt/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): Um rate limit é um teto que uma API impõe a quanto um cliente pode usá-la em um intervalo de tempo, como requisições por minuto ou tokens por minuto. As requisições acima do limite são recusadas, normalmente com HTTP 429 Too Many Requests, até que o uso volte a ficar abaixo dele. - [ReAct (reason and act)](https://broadrouter.com/pt/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (reason and act) é um padrão de prompting, apresentado por Yao et al. em 2022, em que um modelo de linguagem intercala traços de raciocínio com ações, como buscas ou chamadas de ferramentas. Ele lê cada observação, atualiza o plano e repete até conseguir dar uma resposta final. - [Tokens de raciocínio (reasoning tokens)](https://broadrouter.com/pt/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): Tokens de raciocínio são tokens que um modelo de raciocínio gera para trabalhar um problema antes de escrever a resposta visível. Normalmente são cobrados como tokens de saída, contam para o limite de saída e para a janela de contexto, e a API pode ocultá-los ou resumi-los. - [Timeout de requisição (request timeout)](https://broadrouter.com/pt/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): Timeout de requisição é o tempo máximo que um cliente HTTP espera por um servidor antes de abandonar a requisição. Os clientes costumam dividi-lo em timeout de conexão, para abrir a conexão, e timeout de leitura, para esperar os dados. Chamadas a LLMs precisam de timeouts de leitura generosos, porque respostas longas levam tempo. - [Reranking (reordenação de resultados)](https://broadrouter.com/pt/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Reranking é uma segunda etapa de recuperação em que um modelo mais preciso, porém mais lento, em geral um cross-encoder, pontua cada par de consulta e documento candidato vindo de uma primeira busca e os reordena, para que só os poucos mais relevantes sejam passados ao modelo de linguagem. - [Texto RTL (da direita para a esquerda)](https://broadrouter.com/pt/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): Texto da direita para a esquerda (RTL) é a escrita, como a árabe ou a hebraica, exibida da direita para a esquerda. O computador a armazena na ordem lógica de leitura, e o algoritmo bidirecional do Unicode decide a ordem de exibição quando ela se mistura com texto da esquerda para a direita, dígitos, nomes de modelos ou código. - [Taxa de amostragem (sample rate)](https://broadrouter.com/pt/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): Taxa de amostragem (sample rate) é o número de amostras de áudio capturadas por segundo, medido em hertz (Hz). Modelos de reconhecimento de fala costumam esperar áudio mono a 16.000 Hz (16 kHz). Junto com a profundidade de bits e o número de canais, ela define quanto um arquivo de áudio sem compressão ocupa por minuto. - [Parâmetro seed](https://broadrouter.com/pt/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): O parâmetro seed é um inteiro que você envia com a requisição para que a API tente amostrar os mesmos tokens quando a mesma requisição for repetida. A OpenAI o descreve como best effort: o determinismo não é garantido, e o campo system_fingerprint mostra quando a configuração do backend mudou. - [Inferência serverless (serverless inference)](https://broadrouter.com/pt/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): Inferência serverless é uma forma de rodar modelos de machine learning em que a plataforma inicia, escala e desliga automaticamente a computação por trás de um endpoint, muitas vezes até zero quando ocioso, e cobra pelo uso em vez de máquinas reservadas. A contrapartida é um possível atraso de cold start. - [SLA (acordo de nível de serviço)](https://broadrouter.com/pt/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): SLA (service level agreement, ou acordo de nível de serviço) é um compromisso contratual de um fornecedor com um nível de serviço mensurável, em geral um percentual de uptime mensal, com uma compensação definida, em geral créditos de serviço, se não for cumprido. O SLA define o que conta como downtime, o que fica excluído e como reivindicar. - [Speculative decoding (decodificação especulativa)](https://broadrouter.com/pt/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): Speculative decoding (decodificação especulativa) é uma técnica de inferência em que um modelo draft pequeno e rápido propõe vários próximos tokens e o modelo alvo, maior, os verifica em um único forward pass, mantendo os aceitos. Ela acelera a geração sem mudar a distribuição da saída. - [Fala para texto (speech-to-text)](https://broadrouter.com/pt/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): Fala para texto, ou reconhecimento automático de fala (ASR), é a tecnologia que converte áudio falado em texto. Uma API de fala para texto recebe um arquivo ou stream de áudio e retorna uma transcrição, normalmente cobrada por minuto de áudio. A precisão é medida pela taxa de erro de palavras e varia com sotaque, dialeto e qualidade da gravação. - [Limite de gastos (spend limit)](https://broadrouter.com/pt/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): Limite de gastos é um teto para quanto uma conta, um projeto ou uma chave de API pode gastar em um período, em geral um mês. Quando o uso chega nele, o provedor rejeita as próximas requisições, muitas vezes com erro HTTP 429 ou 400, até o período reiniciar ou o limite ser aumentado. - [Legendas SRT (arquivo .srt)](https://broadrouter.com/pt/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) é um formato de legenda em texto puro. Cada legenda tem um número sequencial, um tempo de início e fim escrito como hh:mm:ss,mmm --> hh:mm:ss,mmm, uma ou mais linhas de texto e uma linha em branco. WebVTT é o formato web parecido usado pelo vídeo em HTML. - [Server-sent events (SSE)](https://broadrouter.com/pt/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Server-sent events (SSE) é um padrão web para um servidor enviar a um cliente um fluxo de eventos de texto por uma única resposta HTTP de longa duração, com content type text/event-stream. Cada evento tem uma ou mais linhas de campo, como data:, e termina com uma linha em branco. - [Sequências de parada (stop sequences)](https://broadrouter.com/pt/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): Sequências de parada são strings que você envia com a requisição e que fazem o modelo parar de gerar assim que produz uma delas. A OpenAI chama o campo de stop e aceita até 4 sequências; a Anthropic o chama de stop_sequences. Normalmente o texto encontrado fica fora da saída retornada. - [Streaming](https://broadrouter.com/pt/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): Streaming é uma forma de devolver a resposta de um modelo de linguagem em pequenos pedaços enquanto ela ainda está sendo gerada, em vez de uma resposta única no final. APIs compatíveis com a OpenAI fazem streaming via server-sent events quando a requisição define stream como true. - [Saída estruturada (structured output)](https://broadrouter.com/pt/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): Saída estruturada é um recurso de APIs de LLM que restringe a resposta de um modelo a um formato definido pelo desenvolvedor, normalmente um JSON Schema, para que o código consiga interpretá-la de forma confiável. É mais rigorosa que o JSON mode, que garante um JSON válido, mas não uma estrutura específica. - [Subagente (subagent)](https://broadrouter.com/pt/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): Um subagente é um agente separado que um agente principal inicia para cuidar de uma tarefa específica. Ele roda com a própria janela de contexto, o próprio system prompt e acesso próprio a ferramentas, faz o trabalho e devolve ao agente principal só um resultado ou resumo, o que mantém a conversa principal pequena. - [Suboperador (subprocessor)](https://broadrouter.com/pt/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): Suboperador (subprocessor, ou subcontratante ulterior no texto europeu) é um terceiro que um operador contrata para tratar dados pessoais em nome do controlador. Em uma API de LLM, pode incluir hosts de GPU, provedores de inferência parceiros e capacidade de backup. Pelo artigo 28 do GDPR, exige autorização do controlador e as mesmas obrigações de proteção de dados. - [Prompt de sistema (system prompt)](https://broadrouter.com/pt/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): O system prompt, ou prompt de sistema, é o texto de instruções que uma aplicação envia a um modelo de linguagem antes da conversa, em geral como mensagem com o papel system ou developer. Ele define o papel, as regras, o tom e o formato de saída do modelo, que deve segui-lo mesmo quando mensagens do usuário o contradizem. - [Tā’ marbūṭa (ta marbuta, ة)](https://broadrouter.com/pt/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): Ta marbuta (ة, U+0629) é uma letra árabe que só aparece no fim da palavra, muitas vezes marcando substantivos e adjetivos femininos. Ela parece um heh (ه, U+0647) com dois pontos, por isso o texto digitado mistura as duas, e muitos normalizadores de busca convertem ة em ه. - [Temperatura (temperature)](https://broadrouter.com/pt/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): Temperatura é um parâmetro de amostragem que controla quão aleatória é a saída de um modelo de linguagem. O modelo divide as pontuações dos tokens pela temperatura antes de transformá-las em probabilidades: valores baixos fazem o token mais provável dominar; valores altos achatam a distribuição, e tokens menos prováveis são escolhidos com mais frequência. - [Text-to-speech (TTS, síntese de voz)](https://broadrouter.com/pt/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): Text-to-speech (TTS), ou síntese de voz, converte texto escrito em áudio falado. Uma API de TTS recebe um texto, uma voz e um formato de saída, como MP3 ou WAV, e devolve a fala gerada. É o inverso do speech-to-text, que transforma áudio em texto. - [Throughput (vazão)](https://broadrouter.com/pt/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): Throughput, ou vazão, no serving de LLMs, é a quantidade total de trabalho que um sistema conclui por unidade de tempo em todas as requisições simultâneas, geralmente medida em tokens de saída ou requisições por segundo. É diferente da velocidade por stream, que é a rapidez com que uma única requisição recebe seus tokens. - [Time to first token (TTFT)](https://broadrouter.com/pt/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): Time to first token (TTFT) é o tempo entre o envio de uma requisição a um modelo de linguagem e a chegada do primeiro token de saída. Inclui o tempo de rede, a fila no servidor e o prefill, etapa em que o modelo lê o prompt inteiro, então cresce com o tamanho do prompt e a carga. - [Timestamps (speech-to-text)](https://broadrouter.com/pt/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): Timestamps em speech-to-text são tempos de início e fim, em segundos, associados a partes de uma transcrição. Timestamps de segmento cobrem uma frase ou trecho; timestamps por palavra cobrem cada palavra. Eles são necessários para legendas, busca dentro de gravações e alinhamento do texto com os locutores. - [Token](https://broadrouter.com/pt/glossary/token/) (translation of https://broadrouter.com/glossary/token/): Um token é a unidade de texto que um modelo de linguagem lê e escreve: uma palavra, parte de uma palavra, um sinal de pontuação ou uma sequência de bytes gerada pelo tokenizador do modelo. As APIs de LLM contam, limitam e normalmente cobram o uso em tokens. - [Tokenizer (tokenizador)](https://broadrouter.com/pt/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): Tokenizer (tokenizador) é o componente que converte texto na sequência de IDs inteiros de tokens que um modelo de linguagem lê, e converte os IDs gerados de volta em texto. Cada modelo vem com o próprio tokenizer, então o mesmo texto pode gerar um número diferente de tokens em modelos diferentes. - [Tokens por segundo (tokens per second)](https://broadrouter.com/pt/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): Tokens por segundo (TPS) são a taxa com que um modelo de linguagem gera tokens de saída, geralmente medida para uma requisição depois que o primeiro token chega. Definem a rapidez com que a resposta flui depois de começar. Somados em todas as requisições simultâneas, viram throughput do sistema; a espera pelo primeiro token é medida à parte. - [Tool calling (chamada de ferramentas)](https://broadrouter.com/pt/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): Tool calling, também chamado de function calling ou chamada de ferramentas, é um recurso de APIs de LLM em que o modelo devolve um pedido estruturado para executar uma função definida pelo desenvolvedor, com argumentos JSON. A aplicação executa a função e devolve o resultado; o modelo nunca executa código sozinho. - [tool_choice](https://broadrouter.com/pt/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice é um parâmetro da requisição que controla o uso de ferramentas. No Chat Completions da OpenAI, pode ser auto (o modelo decide), required (precisa chamar uma ou mais ferramentas), none (sem ferramentas) ou uma função específica a forçar. As opções equivalentes da Anthropic são auto, any, tool e none. - [Top-p (nucleus sampling)](https://broadrouter.com/pt/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Top-p, também chamado de nucleus sampling, é um método de decodificação em que o modelo de linguagem escolhe o próximo token apenas entre o menor conjunto de tokens mais prováveis cuja probabilidade somada atinge um limiar p, como 0,9, descartando a cauda improvável. - [Normalização Unicode (Unicode normalization)](https://broadrouter.com/pt/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Normalização Unicode é a conversão de texto para uma forma padrão, para que strings que parecem iguais usem os mesmos code points. As quatro formas são NFD, NFC, NFKD e NFKC; as formas K também reduzem caracteres de compatibilidade, como formas de apresentação árabes e ligaduras, a letras comuns. - [Voice activity detection (VAD)](https://broadrouter.com/pt/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): Voice activity detection (VAD, detecção de atividade de voz) é uma etapa de processamento de sinal que classifica frames curtos de áudio como fala ou não fala. Sistemas de voz a usam para pular silêncios, dividir gravações longas nas pausas, detectar quando alguém terminou de falar e evitar transcrever ruído de fundo. - [Banco de dados vetorial (vector database)](https://broadrouter.com/pt/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): Um banco de dados vetorial é um armazenamento que guarda vetores de embedding, muitas vezes com o texto de origem e metadados, e responde a consultas de vizinhos mais próximos: dado um vetor de consulta, devolve os vetores armazenados mais parecidos, em geral por distância de cosseno, distância euclidiana ou produto interno. - [Modelo de visão e linguagem (VLM)](https://broadrouter.com/pt/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): Um modelo de visão e linguagem (VLM, vision-language model) é um modelo que recebe imagens e texto juntos como entrada e produz texto, então consegue descrever fotos, responder perguntas sobre elas e ler documentos, gráficos ou capturas de tela. VLMs também são chamados de modelos image-text-to-text ou multimodais. - [vLLM](https://broadrouter.com/pt/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM é uma biblioteca open source de inferência e serving de LLMs, desenvolvida originalmente no Sky Computing Lab da UC Berkeley. Ela gerencia a memória do KV cache com PagedAttention, agrupa as requisições em batching contínuo, suporta muitos formatos de quantização e expõe um servidor HTTP compatível com OpenAI. - [Tamanho do vocabulário (vocabulary size)](https://broadrouter.com/pt/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): Tamanho do vocabulário é o número de tokens distintos no tokenizer de um modelo de linguagem, cada um com o seu ID e o seu embedding. Um vocabulário maior pode codificar texto comum em menos tokens, mas aumenta as camadas de embedding e de saída, e só ajuda no tipo de texto com que foi treinado. - [WAV (arquivo de áudio WAV)](https://broadrouter.com/pt/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format, ou WAVE) é um contêiner de áudio baseado no formato RIFF da Microsoft e da IBM. Ele costuma guardar amostras PCM lineares sem compressão, então o tamanho é previsível: taxa de amostragem vezes bytes por amostra vezes canais, por segundo. - [Word error rate (WER)](https://broadrouter.com/pt/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): Word error rate (WER, taxa de erro de palavras) é a medida padrão de precisão do reconhecimento de fala: o número de substituições, deleções e inserções de palavras necessárias para transformar uma transcrição na referência correta, dividido pelo número de palavras da referência. Quanto menor, melhor. - [Retenção zero de dados (zero data retention)](https://broadrouter.com/pt/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): Retenção zero de dados (ZDR) é uma política de tratamento de dados em que um provedor de IA processa prompts e saídas só para atender uma requisição e não guarda esse conteúdo depois, nem em logs, nem em caches, nem em dados de treinamento. Registros sem conteúdo, como contagens de tokens e tempos, costumam continuar sendo guardados. ### French (fr, Français): 294 translated pages - [Articles in French](https://broadrouter.com/fr/blog/) - [Prix des API LLM en octobre 2026 : 3 hausses, aucune baisse](https://broadrouter.com/fr/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 changements sur les pages de tarifs de 5 des 7 fournisseurs d’API LLM en octobre 2026 : 3 hausses, aucune baisse, 9 nouveaux prix et 11 retraits. Liens d’archive. - [Coût d’une exécution d’agent de code IA : étapes, contexte, cache](https://broadrouter.com/fr/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): Pourquoi chaque étape d’un agent de code coûte plus cher à mesure qu’il avance, l’effet du cache sur le total et quand une session au forfait gagne. Calcul détaillé. - [Vocaliser un texte arabe avec un LLM : prompt, contrôles, DER](https://broadrouter.com/fr/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): Restaurer les diacritiques arabes (tachkil) via une API LLM : un prompt qui garde chaque lettre, un contrôle anti-retouches, DER et WER, et quand choisir CATT. - [Correction grammaticale et orthographique de l’arabe par LLM : guide](https://broadrouter.com/fr/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Corriger l’orthographe et la grammaire arabes via une API LLM : prompt à modifications minimales, garde-fou anti-réécriture, éditions par type d’erreur, score M2. - [Comptes rendus de réunion en arabe depuis l’audio : un pipeline API](https://broadrouter.com/fr/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Transformer des enregistrements de réunions en arabe en comptes rendus : transcrire, nettoyer, résumer en JSON avec un LLM, puis vérifier noms et dates. - [Extraction d’entités nommées en arabe par LLM : JSON, dates, contrôles](https://broadrouter.com/fr/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Extraire noms, organisations, lieux et dates hégiriennes arabes en JSON par LLM : prompt, contrôle sans diacritiques, conversion des dates et script d’évaluation. - [Comparatif des API OCR pour l’arabe : Google, Azure, Textract, Mistral](https://broadrouter.com/fr/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): Quelles API OCR prennent officiellement en charge l’arabe imprimé et manuscrit, ce qu’elles renvoient et leur coût pour 1 000 pages, face aux API de LLM vision. - [Sous-titres arabes (SRT) depuis une API de transcription : l’essentiel](https://broadrouter.com/fr/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): Ce qu’exige un fichier SRT, quelles API de transcription renvoient horodatages ou SRT pour l’arabe, et un pipeline testé de sous-titres depuis une API texte. - [Classification de texte arabe par LLM : sentiment, intention, tri](https://broadrouter.com/fr/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): Classer avis, messages et tickets en arabe avec une API LLM : conception des labels, prompt adapté aux dialectes, sortie JSON, test en F1 macro, coût en volume. - [Normaliser le texte arabe pour la recherche et le RAG : quoi changer](https://broadrouter.com/fr/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Tachkil, alif, taa marbouta, supports de hamza, lettres persanes, chiffres, Unicode : quelles règles aident la recherche, lesquelles nuisent, Python testé. - [Résumé de texte arabe par LLM : prompts, longueur et évaluation](https://broadrouter.com/fr/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): Résumer articles, rapports et messages en dialecte arabe via une API LLM : prompts, contrôle du nombre de mots, vérification des chiffres, et pourquoi ROUGE donne 0. - [Alternatives à Artificial Analysis : où comparer les LLM en 2026](https://broadrouter.com/fr/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Les alternatives à Artificial Analysis comparées : Arena, Epoch AI, Vals AI, classements OpenRouter, Hugging Face et données de prix, selon ce que chacune mesure. - [Matériel GPU et serving de BroadNet, tels que BroadNet les décrit](https://broadrouter.com/fr/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): Ce que BroadNet dit du matériel derrière BroadNet Inference : trois niveaux de GPU, du B300 aux appliances client, débit par GPU et leçons tirées de ses tests. - [Mettre en cache les réponses d’API LLM : un cache exact en Python](https://broadrouter.com/fr/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Un cache exact testé pour les réponses d’API LLM en Python : la clé, ce qu’il ne faut jamais cacher, TTL, confidentialité et mesure du taux de succès avant de coder. - [Claude Code et modèles ouverts : proxys, endpoints natifs, limites](https://broadrouter.com/fr/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code parle l’API Messages d’Anthropic. Comment le faire tourner sur des modèles ouverts via un proxy LiteLLM ou un endpoint natif, et ce qui casse en route. - [Intégrer la relecture humaine à votre workflow d’IA documentaire](https://broadrouter.com/fr/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): L’étape de relecture humaine d’un pipeline d’IA documentaire : règles de validation, ce qu’un humain voit, audits aléatoires, piste d’audit et taux de relecture. - [La loi égyptienne sur les données (loi 151 de 2020) et les API LLM](https://broadrouter.com/fr/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Envoyer des données personnelles d’Égyptiens à une API LLM ou speech-to-text à l’étranger : loi 151/2020 et décision 816/2025 sur transferts, licences, violations. - [Avant le fine-tuning : définir le problème à résoudre](https://broadrouter.com/fr/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): Quand prompting, RAG ou évaluation suffisent sans fine-tuning, quelles données d’entraînement réunir et comment tester un fine-tune avant de le demander. - [Corriger le JSON invalide d’un LLM : causes, réparation, relance](https://broadrouter.com/fr/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): Pourquoi les LLM renvoient du JSON cassé (balises de code, texte, coupure à max_tokens, virgules finales) et une fonction Python testée qui ne répare que le sûr. - [Les prix des API LLM changent-ils souvent ? 3 ans d’archives](https://broadrouter.com/fr/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): 308 pages de tarifs archivées de 11 fournisseurs d’API LLM, 2023 à 2026 : 80 % des modèles n’ont jamais changé de prix ; DeepInfra a fait 75 % des changements. - [Réduire les coûts d’API LLM : une checklist en 10 étapes, dans l’ordre](https://broadrouter.com/fr/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): Réduire une facture d’API LLM dans le bon ordre : mesurer, mettre le préfixe en cache, alléger les prompts, plafonner sortie et raisonnement, batch, forfait. - [Failover d’API LLM : un fournisseur de secours dans votre propre code](https://broadrouter.com/fr/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): Basculer entre fournisseurs LLM compatibles OpenAI : sur quelles erreurs, disjoncteur, correspondance des ID de modèle, streaming, coût et confidentialité. - [Baisses de prix des API LLM 2023–2026 : les plus fortes, sourcées](https://broadrouter.com/fr/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): Les plus fortes baisses de prix des API LLM de 2023 à 2026, avec liens d’archive avant et après : o3, Codestral, Gemini 1.5, et les hausses passées inaperçues. - [Les prix des API LLM ne font-ils que baisser ? 52 hausses à budgéter](https://broadrouter.com/fr/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): Les prix des API LLM augmentent aussi : 52 hausses du prix d’un même modèle dans des pages archivées, 2024 à 2026, médiane 40 %. Sources et règle de budget. - [Limites de débit des API LLM par fournisseur : RPM, TPM, simultanéité](https://broadrouter.com/fr/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): Limites de débit d’entrée de gamme de 13 API LLM, lues dans leur doc : OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks, etc., et ce que signifient RPM et TPM. - [SLA des API LLM : ce que promettent Azure, Bedrock, Google et OpenAI](https://broadrouter.com/fr/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): Ce que chaque fournisseur d’API LLM garantit par écrit : disponibilité, définition de la panne, crédits de service et objectifs de latence, lus sur les pages SLA. - [Traitement par lots LLM sans Batch API : un exécuteur reprenable](https://broadrouter.com/fr/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Passer 10 000 à 1 000 000 de prompts dans une API chat completions sans Batch API : concurrence, cadence, nouvelles tentatives, points de reprise et calcul du temps. - [La Safety Map : comment 39 LLM traitent les données sensibles](https://broadrouter.com/fr/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet a étudié 39 LLM de 14 labos sur identifiants, PII, données de santé et financières : ce qui se transfère ou non, et les fuites par outils et raisonnement. - [Tester un LLM sur les données sensibles : la méthode de BroadNet](https://broadrouter.com/fr/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): La méthode de BroadNet pour tester avant déploiement si un LLM divulgue identifiants, PII, santé ou finance : catégories, fuite ou oubli, au moins 3 exécutions. - [Quitter l’API OpenAI : migration pas à pas en Python, Node et cURL](https://broadrouter.com/fr/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): Le code à modifier pour passer d’OpenAI à une API compatible OpenAI : client, Responses vers Chat Completions, outils, streaming, embeddings et test de fumée. - [Les modèles que BroadNet a évalués puis écartés, et pourquoi](https://broadrouter.com/fr/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet publie les modèles testés et rejetés, avec leurs raisons : modèles de chat recalés au test de sécurité des identifiants, modèles vocaux, OCR, embeddings. - [Concevoir une expérience IA multilingue : arabe, anglais, espagnol](https://broadrouter.com/fr/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): Concevoir un produit IA pour l’arabe, l’anglais et l’espagnol : locale de l’interface ou langue du modèle, RTL et texte mixte, nombres, dates, détection et coût. - [Coût API : modèles open-weight ou fermés, ce que disent les prix](https://broadrouter.com/fr/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): Les modèles open-weight sont-ils moins chers via API que GPT, Claude ou Gemini ? Prix au token des fournisseurs, écarts entre hébergeurs et coût par utilisateur. - [Historique des prix des API OpenAI, Anthropic et Google, 2023–2026](https://broadrouter.com/fr/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): Les prix des API GPT, Claude et Gemini côte à côte, d’après des pages archivées de 2023 à 2026 : quelles gammes ont baissé, lesquelles ont augmenté, liens à l’appui. - [URL de base des API compatibles OpenAI : 14 fournisseurs vérifiés](https://broadrouter.com/fr/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): La base_url de 14 API LLM compatibles OpenAI, relevée dans la doc de chaque fournisseur, avec les endpoints listés et les pièges qui cassent un changement direct. - [OpenClaw et Hermes Agent sur un modèle bon marché : config et coûts](https://broadrouter.com/fr/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): Brancher OpenClaw ou Hermes Agent sur une API compatible OpenAI, comprendre pourquoi les tâches planifiées brûlent des tokens, et réduire la facture. - [Un endpoint privé commence par une charge de travail bien définie](https://broadrouter.com/fr/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): Ce qu’il faut définir avant de demander un endpoint LLM privé : charge, tokens par seconde, concurrence, contexte, région, licence et budget, puis comment le tester. - [Injection de prompt et agents de code : des défenses qui tiennent](https://broadrouter.com/fr/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): Par où l’injection de prompt atteint un agent de code, et les défenses en couches que documentent éditeurs et chercheurs : sandbox, réseau, listes, validation. - [Taille des chunks et chevauchement en RAG : effets sur rappel et coût](https://broadrouter.com/fr/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): Comment la taille des chunks et le chevauchement changent le rappel, l’index, le coût d’embedding et les tokens envoyés au LLM, avec résultats publiés et calculs. - [Régler max_tokens pour les modèles de raisonnement : une méthode](https://broadrouter.com/fr/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Pourquoi un modèle de raisonnement renvoie une réponse vide avec un max_tokens bas, et comment fixer le budget : percentiles, fenêtre de contexte, relance. - [Enregistrer l’audio du navigateur, l’envoyer à une API speech-to-text](https://broadrouter.com/fr/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): Enregistrer le micro avec MediaRecorder, convertir le WebM ou le MP4 de Safari en WAV 16 kHz et l’envoyer à une API de transcription via un petit serveur. - [Réduire les tokens du prompt : prompts, outils et données mesurés](https://broadrouter.com/fr/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): Tokens gagnés en allégeant prompt système, exemples few-shot, schémas d’outils, format des données, historique et diacritiques arabes, mesurés sur vrais tokenizers. - [Même modèle, hébergeur différent : l’écart des prix des API LLM](https://broadrouter.com/fr/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): Pour 119 modèles vendus par plusieurs hébergeurs, le prix catalogue le plus cher vaut en médiane 1,76 fois le moins cher. Tarifs de 41 fournisseurs, novembre 2026. - [PDPL saoudienne et émiratie et API LLM : checklist des transferts](https://broadrouter.com/fr/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Envoyer des données personnelles à une API LLM depuis l’Arabie saoudite ou les EAU : règles de transfert de la PDPL, questions au fournisseur, inférence locale. - [Texte, voix ou vision ? Partez de la tâche](https://broadrouter.com/fr/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): Chat texte, entrée image ou transcription audio : choisir selon la tâche, avec les modèles et limites que BroadNet documente, et ce qu’il n’offre pas, comme le TTS. - [Transcrire un audio de plus de 25 Mo : découper, envoyer, recoller](https://broadrouter.com/fr/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): Transcrire des enregistrements au-delà de la limite de fichier d’une API speech-to-text : réduire le fichier, découper aux silences avec ffmpeg, envoyer, recoller. - [Traduire les chaînes d’app en arabe par LLM : variables et pluriels](https://broadrouter.com/fr/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): Traduire des fichiers JSON i18n en arabe via une API LLM : figer les variables, générer les six pluriels arabes, valider l’ICU en CI et isoler les noms insérés. - [Traduire un audio arabe en texte anglais par API : un pipeline testé](https://broadrouter.com/fr/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): Passer d’enregistrements arabes à un texte anglais en deux appels : transcrire avec language=ar, traduire avec un LLM, puis chiffres, découpage, coût et score. - [En ligne, sur demande, recherche : bien lire un catalogue de modèles](https://broadrouter.com/fr/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): Quels modèles d’un catalogue sont appelables aujourd’hui, lesquels sont sur demande et lesquels ont seulement été testés : l’exemple de BroadNet et /v1/models. - [Uno dans vLLM : ce qu’a mesuré l’étude de service de BroadNet](https://broadrouter.com/fr/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet a intégré l’adaptateur d’auto-drafting Uno à vLLM et l’a mesuré sur H100 : débit 2,58x à un flux, 1,74x à 32. La méthode, le DOI et les limites. - [Uno sur Gemma 4 26B : l’étude mixture of experts de BroadNet](https://broadrouter.com/fr/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet a entraîné le premier adaptateur Uno d’auto-drafting pour un modèle mixture of experts, Gemma 4 26B A4B. Mesures, intérêt de l’auto-drafting et DOI. - [Budget de tokens d’un agent : où partent les tokens d’une boucle](https://broadrouter.com/fr/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): Prompt système, schémas d’outils, lectures de fichiers, historique et raisonnement : où partent les tokens d’un agent, comment mesurer chaque poste et quoi réduire. - [Taux d’erreur de mots (WER) en reconnaissance vocale arabe](https://broadrouter.com/fr/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): Ce que mesure le taux d’erreur de mots, comment le calculer et pourquoi le WER arabe réagit tant aux diacritiques, à l’orthographe et aux dialectes. Avec Python. - [Reconnaissance vocale arabe : Golfe, Égypte, Levant, Maghreb](https://broadrouter.com/fr/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Pourquoi le speech-to-text peine sur l’arabe du Golfe, égyptien, levantin et maghrébin, et comment un jeu de test par dialecte révèle le niveau d’un fournisseur. - [Traduction arabe–anglais par LLM : prompts, glossaires, évaluation](https://broadrouter.com/fr/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): Traduire de l’arabe vers l’anglais et inversement via une API LLM : prompts qui gardent le ton, glossaires de termes fixes, nombres et noms, évaluation. - [Arabe du Golfe, égyptien et levantin : les LLM face aux dialectes](https://broadrouter.com/fr/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): Comment les LLM gèrent l’arabe du Golfe, égyptien et levantin : où ils échouent, comment prompter et tester un dialecte, et ce que BroadNet annonce pour Qwen3.8 27B. - [OCR arabe ou LLM de vision : lire documents et factures en arabe](https://broadrouter.com/fr/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): Quand choisir l’OCR classique ou un LLM de vision pour factures, reçus et formulaires arabes : pièges, sortie JSON, coût par page et pipeline de validation. - [Prompt engineering en arabe : registre, RTL, chiffres et texte mixte](https://broadrouter.com/fr/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): Des prompts qui obtiennent un bon arabe d’un LLM : registre et dialecte, chiffres et dates maîtrisés, mise en forme RTL propre, mélange arabe–anglais maîtrisé. - [RAG en arabe : normalisation, chunking et recherche qui fonctionnent](https://broadrouter.com/fr/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): Le RAG sur des documents arabes : normaliser alef, ya et ta marbuta, découper avec bon sens, choisir des embeddings multilingues, ajouter la recherche par mots-clés. - [API speech-to-text pour l’arabe : comment la choisir en 2026](https://broadrouter.com/fr/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): Choisir une API de transcription arabe : WER, dialectes, limites de fichiers, prix par heure d’audio, confidentialité, et un plan de test sur vos enregistrements. - [Chatbot WhatsApp en arabe : du message vocal à la réponse](https://broadrouter.com/fr/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): Une architecture éprouvée pour un assistant WhatsApp en arabe : webhook, file, transcription des vocaux, réponse LLM, règles et coûts, avec du Python à adapter. - [L’arabizi et les LLM : traiter le franco-arabe et ses chiffres 3 et 7](https://broadrouter.com/fr/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): L’arabizi écrit l’arabe en lettres latines, avec 3 pour ع et 7 pour ح. Comment le détecter, le convertir en écriture arabe avec un LLM et bien répondre. - [Meilleur LLM pour Aider : choisir un modèle et le connecter](https://broadrouter.com/fr/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Choisir le meilleur LLM pour Aider : rigueur du format d’édition, contexte, coût horaire, un test d’une heure sur votre dépôt et la configuration OPENAI_API_BASE. - [Meilleur modèle ouvert pour agents de code : outils, contexte, coût](https://broadrouter.com/fr/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B ou Gemma 4 pour votre agent de code ? Choisissez selon l’appel d’outils, le contexte, le raisonnement et le coût par tâche. - [Tokens d’entrée en cache : réduire la facture d’une API LLM](https://broadrouter.com/fr/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): L’entrée en cache coûte une fraction de l’entrée normale. Facturation du prompt caching, lecture de cached_tokens dans usage, et prompts structurés pour en profiter. - [API LLM de vision la moins chère : coût par image et par page, calculé](https://broadrouter.com/fr/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): Coût d’un LLM de vision par image et par page : calcul en tokens pour Gemma 4, Qwen3.8 27B et DeepSeek V4.1 Flash, et où DeepInfra est moins cher. - [Comment choisir un fournisseur d’API LLM : checklist en 20 points](https://broadrouter.com/fr/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): Vingt questions avant de choisir une API LLM : prix, limites, compatibilité OpenAI, confidentialité, modèles, support, disponibilité, grille et plan de test. - [Choisir une API LLM pour l’arabe : ce qu’il faut tester](https://broadrouter.com/fr/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): Choisir un LLM pour une appli en arabe : compréhension des dialectes, qualité rédactionnelle, traduction, documents et images, coût, et un jeu de test en un jour. - [Cline, Roo Code ou Aider pour une équipe sur modèles ouverts (2026)](https://broadrouter.com/fr/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (archivé, poursuivi par Zoo Code) et Aider comparés sur ce qui compte avec des modèles ouverts : appels d’outils, contexte, MCP, git et validations. - [Compter les tokens avant d’envoyer une requête LLM](https://broadrouter.com/fr/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): Estimer les tokens d’un prompt par langue et modèle avant d’appeler une API LLM : écarts entre tokenizers, mesures en cinq langues et marges sûres pour max_tokens. - [Réduire le coût LLM d’un agent de code : 10 leviers chiffrés](https://broadrouter.com/fr/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Dix façons concrètes de réduire ce que Cline, Aider, opencode ou votre propre agent de code dépense en tokens, avec le calcul de l’économie et sa mesure. - [Alternatives à l’API DeepSeek : 5 façons d’utiliser DeepSeek](https://broadrouter.com/fr/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): DeepSeek via l’API officielle, un agrégateur, un catalogue cloud, vos GPU ou BroadNet Inference : compromis sur le prix, la confidentialité et les limites. - [Prix de l’API DeepSeek V4.1 Flash : au token, en cache et forfaits](https://broadrouter.com/fr/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): Prix de l’API DeepSeek V4.1 Flash par million de tokens (entrée, cache, sortie), forfait au mois, à la semaine ou à l’heure, face à DeepSeek, DeepInfra et Fireworks. - [Estimer le coût d’une API LLM avant de lancer une fonctionnalité](https://broadrouter.com/fr/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): Modèle de coût pas à pas d’une fonctionnalité IA : utilisateurs × actions × appels × tokens, avec cache, raisonnement, relances, limites et un exemple chiffré. - [Comment évaluer un LLM avant de changer de fournisseur](https://broadrouter.com/fr/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): Construisez un petit jeu d’évaluation à partir du trafic réel, notez-le avec des tests en code et un LLM juge débiaisé, puis comparez qualité, coût et latence. - [Extraire factures, reçus et formulaires en JSON avec un LLM vision](https://broadrouter.com/fr/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): Un pipeline de production pour factures, reçus et formulaires : préparation des images, prompt JSON, validation Pydantic, contrôles de calcul, relances, revue. - [API LLM : forfait ou paiement au token ? Le calcul de rentabilité](https://broadrouter.com/fr/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): Quand un forfait LLM bat-il le paiement au token ? Quatre factures mensuelles pour DeepSeek V4.1 Flash, le volume de rentabilité et comment mesurer le vôtre. - [Crédit gratuit pour API LLM : jusqu’où va vraiment un crédit de départ](https://broadrouter.com/fr/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): Ce que permet vraiment un crédit gratuit d’API LLM : tokens, minutes d’audio et clés à l’heure avec le crédit de départ de BroadNet, face à Gemini et OpenRouter. - [RGPD et API LLM : 15 questions à poser à votre fournisseur](https://broadrouter.com/fr/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): Checklist RGPD pour les API LLM : DPA, sous-traitants ultérieurs, conservation, entraînement, transferts hors UE et notification des violations, avec les articles. - [GLM 5.3 ou DeepSeek V4.1 Flash : lequel utiliser, et quand ?](https://broadrouter.com/fr/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 vs DeepSeek V4.1 Flash : prix, contexte, outils, sortie structurée, images et raisonnement, avec des coûts chiffrés et une règle simple pour choisir. - [Gérer les erreurs 429 d’une API LLM : backoff, jitter et limiteurs](https://broadrouter.com/fr/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 ou 5xx, backoff exponentiel avec full jitter, limiteurs de simultanéité et files d’attente pour les API LLM, avec du code Python et Node prêt pour la production. - [Clé d’API LLM horaire : un prix fixe à l’heure, expliqué](https://broadrouter.com/fr/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): La clé horaire de BroadNet : une heure de Gemma 4 VL 26B sur un flux à prix fixe. Ce qu’elle inclut, ses limites, à qui elle convient, quand elle est rentable. - [Conservation des données API LLM : OpenAI, Anthropic, Gemini, Mistral](https://broadrouter.com/fr/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): Ce qu’OpenAI, Anthropic, l’API Gemini, Mistral et BroadNet disent officiellement de la conservation des prompts d’API et de l’entraînement, cité en novembre 2026. - [Sécuriser une clé d’API LLM : stockage, rotation et listes d’IP](https://broadrouter.com/fr/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): Gardez les clés d’API LLM hors du code et des logs, une clé par usage, faites-les tourner, liez-les à vos IP serveur et révoquez une fuite en quelques secondes. - [Tarification des API LLM : entrée, cache, sortie et raisonnement](https://broadrouter.com/fr/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): Comment fonctionne le prix au token des API LLM : entrée, entrée en cache, sortie, tokens de raisonnement, images et audio à la minute, avec des exemples chiffrés. - [Données sensibles et API LLM : masquer, minimiser, moins journaliser](https://broadrouter.com/fr/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): Envoyer des données clients, de santé ou financières à une API LLM en sécurité : classer les champs, pseudonymiser de façon réversible, zéro rétention, logs propres. - [Maîtriser les dépenses d’API LLM : budgets, limites de clés, boucles](https://broadrouter.com/fr/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): Empêcher une facture d’API LLM de s’emballer : solde prépayé, une clé par application, listes d’IP autorisées, logs, budgets en code, plafonds et prix fixes. - [Journaliser l’usage d’une API LLM : coût par utilisateur, sans PII](https://broadrouter.com/fr/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): Journalisez chaque appel d’API LLM avec tokens, coût, latence et ID de requête, attribuez la dépense par utilisateur et fonctionnalité, masquez les PII. Code Python. - [Coût LLM par utilisateur : l’économie unitaire d’une fonction IA SaaS](https://broadrouter.com/fr/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): Calculez le coût d’une fonction IA par utilisateur actif et par mois, pourquoi les gros utilisateurs comptent plus que la moyenne, et comment fixer vos offres SaaS. - [Prix API des modèles ouverts : 6 fournisseurs d’inférence comparés](https://broadrouter.com/fr/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): Comment OpenRouter, Together AI, Fireworks, Groq, DeepInfra et BroadNet tarifent les modèles ouverts : au token, cache, batch, heures de GPU, forfaits, avec sources. - [Latence LLM : TTFT, tokens par seconde et temps de bout en bout](https://broadrouter.com/fr/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): Ce que mesurent le time to first token, les tokens par seconde et la latence de bout en bout, comment les chronométrer côté client, et quel levier agit sur quoi. - [Le streaming LLM expliqué : chunks SSE, deltas, usage et annulation](https://broadrouter.com/fr/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Le streaming des chat completions sur le réseau : rebâtir la réponse à partir des deltas, lire l’usage du dernier chunk, annuler proprement (Python, Node, curl). - [Contexte long ou RAG : quand 1M de tokens remplace la base vectorielle](https://broadrouter.com/fr/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): Quand une fenêtre de contexte de 1M de tokens peut remplacer une base vectorielle, et quand le RAG l’emporte : coût par requête, latence, rappel et maintenance. - [Gérer le contexte des agents longue durée avec une fenêtre de 1M](https://broadrouter.com/fr/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): Des sessions d’agent longues, précises et abordables avec une fenêtre de 1M tokens : historique en ajout seul, prompts compatibles cache, compactage. - [Agents de longue durée : timeouts, retries et checkpoints qui tiennent](https://broadrouter.com/fr/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): Faire tourner un agent LLM des heures : timeout par appel, streaming contre les blocages, retries sûrs, effets de bord idempotents et checkpoint à chaque étape. - [Serveurs MCP avec des modèles compatibles OpenAI (tout fournisseur)](https://broadrouter.com/fr/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP marche avec tout LLM : le client transforme les outils MCP en appels de fonction. MCP avec des modèles ouverts dans Cline, Continue, opencode, n8n ou en Python. - [Quand une fenêtre de contexte de 1M de tokens vaut le coup (ou pas)](https://broadrouter.com/fr/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): Ce qui tient dans 1 048 576 tokens, ce que cela coûte sur DeepSeek V4.1 Flash, comment raisonnement et cache changent le calcul, et quand le RAG reste meilleur. - [Ce que coûte vraiment un système multi-agents : au token ou au forfait](https://broadrouter.com/fr/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): Un système multi-agents consomme plusieurs fois les tokens d’un agent seul. Forfait ou token : le calcul, les multiplicateurs cachés et quand chacun l’emporte. - [API speech-to-text multilingue : 14 langues, un seul endpoint](https://broadrouter.com/fr/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): Transcrire anglais, espagnol, français, chinois, arabe et 9 autres langues via un endpoint compatible OpenAI facturé à la minute ; quand choisir le modèle arabe. - [Agent IA n8n sur un LLM bon marché : configuration, modèle, coût](https://broadrouter.com/fr/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): Workflows AI Agent n8n sur un modèle compatible OpenAI peu cher : quel modèle pour les outils, les réglages qui économisent des tokens, coût pour 10 000 exécutions. - [IA open-weight ou fermée en entreprise : coût, contrôle, risque](https://broadrouter.com/fr/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): LLM open-weight ou fermés pour votre entreprise ? Coût, contrôle, confidentialité, écart de qualité et dépendance, avec trois façons d’exécuter les modèles ouverts. - [Passer à une API compatible OpenAI : la checklist de migration](https://broadrouter.com/fr/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): Changer base_url, c’est facile. Ce qui casse : ID de modèles, tokens de raisonnement, usage en streaming, erreurs, limites de débit et clés. Checklist avec code. - [OpenRouter ou forfait pour vos agents IA : lequel coûte moins cher ?](https://broadrouter.com/fr/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter facture au token plus des frais sur les crédits ; un forfait, à la session. Le calcul pour un agent permanent, les frais, les limites, et quoi choisir. - [Laisser tourner des agents de code toute la nuit, en toute sécurité](https://broadrouter.com/fr/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): Une configuration pratique pour faire tourner un agent de code sans surveillance : sandbox, points de contrôle git, fichiers de tâche, limites, relecture et forfait. - [Structurer ses prompts pour que le préfixe soit mis en cache](https://broadrouter.com/fr/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): Organiser un prompt LLM pour garder le préfixe en cache identique : prompt système, schémas d’outils, documents, historique en ajout seul, et un test qui le prouve. - [Tokens de raisonnement : pourquoi vous payez plus de sortie que prévu](https://broadrouter.com/fr/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): Les modèles de raisonnement facturent leur réflexion cachée en sortie. La mesurer, son coût par modèle, la réduire avec reasoning_effort, max_tokens et le routage. - [Lancer des agents de code en parallèle sur une API LLM au forfait](https://broadrouter.com/fr/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Sous-agents, orchestrateurs et workers batch sur DeepSeek V4.1 Flash illimité : combien de sessions, comment les répartir entre clés, limiter le chevauchement. - [Auto-héberger un LLM ou passer par une API : coût réel et rentabilité](https://broadrouter.com/fr/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): Prix de location de GPU (pages officielles), coût par million de tokens sur votre GPU, taux d’utilisation pour battre une API et coûts d’exploitation oubliés. - [Prix des API speech-to-text comparés : tarif à la minute en 2026](https://broadrouter.com/fr/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): Prix à la minute et pour 1 000 heures d’OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq et BroadNet Transcribe, d’après leurs pages, options comprises. - [Mode JSON ou json_schema : une sortie structurée qui tient](https://broadrouter.com/fr/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): Le mode JSON donne du JSON valide ; json_schema donne votre structure. Les demander sur des modèles compatibles OpenAI, valider avec Pydantic et réparer en boucle. - [Tokens par mot, arabe ou anglais : pourquoi l’arabe coûte plus cher](https://broadrouter.com/fr/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): Pourquoi l’arabe prend souvent plus de tokens que l’anglais pour un même sens, pourquoi les tokens par mot trompent, et un script qui mesure votre ratio via usage. - [Appel d’outils et modèles ouverts : schémas, parallélisme, relances](https://broadrouter.com/fr/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): Function calling sur modèles ouverts compatibles OpenAI : schémas, appels parallèles, les cinq échecs typiques et une boucle Python qui valide et relance. - [Transcrire les enregistrements d’un centre d’appels, arabe compris](https://broadrouter.com/fr/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): Séparer les canaux stéréo, découper les tours de parole, transcrire l’arabe et l’anglais, masquer les PII et résumer chaque appel par LLM. Pipeline Python et coûts. - [Modèles ouverts et agents de code : Aider, Cline, opencode, Codex](https://broadrouter.com/fr/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Connectez Aider, Cline, Roo Code, opencode, Continue et Codex CLI à un endpoint compatible OpenAI : les noms de champs exacts de chacun et les réglages qui cassent. - [Messages vocaux en texte : de WhatsApp et Telegram aux tickets](https://broadrouter.com/fr/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): Transformer les messages vocaux de WhatsApp Business et des bots Telegram en transcriptions et tickets : télécharger, transcrire, extraire par LLM, router. - [Forfait LLM à la semaine pour projets courts : Orbit, heure ou mois](https://broadrouter.com/fr/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): Quand une semaine de DeepSeek V4.1 Flash illimité bat une session horaire, un forfait mensuel ou le paiement au token : calculs pour hackathons, sprints, migrations. - [Ce que « illimité » veut vraiment dire dans un forfait d’API LLM](https://broadrouter.com/fr/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Compteur de tokens, fenêtres de 5 h, plafonds hebdo, limites de débit, simultanéité : 7 questions pour savoir si un forfait IA « illimité » l’est pour votre agent. - [Alternative à l’API Whisper : quand changer et comment tester](https://broadrouter.com/fr/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): Quand une alternative hébergée à l’API Whisper d’OpenAI a du sens : taux d’erreur sur vos audios, langues, prix à l’heure, limites, confidentialité, code de test. - [API LLM à conservation zéro des données : checklist en 10 questions](https://broadrouter.com/fr/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): Ce que « zero data retention » devrait signifier pour une API LLM : prompts, sorties, logs, entraînement, sous-traitants, audits. Dix questions avant tout envoi. - [Guides in French](https://broadrouter.com/fr/guides/) - [Utiliser le SDK Python d’OpenAI avec BroadNet Inference](https://broadrouter.com/fr/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): Branchez le SDK Python officiel d’OpenAI sur BroadNet Inference en changeant la base_url : installation, clé, première requête et réponse en streaming. - [Utiliser le SDK Node.js d’OpenAI avec BroadNet Inference](https://broadrouter.com/fr/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): Branchez le SDK officiel d’OpenAI pour Node.js et TypeScript sur BroadNet Inference via baseURL : installation, première requête, streaming et ID de requête. - [Tester BroadNet Inference dans le navigateur, sans compte](https://broadrouter.com/fr/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): Testez le speech-to-text arabe, la lecture de documents et le chat Gemma 4 dans le playground de BroadNet, sans inscription ni carte, puis via l’API. - [Première requête à BroadNet Inference avec curl](https://broadrouter.com/fr/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): Premiers appels à BroadNet Inference avec curl : chat completion, réponse en streaming, transcription audio, puis lecture de la réponse, de l’usage et des erreurs. - [Streamer des chat completions et gérer les connexions coupées](https://broadrouter.com/fr/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): Le streaming sur BroadNet Inference : Server-Sent Events, le marqueur [DONE], l’usage en tokens dans le dernier chunk, et que faire quand un flux s’arrête trop tôt. - [Pourquoi un modèle de raisonnement renvoie une réponse vide](https://broadrouter.com/fr/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3 et Qwen3.8 27B réfléchissent avant de répondre, et ce raisonnement compte dans max_tokens. Comment éviter les réponses vides. - [Codes d’erreur de BroadNet Inference : quand réessayer ?](https://broadrouter.com/fr/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): Ce que signifie chaque statut HTTP de l’API BroadNet Inference, lesquels réessayer avec backoff, lesquels corriger d’abord, et quoi envoyer au support avec un ID. - [Limiter vos clés API à vos serveurs avec une liste d’IP autorisées](https://broadrouter.com/fr/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): Protégez vos clés BroadNet Inference avec une liste d’IP autorisées, séparez clés d’inférence et identifiants du portail, révoquez une fuite en secondes. - [Comment fonctionnent les forfaits API LLM illimités pour agents](https://broadrouter.com/fr/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): Les forfaits illimités DeepSeek V4.1 Flash n’ont ni compteur de tokens ni limite hebdomadaire. Sessions, agents parallèles, clés et facturation expliqués. - [Faire tourner Aider sur DeepSeek V4.1 Flash en illimité](https://broadrouter.com/fr/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Connectez l’agent de code Aider à BroadNet Inference via son réglage compatible OpenAI et utilisez DeepSeek V4.1 Flash en illimité, à prix forfaitaire. - [Utiliser LangChain avec BroadNet Inference](https://broadrouter.com/fr/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): Connectez ChatOpenAI de LangChain à BroadNet Inference avec base_url et api_key, et voyez ce que LangChain garde ou non des modèles de raisonnement. - [Utiliser Cline avec BroadNet Inference (fournisseur OpenAI Compatible)](https://broadrouter.com/fr/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Configurez l’extension VS Code Cline avec BroadNet Inference via le fournisseur OpenAI Compatible : champs exacts, ID de modèles, tailles de contexte et correctifs. - [Utiliser Roo Code avec BroadNet Inference (OpenAI Compatible)](https://broadrouter.com/fr/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code a fermé en mai 2026. Réglages pour les installations existantes et les forks, et où migrer (Cline ou le fork Zoo Code) avec BroadNet Inference. - [Utiliser Continue avec BroadNet Inference (config.yaml et apiBase)](https://broadrouter.com/fr/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): Ajoutez les modèles BroadNet Inference à l’extension Continue avec provider openai et apiBase dans config.yaml, capacités tool_use et longueur de contexte comprises. - [Ajouter BroadNet Inference à opencode comme fournisseur personnalisé](https://broadrouter.com/fr/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): Ajoutez BroadNet Inference à l’agent de code opencode comme fournisseur compatible OpenAI dans opencode.jsonc, avec alias de modèles, capacités et limites. - [Connecter Open WebUI à BroadNet Inference](https://broadrouter.com/fr/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Ajoutez BroadNet Inference à Open WebUI comme connexion compatible OpenAI, inscrivez les ID de modèles et corrigez vérification, réponses vides et 429. - [Ajouter BroadNet Inference à LibreChat comme custom endpoint](https://broadrouter.com/fr/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): Ajoutez BroadNet Inference à LibreChat comme custom endpoint compatible OpenAI dans librechat.yaml : clé dans .env, liste de modèles, modèle de titre et dépannage. - [Utiliser BroadNet Inference dans n8n (credential OpenAI, Base URL)](https://broadrouter.com/fr/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): Pointez le credential OpenAI de n8n vers BroadNet Inference avec le champ Base URL, puis utilisez le nœud OpenAI Chat Model par ID, Responses API désactivée. - [Utiliser le Vercel AI SDK avec BroadNet Inference](https://broadrouter.com/fr/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): Appelez BroadNet Inference depuis le Vercel AI SDK avec createOpenAICompatible (@ai-sdk/openai-compatible) : generateText, streamText, outils, raisonnement. - [API speech-to-text pour l’arabe, étape par étape](https://broadrouter.com/fr/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): Transcrivez de l’audio arabe avec BroadNet Transcribe Arabic via un endpoint compatible OpenAI : formats, champ language, exemples Python et curl, et prix. - [Router LiteLLM vers BroadNet Inference (SDK et proxy)](https://broadrouter.com/fr/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): Appelez BroadNet Inference via LiteLLM avec le préfixe openai/ et api_base, dans le SDK Python ou le config.yaml du proxy, avec des limites de débit adaptées. - [Faire tourner l’OpenAI Agents SDK sur BroadNet Inference](https://broadrouter.com/fr/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Utilisez l’OpenAI Agents SDK pour Python avec BroadNet Inference via AsyncOpenAI et OpenAIChatCompletionsModel, et coupez le tracing qui exige une clé OpenAI. - [Transcrire les messages vocaux WhatsApp en texte avec une API](https://broadrouter.com/fr/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): Convertissez les messages vocaux WhatsApp (.ogg) ou les fichiers .m4a en WAV avec ffmpeg, puis transcrivez-les en arabe ou en 13 autres langues par API. - [Transcrire les messages vocaux WhatsApp dans n8n](https://broadrouter.com/fr/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): Un workflow n8n qui reçoit un message vocal WhatsApp, le télécharge chez Meta, le convertit avec ffmpeg et le transcrit en arabe ou dans 13 autres langues. - [Lire des factures et reçus arabes avec un modèle de vision](https://broadrouter.com/fr/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Envoyez la photo d’une facture ou d’un reçu à un modèle de vision sur BroadNet Inference et récupérez total, date et fournisseur en JSON, en arabe ou en anglais. - [Utiliser Zed avec BroadNet Inference, fournisseur compatible OpenAI](https://broadrouter.com/fr/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Ajoutez BroadNet Inference à l’éditeur Zed comme fournisseur compatible OpenAI dans settings.json : ID de modèle, limites de tokens, capacités et variable de clé. - [Kilo Code avec BroadNet Inference (fournisseur OpenAI Compatible)](https://broadrouter.com/fr/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Connectez Kilo Code à BroadNet Inference comme fournisseur OpenAI Compatible, dans VS Code ou kilo.json, avec ID de modèle, limites et options d’outils. - [Ajouter BroadNet Inference à goose comme fournisseur personnalisé](https://broadrouter.com/fr/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): Connectez l’agent IA goose à BroadNet Inference comme fournisseur personnalisé OpenAI Compatible, depuis goose Desktop, goose configure ou un fichier JSON. - [Pydantic AI avec BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/fr/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Faites tourner des agents Pydantic AI sur BroadNet Inference avec OpenAIChatModel et OpenAICompatibleProvider : outils, sortie structurée, profils de modèle. - [CrewAI avec BroadNet Inference comme endpoint OpenAI personnalisé](https://broadrouter.com/fr/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): Branchez vos agents CrewAI sur BroadNet Inference avec LLM(custom_openai=True, base_url=...), puis choisissez les modèles pour outils, long contexte et forfait. - [Utiliser LlamaIndex avec BroadNet Inference via OpenAILike](https://broadrouter.com/fr/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): Connectez LlamaIndex à BroadNet Inference avec la classe OpenAILike, en réglant correctement api_base, context_window, is_chat_model et is_function_calling_model. - [Ajouter BroadNet Inference à Dify avec le plugin OpenAI-API-compatible](https://broadrouter.com/fr/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): Connectez Dify à BroadNet Inference via le plugin de fournisseur OpenAI-API-compatible : valeurs exactes des champs, tailles de contexte et function calling. - [Utiliser Jan avec BroadNet Inference, fournisseur compatible OpenAI](https://broadrouter.com/fr/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Ajoutez BroadNet Inference à l’application de bureau Jan comme fournisseur compatible OpenAI, saisissez les ID de modèle et activez outils et vision par modèle. - [Connecter SillyTavern à BroadNet Inference (Custom OpenAI-compatible)](https://broadrouter.com/fr/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): Connectez SillyTavern à BroadNet Inference via la source Custom (OpenAI-compatible) : bons ID de modèles, Prompt Post-Processing et limites de tokens. - [Utiliser Spring AI avec BroadNet Inference (base-url OpenAI)](https://broadrouter.com/fr/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Pointez le starter de chat OpenAI de Spring AI vers BroadNet Inference avec spring.ai.openai.base-url, un ID BroadNet et max-tokens, puis appelez ChatClient. - [Appeler BroadNet Inference depuis Go avec le SDK openai-go](https://broadrouter.com/fr/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): Utilisez le SDK officiel openai-go v3 avec BroadNet Inference via option.WithBaseURL : chat completions, streaming avec usage et gestion des erreurs en Go. - [Créer des agents LangGraph sur BroadNet Inference](https://broadrouter.com/fr/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): Faites tourner un StateGraph LangGraph ou create_agent sur BroadNet Inference avec ChatOpenAI, base_url et un ID BroadNet, plus outils, streaming et mémoire. - [Utiliser les pipelines et agents Haystack avec BroadNet Inference](https://broadrouter.com/fr/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Pointez l’OpenAIChatGenerator de Haystack vers BroadNet Inference avec api_base_url, un ID BroadNet et un Secret : pipelines, outils, streaming et agents. - [Utiliser Semantic Kernel avec BroadNet Inference](https://broadrouter.com/fr/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Connectez Microsoft Semantic Kernel à BroadNet Inference : en Python via un client AsyncOpenAI, en C# via un endpoint personnalisé, pour chat, plugins et agents. - [Utiliser des agents Mastra avec BroadNet Inference](https://broadrouter.com/fr/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): Faites tourner des agents Mastra sur BroadNet Inference avec un objet model qui définit id, url et apiKey, y compris le préfixe broadnet/ doublé qu’exige l’ID. - [Obtenir une sortie structurée avec Instructor sur BroadNet Inference](https://broadrouter.com/fr/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Extrayez des objets Pydantic de BroadNet Inference avec Instructor, via from_openai ou from_provider et une base_url personnalisée, et choisissez le bon Mode. - [Utiliser DSPy avec BroadNet Inference (définir model_type="chat")](https://broadrouter.com/fr/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): Configurez dspy.LM pour BroadNet Inference avec le préfixe openai/, api_base et model_type="chat", car DSPy 3.5 envoie les modèles OpenAI vers la Responses API. - [Évaluer les modèles BroadNet avec promptfoo](https://broadrouter.com/fr/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): Lancez des evals promptfoo et comparez des modèles côte à côte sur BroadNet Inference avec le fournisseur openai:chat, apiBaseUrl, apiKeyEnvar et des graders. - [Faire tourner smolagents sur BroadNet Inference avec OpenAIModel](https://broadrouter.com/fr/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Connectez smolagents de Hugging Face à BroadNet Inference avec OpenAIModel et api_base : CodeAgent sur tout modèle, ToolCallingAgent sur GLM 5.3 et Qwen3.8 27B. - [Connecter Flowise à BroadNet Inference (OpenAI Custom Model)](https://broadrouter.com/fr/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Utilisez BroadNet Inference dans vos chatflows Flowise avec le nœud OpenAI Custom Model, un Base Path, un identifiant OpenAI API et un nom de modèle BroadNet saisi. - [Connecter Langflow à BroadNet Inference](https://broadrouter.com/fr/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Utilisez BroadNet Inference dans Langflow via le champ OpenAI API Base du composant OpenAI et un nom de modèle saisi, et pourquoi OpenAI Compatible ne le liste pas. - [Glossary entries in French](https://broadrouter.com/fr/glossary/) - [Agent IA (AI agent)](https://broadrouter.com/fr/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): Un agent IA est un système dans lequel un modèle de langage dirige lui-même ses étapes vers un objectif : il décide quels outils appeler, lit les résultats et recommence jusqu’à produire une réponse finale ou atteindre une limite. Un script fixe qui appelle un modèle à des moments prévus s’appelle plutôt un workflow. - [Clé d’API (API key)](https://broadrouter.com/fr/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): Une clé d’API est une chaîne secrète qui identifie et autorise un client qui appelle une API. Les API LLM l’attendent généralement dans un en-tête Authorization: Bearer, et quiconque détient la clé peut envoyer des requêtes facturées au compte qui la possède. - [Diacritiques arabes (tashkeel)](https://broadrouter.com/fr/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): Les diacritiques arabes, ou tashkeel, sont de petits signes placés au-dessus ou au-dessous des lettres pour noter les voyelles brèves, les consonnes doublées et l’absence de voyelle. Ils sont facultatifs : la plupart des textes arabes, presse comprise, les omettent, et lecteurs comme logiciels déduisent les voyelles du contexte. - [Identification des dialectes arabes](https://broadrouter.com/fr/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): L’identification des dialectes arabes consiste à déterminer dans quelle variété d’arabe est un texte ou un enregistrement, par exemple égyptien, du Golfe, levantin, maghrébin ou arabe standard moderne, au niveau de la région, du pays ou de la ville. Elle sert à router les messages, choisir les prompts et découper les jeux d’évaluation. - [Normalisation du texte arabe](https://broadrouter.com/fr/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): La normalisation du texte arabe est l’étape qui ramène les variantes orthographiques d’un même mot à une seule forme avant la comparaison : suppression des diacritiques et du tatweel, unification des formes d’alif, et souvent ى vers ي et ة vers ه. Elle s’applique aux clés de recherche et à l’évaluation, pas au texte affiché. - [Arabizi (franco-arabe)](https://broadrouter.com/fr/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): L’arabizi est de l’arabe écrit de façon informelle en lettres latines, avec des chiffres pour les sons arabes sans équivalent latin, comme 3 pour ع. Courant dans les messageries et sur les réseaux sociaux, il est surtout dialectal, n’a pas d’orthographe standard et se mêle souvent d’anglais ou de français. - [AWQ (activation-aware weight quantization)](https://broadrouter.com/fr/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): L’AWQ (activation-aware weight quantization) est une méthode post-entraînement qui compresse les poids d’un LLM en basse précision, généralement des entiers 4 bits. Elle utilise les statistiques d’activation d’un petit jeu de calibration pour repérer les canaux de poids les plus importants et les met à l’échelle pour qu’ils perdent moins de précision. Elle ne nécessite aucune rétropropagation. - [URL de base (base URL)](https://broadrouter.com/fr/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): L’URL de base est l’adresse racine fixe d’une API, comme https://api.example.com/v1, que le client place devant chaque chemin d’endpoint, comme /chat/completions. Dans les SDK compatibles OpenAI, changer l’URL de base, avec la clé d’API et l’ID du modèle, oriente le même code vers un autre fournisseur. - [API batch (traitement par lots)](https://broadrouter.com/fr/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Une API batch est un endpoint LLM asynchrone qui accepte de nombreuses requêtes d’un coup, généralement sous forme de fichier, les traite dans une fenêtre définie plutôt qu’immédiatement et renvoie les résultats ensemble. Des fournisseurs comme OpenAI et Anthropic facturent les requêtes batch avec une remise par rapport à leurs API standard. - [BM25](https://broadrouter.com/fr/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) est une fonction de classement pour la recherche par mots-clés qui note un document selon la fréquence des termes de la requête dans ce document, la rareté de ces termes dans la collection et la longueur du document. C’est la similarité par défaut de Lucene et d’Elasticsearch. - [Byte pair encoding (BPE)](https://broadrouter.com/fr/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Le byte pair encoding (BPE) est un algorithme de tokenisation en sous-mots qui part de caractères ou d’octets isolés et fusionne de façon répétée la paire adjacente la plus fréquente en une nouvelle entrée de vocabulaire, jusqu’à atteindre une taille de vocabulaire cible. La plupart des tokenizers de LLM actuels utilisent le BPE ou une variante au niveau octet. - [Tokens d’entrée en cache (cached input tokens)](https://broadrouter.com/fr/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): Les tokens d’entrée en cache sont des tokens de prompt qu’un fournisseur de LLM a traités récemment et peut réutiliser depuis un cache au lieu de les recalculer, en général un préfixe répété comme un prompt système ou un document. Les fournisseurs les facturent moins cher que les nouveaux tokens d’entrée. - [Chaîne de pensée (chain-of-thought, CoT)](https://broadrouter.com/fr/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): La chaîne de pensée (chain-of-thought, CoT) est une technique de prompting dans laquelle on montre à un modèle de langage, ou on lui demande d’écrire, des étapes de raisonnement intermédiaires avant sa réponse finale. Elle améliore la précision sur les problèmes à plusieurs étapes comme l’arithmétique et la logique, et les modèles de raisonnement la génèrent désormais automatiquement. - [Taux d’erreur sur les caractères (CER)](https://broadrouter.com/fr/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Le taux d’erreur sur les caractères (CER) mesure la précision d’une transcription ou d’un OCR au niveau du caractère : substitutions, suppressions et insertions nécessaires pour transformer la sortie en texte de référence, divisées par le nombre de caractères de la référence. Plus il est bas, mieux c’est ; 0 signifie une correspondance parfaite. - [API Chat Completions](https://broadrouter.com/fr/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): L’API Chat Completions est l’endpoint HTTP d’OpenAI, POST /v1/chat/completions, qui reçoit un nom de modèle et une liste de messages associés à un rôle, et renvoie le message suivant du modèle. Sa forme de requête et de réponse est devenue le format commun que reprennent les fournisseurs et SDK compatibles OpenAI. - [Chat template (modèle de conversation)](https://broadrouter.com/fr/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): Un chat template est la règle de mise en forme, généralement écrite en Jinja et livrée avec le tokenizer d’un modèle, qui convertit une liste de messages de chat avec leurs rôles en la séquence unique de tokens attendue par le modèle, tokens de contrôle spéciaux compris. Un mauvais template peut fortement dégrader les réponses d’un modèle. - [Chunking (découpage pour le RAG)](https://broadrouter.com/fr/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Le chunking est l’étape d’un pipeline de récupération qui découpe les documents en morceaux plus petits, les chunks, avant de les vectoriser et de les indexer. La taille de chunk fixe la quantité de texte par morceau, et le chevauchement répète un peu de texte entre chunks voisins pour ne pas perdre les idées coupées à une frontière. - [Code-switching (alternance codique)](https://broadrouter.com/fr/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Le code-switching, ou alternance codique, est l’usage de deux langues ou plus, ou de plusieurs variétés d’une même langue, au sein d’une même conversation ou phrase, à l’oral comme à l’écrit. Par exemple, des arabophones qui glissent des mots anglais dans leur dialecte, ou qui passent de l’arabe standard moderne à leur dialecte. - [Agent de code IA (AI coding agent)](https://broadrouter.com/fr/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): Un agent de code IA est un outil logiciel dans lequel un modèle de langage travaille en boucle sur une base de code : il lit des fichiers, les modifie, lance des commandes comme les tests, lit la sortie et décide de l’étape suivante jusqu’à ce que la tâche soit terminée ou qu’une limite soit atteinte. - [Concurrence (requêtes simultanées)](https://broadrouter.com/fr/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): La concurrence est le nombre de requêtes qu’un client a en cours au même moment. Les API de LLM la plafonnent souvent par clé ou par forfait, indépendamment des requêtes par minute, car chaque requête ouverte occupe de la capacité GPU jusqu’à la fin de sa réponse. - [Compaction du contexte (context compaction)](https://broadrouter.com/fr/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): La compaction du contexte consiste à résumer une conversation qui approche de la limite de la fenêtre de contexte et à poursuivre la session à partir de ce résumé, afin qu’un agent conserve les décisions clés et l’état tout en abandonnant l’historique brut, comme les anciennes sorties d’outils. - [Context engineering (ingénierie du contexte)](https://broadrouter.com/fr/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): Le context engineering est le travail qui consiste à choisir et à entretenir l’ensemble des tokens qu’un modèle de langage voit à chaque requête (prompt système, définitions d’outils, historique, documents récupérés et notes) pour que le modèle reçoive le plus petit ensemble d’informations pertinentes dont il a besoin pour bien faire la tâche. - [Context rot (dégradation du contexte)](https://broadrouter.com/fr/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): Le context rot est la perte progressive de précision et de fiabilité que montrent les modèles de langage quand leur entrée s’allonge, même si la tâche reste la même et que l’entrée tient largement dans la fenêtre de contexte. Le terme a été popularisé par un rapport technique de Chroma de 2025 qui a testé 18 modèles. - [Fenêtre de contexte (context window)](https://broadrouter.com/fr/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): La fenêtre de contexte est le nombre maximal de tokens qu’un modèle de langage peut traiter dans une seule requête. Elle couvre le prompt, l’historique de conversation, les définitions d’outils, les images et la sortie générée, tokens de raisonnement compris. Ce qui ne tient pas doit être coupé, résumé ou réparti sur plusieurs requêtes. - [Résidence des données (data residency)](https://broadrouter.com/fr/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): La résidence des données est une exigence ou un engagement selon lequel les données sont stockées et traitées dans un pays ou une région précis. Pour une API LLM, elle couvre le lieu où les prompts sont traités pendant l’inférence et celui où se trouve tout contenu stocké ; elle exclut souvent les données de compte et de facturation. - [Arabe dialectal (dialectal Arabic)](https://broadrouter.com/fr/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): L’arabe dialectal est l’arabe parlé au quotidien dans une région ou une ville, comme l’égyptien, l’arabe du Golfe, le levantin ou le maghrébin. Les dialectes diffèrent de l’arabe standard moderne par les sons, les mots et la grammaire, n’ont pas d’orthographe standard et s’écrivent de plus en plus en ligne, ce qui complique la tâche des modèles de langage. - [Diarisation du locuteur (speaker diarization)](https://broadrouter.com/fr/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): La diarisation du locuteur consiste à découper un enregistrement audio en segments par locuteur, pour répondre à la question « qui parle quand ». Elle attribue à chaque passage de parole un identifiant anonyme comme SPEAKER_00 et se combine généralement avec une transcription pour montrer qui a dit chaque ligne. - [Distillation (knowledge distillation)](https://broadrouter.com/fr/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): La distillation, ou distillation des connaissances, est une méthode d’entraînement dans laquelle un modèle élève plus petit apprend à reproduire les sorties d’un modèle maître plus grand, comme ses distributions de probabilités ou ses réponses générées, afin de conserver une grande partie de sa qualité pour un coût d’exécution moindre. - [Accord de traitement des données (DPA)](https://broadrouter.com/fr/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): Un accord de traitement des données (DPA, data processing agreement) est le contrat entre un responsable du traitement et un sous-traitant qui encadre la manière dont le sous-traitant traite les données personnelles. L’article 28, paragraphe 3, du RGPD l’impose et en liste le contenu : instructions documentées, confidentialité, sécurité, règles de sous-traitance ultérieure, suppression ou restitution des données, et audits. - [Embeddings (plongements vectoriels)](https://broadrouter.com/fr/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): Les embeddings sont des listes de nombres à virgule flottante (des vecteurs) qu’un modèle produit pour représenter un texte, une image ou une autre entrée. Des entrées de sens proche obtiennent des vecteurs voisins : la distance entre vecteurs permet de classer des résultats de recherche, de regrouper des documents ou de trouver des éléments apparentés. - [Eval (évaluation de LLM)](https://broadrouter.com/fr/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Une eval (abréviation d’évaluation) est un test reproductible d’un modèle de langage ou d’un prompt : un ensemble fixe d’entrées, une description de ce qu’est une bonne sortie et un correcteur (du code, un humain ou un autre modèle) qui note chaque sortie pour pouvoir comparer les versions. - [Backoff exponentiel (exponential backoff)](https://broadrouter.com/fr/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): Le backoff exponentiel est une stratégie de nouvelle tentative dans laquelle un client attend plus longtemps après chaque échec, en doublant généralement le délai jusqu’à un plafond, avant de réessayer. Ajouter un jitter aléatoire à chaque attente étale les tentatives de nombreux clients, pour qu’elles ne frappent pas le serveur au même instant. - [Few-shot prompting (prompt avec exemples)](https://broadrouter.com/fr/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): Le few-shot prompting est une technique où le prompt contient quelques exemples entrée-sortie résolus avant l’entrée réelle, pour qu’un modèle de langage en déduise la tâche et le format. Les poids du modèle ne changent pas ; le zero-shot n’utilise aucun exemple et le one-shot un seul. - [Fine-tuning (affinage d’un modèle)](https://broadrouter.com/fr/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): Le fine-tuning consiste à poursuivre l’entraînement d’un modèle de langage pré-entraîné sur un jeu de données plus petit et propre à une tâche, comme des conversations d’exemple, pour que ses poids changent et qu’il exécute cette tâche ou ce style de façon plus fiable, sans longues instructions ni exemples dans chaque prompt. - [finish_reason](https://broadrouter.com/fr/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason est un champ présent sur chaque choice d’une réponse de chat completion, qui indique pourquoi le modèle a arrêté de générer. Au format OpenAI, les valeurs sont stop, length, tool_calls, content_filter et function_call, obsolète. L’API Messages d’Anthropic exprime la même idée avec stop_reason. - [Forfait LLM (flat rate)](https://broadrouter.com/fr/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): Un forfait LLM facture un prix fixe pour une période, par exemple un mois, en échange de l’accès à une API de modèle de langage, au lieu de facturer au token. Ces forfaits limitent généralement autre chose, comme la concurrence, la vitesse, les modèles ou des fenêtres d’usage : les plafonds comptent donc autant que le prix. - [FP8](https://broadrouter.com/fr/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): Le FP8 est un format de nombres à virgule flottante sur 8 bits utilisé en deep learning, défini en deux encodages : E4M3, avec 4 bits d’exposant et 3 de mantisse, et E5M2, avec 5 bits d’exposant et 2 de mantisse. Il divise par deux la mémoire par rapport aux formats 16 bits et s’exécute nativement sur les GPU récents. - [Frequency penalty et presence penalty](https://broadrouter.com/fr/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): La frequency penalty et la presence penalty sont des réglages d’échantillonnage qui abaissent le score des tokens déjà présents dans la sortie. La presence penalty s’applique une fois dès qu’un token est apparu ; la frequency penalty augmente à chaque répétition. Dans l’API d’OpenAI, les deux vont de -2.0 à 2.0, avec 0 par défaut. - [RGPD (GDPR)](https://broadrouter.com/fr/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): Le RGPD (règlement général sur la protection des données, règlement (UE) 2016/679, GDPR en anglais) est la loi de l’UE sur le traitement des données personnelles. Applicable depuis le 25 mai 2018, il couvre les organismes établis dans l’UE et certains organismes hors UE. Pour les API LLM, il encadre la sous-traitance, les transferts internationaux et la minimisation des données. - [GGUF](https://broadrouter.com/fr/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): Le GGUF est un format de fichier binaire qui stocke des modèles de langage pour l’inférence avec des runtimes basés sur GGML comme llama.cpp. Un fichier GGUF contient les poids, souvent quantifiés, ainsi que les métadonnées nécessaires au chargement, dont le tokenizer et l’architecture, et peut être mappé en mémoire pour un chargement rapide. - [Heure GPU (GPU-hour)](https://broadrouter.com/fr/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): Une heure GPU correspond à un processeur graphique utilisé pendant une heure. Les fournisseurs cloud facturent les instances GPU louées à l’heure ou à la seconde, et les chercheurs expriment en heures GPU la quantité de calcul qu’a demandée un travail. Vous payez le temps de la machine, qu’elle soit occupée ou non. - [Grounding (ancrage dans les sources)](https://broadrouter.com/fr/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): Le grounding consiste à rattacher la sortie d’un modèle de langage à des sources d’information précises et vérifiables, comme des documents fournis, des résultats de recherche ou des enregistrements de base de données, et à lui demander d’y puiser sa réponse. Une réponse ancrée peut être retracée jusqu’à une source et contient moins de faits inventés. - [Guardrails (garde-fous pour LLM)](https://broadrouter.com/fr/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): Les guardrails sont des contrôles qu’une application exécute autour d’un appel à un modèle de langage pour le maintenir dans des limites définies. Ceux d’entrée filtrent ce qui entre dans le modèle, ceux de sortie vérifient ce qui en sort, et ceux d’outils valident les actions avant leur exécution. Un contrôle échoué bloque, réécrit ou signale la requête. - [Hallucination (IA)](https://broadrouter.com/fr/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): Une hallucination est une sortie d’un modèle de langage qui paraît plausible mais qui est fausse, inventée ou non étayée par l’entrée fournie : une référence inventée, un chiffre erroné ou une API qui n’existe pas. Elle vient du fait que le modèle prédit un texte probable au lieu de vérifier des faits. - [Hamza](https://broadrouter.com/fr/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): La hamza (ء) est la lettre arabe qui note le coup de glotte. Elle s’écrit seule ou sur un support : alif (أ إ), waw (ؤ) ou ya (ئ), et آ combine alif et madda. Comme le texte saisi l’écrit de façon irrégulière, les moteurs de recherche et l’évaluation de la parole ramènent souvent les formes d’alif à ا. - [Recherche hybride (hybrid search)](https://broadrouter.com/fr/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): La recherche hybride est une méthode de récupération qui exécute, sur la même requête, une recherche par mots-clés comme BM25 et une recherche vectorielle sur des embeddings, puis fusionne les deux listes classées en une seule, souvent par reciprocal rank fusion (RRF) ou par somme pondérée de scores normalisés. - [Clé d’idempotence (idempotency key)](https://broadrouter.com/fr/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Une clé d’idempotence est une valeur unique, généralement un identifiant aléatoire comme un UUID, qu’un client joint à une requête qui modifie un état, comme un paiement ou un achat. Si la même requête est renvoyée avec la même clé, le serveur renvoie le résultat d’origine au lieu d’exécuter l’action deux fois. - [Tokens d’image (image tokens)](https://broadrouter.com/fr/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): Les tokens d’image sont les tokens d’entrée qu’un modèle vision-langage utilise pour représenter une image dans un prompt. Les fournisseurs les calculent à partir de la taille de l’image, en général en la découpant en patchs ou tuiles de taille fixe après redimensionnement, et les facturent comme des tokens d’entrée texte : une image plus grande coûte plus cher. - [Fournisseur d’inférence (inference provider)](https://broadrouter.com/fr/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): Un fournisseur d’inférence est une entreprise qui fait tourner des modèles de machine learning, souvent des LLM open-weight, sur des GPU qu’elle possède, loue ou prend en location, et qui en vend l’accès via une API, généralement facturée au token, à la requête ou au temps. Il gère le service, la montée en charge et le matériel. - [Tokens d’entrée et de sortie](https://broadrouter.com/fr/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): Les tokens d’entrée sont les tokens qu’un modèle de langage lit dans une requête, dont le prompt système, l’historique de conversation, les définitions d’outils et les images. Les tokens de sortie sont ceux qu’il génère en réponse, raisonnement compris. Les API LLM les facturent généralement séparément, et la sortie coûte en général plus cher. - [INT4 (quantification 4 bits)](https://broadrouter.com/fr/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 est un format d’entier sur 4 bits qui représente 16 niveaux. En inférence de LLM, il désigne en général la quantification des poids seuls sur 4 bits (W4A16) : les poids sont stockés en INT4 avec des facteurs d’échelle et repassés en 16 bits pendant le calcul, pour environ un quart de la mémoire de poids en 16 bits. - [Liste d’IP autorisées (IP allow list)](https://broadrouter.com/fr/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): Une liste d’IP autorisées est un contrôle de sécurité qui n’accepte que les requêtes provenant d’adresses IP ou de plages CIDR listées et refuse toutes les autres. Sur une API, elle limite les dégâts d’une clé divulguée, car la clé ne fonctionne pas depuis des réseaux absents de la liste. - [Jailbreak (LLM)](https://broadrouter.com/fr/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): Un jailbreak est une entrée conçue pour pousser un modèle de langage à ignorer son entraînement de sécurité ou les règles fixées par l’application qui le déploie, afin qu’il produise des contenus ou exécute des actions qu’il refuserait normalement. L’OWASP considère le jailbreaking comme une forme de prompt injection. - [Mode JSON (JSON mode)](https://broadrouter.com/fr/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): Le mode JSON est un réglage des API de LLM, response_format {"type": "json_object"} dans les API de style OpenAI, qui fait de la réponse du modèle un JSON valide et analysable. Contrairement aux Structured Outputs avec un JSON Schema, il ne garantit pas les clés, les types ni les valeurs de l’objet. - [JSON Schema](https://broadrouter.com/fr/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema est un vocabulaire, écrit en JSON, qui sert à décrire et à valider la structure de données JSON : types, champs obligatoires, valeurs autorisées et imbrication. La version actuelle est la 2020-12. Les API de LLM l’utilisent pour définir les paramètres des outils et la forme exacte que doit suivre une sortie structurée. - [Cache KV (KV cache)](https://broadrouter.com/fr/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): Le cache KV est la mémoire où un modèle de langage transformer stocke les clés et les valeurs d’attention calculées pour chaque token déjà traité, afin que chaque nouveau token les réutilise au lieu de recalculer toute la séquence. Il grossit avec la longueur du contexte et limite souvent le nombre de requêtes qu’un GPU peut servir. - [Latence (LLM)](https://broadrouter.com/fr/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): Pour une API de modèle de langage, la latence est le temps que prend une requête entre son envoi et la réception du dernier token de sortie. C’est la somme du time to first token et du temps de génération, généralement exprimée en percentiles comme le p50 et le p95 plutôt qu’en moyenne. - [LLM-as-a-judge (LLM juge)](https://broadrouter.com/fr/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge est une méthode d’évaluation dans laquelle un modèle de langage note les sorties d’autres modèles, soit en évaluant une réponse unique selon une grille, soit en choisissant la meilleure de deux réponses. Elle approche le jugement humain pour un coût bien moindre, mais présente des biais connus, comme une préférence pour la première réponse ou la plus longue. - [Benchmark de LLM](https://broadrouter.com/fr/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): Un benchmark de LLM est un ensemble fixe de tâches, généralement public, avec une méthode de notation, qui mesure et compare des modèles de langage sur une compétence : connaissances, code ou reconnaissance vocale. Un score de benchmark indique un niveau relatif sur ce test, pas les résultats d’un modèle sur votre propre charge de travail. - [Routeur LLM (LLM router)](https://broadrouter.com/fr/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): Un routeur LLM est un logiciel ou un service placé entre une application et une ou plusieurs API de modèles de langage, qui décide pour chaque requête quel modèle ou quel fournisseur la traite, selon des règles comme le coût, la latence, la disponibilité ou le type de tâche. - [Logprobs (log-probabilités)](https://broadrouter.com/fr/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Les logprobs sont les log-probabilités qu’un modèle de langage a attribuées aux tokens qu’il a générés. Une API qui les renvoie montre, pour chaque token de sortie, à quel point le modèle le jugeait probable et, en option, les tokens alternatifs les plus probables à cette position. - [LoRA (low-rank adaptation)](https://broadrouter.com/fr/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation) est une méthode de fine-tuning économe en paramètres qui gèle les poids d’un modèle pré-entraîné et entraîne une paire de petites matrices de bas rang pour certaines couches. La mise à jour apprise peut être stockée comme un petit adaptateur ou fusionnée dans les poids, sans latence d’inférence supplémentaire. - [Lost in the middle (perdu au milieu)](https://broadrouter.com/fr/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle désigne la tendance observée des modèles de langage à mieux exploiter une information pertinente lorsqu’elle se trouve près du début ou de la fin d’une longue entrée, et moins bien lorsqu’elle se situe au milieu. Le phénomène a été nommé dans une étude de Liu et al. en 2023. - [max_tokens](https://broadrouter.com/fr/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens est un paramètre des API LLM qui fixe le nombre maximal de tokens qu’un modèle peut générer dans une réponse. Quand la génération atteint la limite, elle s’arrête et la réponse est marquée comme coupée, par exemple avec finish_reason égal à « length ». - [MCP (Model Context Protocol)](https://broadrouter.com/fr/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol) est un protocole ouvert, fondé sur JSON-RPC 2.0, qui normalise la façon dont les applications LLM se connectent à des outils et des données externes. Un serveur MCP expose des outils, des ressources et des prompts ; un client MCP, dans l’application hôte, les découvre et transmet au serveur les appels d’outils du modèle. - [Mixture of experts (MoE, mélange d’experts)](https://broadrouter.com/fr/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Le mixture of experts (MoE) est une architecture de réseau de neurones dans laquelle une couche contient de nombreux sous-réseaux experts et un routeur appris envoie chaque token vers quelques-uns seulement. Un modèle de langage MoE a un grand nombre total de paramètres mais n’en utilise qu’un petit sous-ensemble actif par token. - [Dépréciation de modèle (model deprecation)](https://broadrouter.com/fr/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): La dépréciation d’un modèle, c’est l’annonce par un fournisseur d’API qu’un modèle cessera d’être proposé, en général avec une date de retrait ou d’arrêt et un remplaçant recommandé. Après cette date, les requêtes vers le modèle échouent : les applications doivent donc migrer vers un autre ID de modèle avant. - [Fallback de modèle (model fallback)](https://broadrouter.com/fr/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): Le fallback de modèle est un schéma de fiabilité où une requête LLM échouée est relancée sur un autre modèle ou fournisseur d’une liste ordonnée. Déclencheurs possibles : erreur, limite de débit, contexte trop long ou refus lié à la politique de contenu. Il s’exécute dans votre client ou dans une passerelle devant les API. - [Licence de modèle (model license)](https://broadrouter.com/fr/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): Une licence de modèle désigne les conditions juridiques selon lesquelles les poids, le code et les sorties d’un modèle peuvent être utilisés, modifiés et partagés. Les modèles open-weight utilisent des licences logicielles standard comme Apache 2.0 ou MIT, ou des licences propres pouvant ajouter des conditions, par exemple pour les grands services commerciaux. - [Arabe standard moderne (MSA)](https://broadrouter.com/fr/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): L’arabe standard moderne (MSA) est la forme standardisée de l’arabe utilisée à l’écrit, dans l’information, l’enseignement et le discours formel dans tout le monde arabe. Ce n’est la langue maternelle de personne : on parle au quotidien des dialectes régionaux, comme l’égyptien ou l’arabe du Golfe, et on passe au MSA dans les contextes formels. - [Système multi-agents (multi-agent system)](https://broadrouter.com/fr/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): Un système multi-agents est une application LLM dans laquelle plusieurs agents, chacun avec son prompt, ses outils et son contexte, traitent les parties d’une même tâche. Un schéma courant est orchestrateur-workers : un agent principal planifie, lance des agents workers en parallèle et combine leurs résultats. - [IA multimodale (multimodal AI)](https://broadrouter.com/fr/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): L’IA multimodale désigne un modèle ou un système qui travaille avec plusieurs types de données, comme le texte, les images, la vidéo et l’audio. Un modèle multimodal peut lire une photo et une question ensemble, ou recevoir de la parole et répondre en texte. Le modèle vision-langage en est le type le plus courant. - [Reconnaissance d’entités nommées (NER)](https://broadrouter.com/fr/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): La reconnaissance d’entités nommées (NER) est la tâche qui consiste à trouver les noms dans un texte et à attribuer à chacun un type, comme personne, organisation ou lieu. Les systèmes classiques étiquettent chaque token ; on demande plutôt aux LLM une liste d’entités en JSON, ensuite vérifiée par rapport au texte source. - [Reconnaissance optique de caractères (OCR)](https://broadrouter.com/fr/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): La reconnaissance optique de caractères (OCR) est une technologie qui détecte le texte imprimé ou manuscrit d’une image, d’un scan ou d’un PDF et le convertit en caractères lisibles par une machine. L’OCR classique produit le texte et sa position sur la page ; les modèles vision-langage peuvent lire le texte et répondre directement à des questions dessus. - [Ogg Opus](https://broadrouter.com/fr/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): L’Ogg Opus est de l’audio encodé avec le codec Opus (RFC 6716) et stocké dans un conteneur Ogg (RFC 7845). Conçu pour la parole et l’audio interactif à faible débit, c’est le format qu’utilise WhatsApp pour les messages vocaux, en général sous forme de fichiers .ogg ou .opus. - [Modèle open-weight (à poids ouverts)](https://broadrouter.com/fr/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): Un modèle open-weight est un modèle d’apprentissage automatique dont les poids entraînés sont publiés pour que chacun puisse les télécharger et les exécuter, généralement sous une licence assortie de conditions. Il se distingue de l’IA entièrement open source, qui publie aussi le code d’entraînement et des informations détaillées sur les données, pour que d’autres puissent étudier et reproduire le modèle. - [API compatible OpenAI](https://broadrouter.com/fr/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): Une API compatible OpenAI est une API HTTP qui accepte les mêmes formats de requête et de réponse que l’API OpenAI, comme POST /v1/chat/completions. Les SDK et outils OpenAI peuvent donc l’utiliser en changeant seulement l’URL de base, la clé d’API et le nom du modèle. - [Appels d’outils parallèles (parallel tool calls)](https://broadrouter.com/fr/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): On parle d’appels d’outils parallèles quand un modèle renvoie plus d’un appel d’outil dans une même réponse, par exemple trois lectures de fichiers d’un coup. Au format OpenAI, ils arrivent comme plusieurs entrées dans tool_calls, et parallel_tool_calls à false limite le modèle à zéro ou un appel par tour. - [Paiement à l’usage (pay-as-you-go)](https://broadrouter.com/fr/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): Le paiement à l’usage (pay-as-you-go) est un modèle tarifaire où vous ne payez que ce que vous consommez, mesuré par unité, comme les tokens, les requêtes ou les minutes d’audio, sans abonnement ni engagement minimal. La plupart des API LLM facturent ainsi, à partir d’un solde prépayé ou d’une facture périodique. - [PDPL (lois saoudienne et émiratie sur les données)](https://broadrouter.com/fr/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): La PDPL (Personal Data Protection Law) désigne la loi saoudienne sur les données personnelles, promulguée par le décret royal M/19 et appliquée par la SDAIA, et le décret-loi fédéral n° 45 de 2021 des Émirats. Les deux encadrent le traitement des données personnelles et leurs transferts à l’étranger, y compris les prompts envoyés à des API LLM étrangères. - [PII (données d’identification personnelle)](https://broadrouter.com/fr/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): Les PII (personally identifiable information) sont les informations permettant de distinguer ou de retrouver l’identité d’une personne, comme un nom, un numéro d’identité ou des données biométriques, et toute information qui lui est liée ou rattachable. Le RGPD parle plus largement de données personnelles. Avec les LLM, le masquage des PII les retire des prompts avant qu’ils quittent vos systèmes. - [Prefill et decode](https://broadrouter.com/fr/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Le prefill et le decode sont les deux phases de l’inférence d’un LLM. Pendant le prefill, le modèle traite tout le prompt en parallèle et remplit son cache KV ; pendant le decode, il génère la réponse token par token en réutilisant ce cache. Le prefill détermine surtout le time to first token, le decode la vitesse de sortie. - [Crédit prépayé (prepaid credit)](https://broadrouter.com/fr/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): Le crédit prépayé est de l’argent versé sur un compte API avant utilisation. Chaque requête déduit son coût du solde, et quand le solde est épuisé, les requêtes facturées s’arrêtent jusqu’à la prochaine recharge. Les dépenses sont ainsi plafonnées à ce que vous avez versé. - [Prompt engineering (ingénierie de prompt)](https://broadrouter.com/fr/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): Le prompt engineering est le travail qui consiste à écrire, structurer et tester les instructions données à un modèle de langage pour qu’il produise de façon constante une sortie conforme à vos exigences. Il couvre les messages de rôle, des instructions claires, des exemples, la mise en forme et la mesure des résultats sur des cas de test. - [Injection de prompt (prompt injection)](https://broadrouter.com/fr/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): L’injection de prompt est une attaque dans laquelle un texte fourni à un modèle de langage, par un utilisateur ou caché dans une page web, un e-mail, un fichier ou un résultat d’outil, modifie le comportement du modèle contre l’intention du développeur. L’OWASP la classe en tête (LLM01:2025) de son Top 10 pour les applications LLM. - [Quantification (quantization)](https://broadrouter.com/fr/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): La quantification consiste à stocker les nombres d’un réseau de neurones, surtout ses poids, dans un format de moindre précision comme 8 ou 4 bits au lieu de 16 bits, pour que le modèle utilise moins de mémoire et tourne souvent plus vite. Les bonnes méthodes limitent la perte de précision. - [RAG (génération augmentée par récupération)](https://broadrouter.com/fr/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): Le RAG (retrieval-augmented generation, génération augmentée par récupération) est une approche dans laquelle une application récupère d’abord, dans une source externe comme un index de recherche ou un index vectoriel, les passages pertinents pour une question, puis les ajoute au prompt pour qu’un modèle de langage réponde à partir de ce texte et pas seulement de sa mémoire. - [Limite de débit (rate limit)](https://broadrouter.com/fr/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): Une limite de débit (rate limit) est un plafond qu’une API impose à l’usage d’un client sur une période, par exemple en requêtes ou en tokens par minute. Les requêtes au-delà sont refusées, généralement avec HTTP 429 Too Many Requests, jusqu’à ce que l’usage repasse sous la limite. - [ReAct (reason and act)](https://broadrouter.com/fr/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (reason and act) est un schéma de prompting, présenté par Yao et al. en 2022, dans lequel un modèle de langage entrelace des traces de raisonnement et des actions, comme une recherche ou un appel d’outil. Il lit chaque observation, met à jour son plan et recommence jusqu’à pouvoir donner une réponse finale. - [Tokens de raisonnement (reasoning tokens)](https://broadrouter.com/fr/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): Les tokens de raisonnement sont les tokens qu’un modèle de raisonnement génère pour résoudre un problème avant d’écrire sa réponse visible. Ils sont généralement facturés comme des tokens de sortie, comptent dans la limite de sortie et dans la fenêtre de contexte, et peuvent être masqués ou résumés par l’API. - [Timeout de requête (délai d’expiration)](https://broadrouter.com/fr/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): Un timeout de requête est la durée maximale pendant laquelle un client HTTP attend un serveur avant d’abandonner la requête. Les clients le divisent souvent en un timeout de connexion, pour ouvrir la connexion, et un timeout de lecture, pour attendre les données. Les appels de LLM exigent des timeouts de lecture généreux, car les longues réponses prennent du temps. - [Reranking (reclassement)](https://broadrouter.com/fr/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Le reranking est une seconde étape de recherche dans laquelle un modèle plus précis mais plus lent, en général un cross-encoder, note chaque paire requête-document candidat issue d’une première recherche et les réordonne, afin que seuls les quelques plus pertinents soient transmis au modèle de langage. - [Texte de droite à gauche (RTL)](https://broadrouter.com/fr/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): Le texte de droite à gauche (RTL) est une écriture, comme l’arabe ou l’hébreu, affichée de droite à gauche. Les ordinateurs la stockent dans l’ordre logique de lecture, et l’algorithme bidirectionnel Unicode décide de l’ordre d’affichage quand le texte RTL se mêle à du texte de gauche à droite, des chiffres, des noms de modèles ou du code. - [Fréquence d’échantillonnage (sample rate)](https://broadrouter.com/fr/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): La fréquence d’échantillonnage (sample rate) est le nombre d’échantillons audio captés par seconde, mesuré en hertz (Hz). Les modèles de reconnaissance vocale attendent souvent un audio mono à 16 000 Hz (16 kHz). Avec la profondeur de bits et le nombre de canaux, elle fixe la taille d’un fichier audio non compressé par minute. - [Paramètre seed (graine aléatoire)](https://broadrouter.com/fr/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): Le paramètre seed est un entier envoyé avec une requête pour que l’API essaie de tirer les mêmes tokens quand la même requête est répétée. OpenAI parle de « best effort » : le déterminisme n’est pas garanti, et le champ system_fingerprint signale quand la configuration du backend a changé. - [Inférence serverless (serverless inference)](https://broadrouter.com/fr/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): L’inférence serverless est une façon d’exécuter des modèles de machine learning où la plateforme démarre, met à l’échelle et arrête automatiquement le calcul derrière un endpoint, souvent jusqu’à zéro au repos, et facture l’usage plutôt que des machines réservées. La contrepartie est un possible délai de démarrage à froid (cold start). - [SLA (accord de niveau de service)](https://broadrouter.com/fr/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): Un SLA (accord de niveau de service) est un engagement contractuel d’un fournisseur sur un niveau de service mesurable, le plus souvent un pourcentage de disponibilité mensuelle, avec une compensation définie, généralement des crédits de service, s’il n’est pas atteint. Un SLA précise ce qui compte comme indisponibilité, ce qui est exclu et comment réclamer. - [Décodage spéculatif (speculative decoding)](https://broadrouter.com/fr/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): Le décodage spéculatif est une technique d’inférence dans laquelle un petit modèle brouillon rapide propose plusieurs tokens suivants, que le grand modèle cible vérifie en une seule passe avant (forward pass) en gardant ceux qu’il accepte. Il accélère la génération sans changer la distribution de la sortie. - [Speech-to-text (reconnaissance vocale)](https://broadrouter.com/fr/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): Le speech-to-text, ou reconnaissance automatique de la parole (ASR), est une technologie qui convertit l’audio parlé en texte écrit. Une API speech-to-text accepte un fichier ou un flux audio et renvoie une transcription, généralement facturée à la minute d’audio. La précision se mesure avec le taux d’erreur de mots et varie selon l’accent, le dialecte et la qualité d’enregistrement. - [Plafond de dépenses (spend limit)](https://broadrouter.com/fr/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): Un plafond de dépenses (spend limit) limite ce qu’un compte, un projet ou une clé API peut dépenser sur une période, généralement un mois. Une fois atteint, le fournisseur rejette les requêtes suivantes, souvent avec une erreur HTTP 429 ou 400, jusqu’à la réinitialisation de la période ou au relèvement de la limite. - [Sous-titres SRT](https://broadrouter.com/fr/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) est un format de sous-titres en texte brut. Chaque sous-titre comprend un numéro d’ordre, une heure de début et de fin écrite hh:mm:ss,mmm --> hh:mm:ss,mmm, une ou plusieurs lignes de texte et une ligne vide. WebVTT est le format web similaire utilisé par la vidéo HTML. - [Server-sent events (SSE)](https://broadrouter.com/fr/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Les server-sent events (SSE) sont un standard du web qui permet à un serveur de pousser un flux d’événements texte vers un client, via une seule réponse HTTP de longue durée de type text/event-stream. Chaque événement se compose d’une ou plusieurs lignes de champ, comme data:, et se termine par une ligne vide. - [Séquences d’arrêt (stop sequences)](https://broadrouter.com/fr/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): Les séquences d’arrêt (stop sequences) sont des chaînes envoyées avec une requête qui font cesser la génération dès que le modèle produit l’une d’elles. OpenAI appelle ce champ stop et accepte jusqu’à 4 séquences ; Anthropic l’appelle stop_sequences. Le texte trouvé est normalement exclu de la sortie renvoyée. - [Streaming](https://broadrouter.com/fr/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): Le streaming est une manière de renvoyer la réponse d’un modèle de langage par petits chunks pendant sa génération, plutôt qu’en un seul bloc à la fin. Les API compatibles OpenAI streament via des server-sent events quand la requête règle stream sur true. - [Sortie structurée (structured output)](https://broadrouter.com/fr/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): La sortie structurée est une fonction des API LLM qui contraint la réponse d’un modèle à un format défini par le développeur, en général un JSON Schema, pour que le code puisse l’analyser de façon fiable. Elle est plus stricte que le mode JSON, qui garantit un JSON valide mais pas une structure précise. - [Sous-agent (subagent)](https://broadrouter.com/fr/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): Un sous-agent est un agent distinct que l’agent principal lance pour traiter une tâche ciblée. Il s’exécute avec sa propre fenêtre de contexte, son prompt système et ses accès aux outils, fait le travail et ne renvoie qu’un résultat ou un résumé à l’agent principal, ce qui garde la conversation principale légère. - [Sous-traitant ultérieur (subprocessor)](https://broadrouter.com/fr/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): Un sous-traitant ultérieur (subprocessor) est un tiers qu’un sous-traitant engage pour effectuer des traitements de données personnelles pour le compte du responsable du traitement. Pour une API LLM, il peut s’agir d’hébergeurs GPU, de fournisseurs d’inférence partenaires ou de capacité de secours. Selon l’article 28 du RGPD, il faut l’autorisation du responsable et les mêmes obligations de protection des données. - [Prompt système (system prompt)](https://broadrouter.com/fr/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): Le prompt système est le texte d’instructions qu’une application envoie à un modèle de langage avant la conversation, généralement sous forme de message au rôle system ou developer. Il fixe le rôle, les règles, le ton et le format de sortie du modèle, qui est censé le suivre même face à des messages contradictoires de l’utilisateur. - [Tâ’ marbûta (ة)](https://broadrouter.com/fr/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): Le tâ’ marbûta (ة, U+0629) est une lettre arabe qui n’apparaît qu’en fin de mot et marque souvent les noms et adjectifs féminins. Il ressemble au hâ’ (ه, U+0647) surmonté de deux points : les textes saisis mélangent donc les deux, et de nombreux normaliseurs de recherche remplacent ة par ه. - [Température (temperature)](https://broadrouter.com/fr/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): La température est un paramètre d’échantillonnage qui règle le degré de hasard de la sortie d’un modèle de langage. Le modèle divise les scores des tokens par la température avant de les convertir en probabilités : une valeur basse fait dominer le token le plus probable, une valeur haute aplatit la distribution et favorise les tokens moins probables. - [Synthèse vocale (text-to-speech, TTS)](https://broadrouter.com/fr/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): La synthèse vocale (text-to-speech, TTS) convertit un texte écrit en audio parlé. Une API TTS reçoit un texte, une voix et un format de sortie comme MP3 ou WAV, et renvoie la parole générée. C’est l’inverse du speech-to-text, qui transforme l’audio en texte. - [Débit (throughput)](https://broadrouter.com/fr/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): Le débit (throughput), dans le service de LLM, est la quantité totale de travail qu’un système accomplit par unité de temps sur toutes les requêtes simultanées, généralement mesurée en tokens de sortie par seconde ou en requêtes par seconde. Il se distingue de la vitesse par flux, c’est-à-dire la vitesse à laquelle une seule requête reçoit ses tokens. - [Time to first token (TTFT)](https://broadrouter.com/fr/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): Le time to first token (TTFT) est le temps écoulé entre l’envoi d’une requête à un modèle de langage et l’arrivée du premier token de sortie. Il comprend le temps réseau, la file d’attente côté serveur et le prefill, où le modèle lit tout le prompt : il augmente donc avec la longueur du prompt et la charge. - [Horodatages (speech-to-text)](https://broadrouter.com/fr/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): En speech-to-text, les horodatages (timestamps) sont des heures de début et de fin, en secondes, attachées à des parties d’une transcription. Les horodatages de segment couvrent une phrase ou une expression ; les horodatages de mot couvrent chaque mot. Ils servent aux sous-titres, à la recherche dans les enregistrements et à l’alignement du texte avec les locuteurs. - [Token](https://broadrouter.com/fr/glossary/token/) (translation of https://broadrouter.com/glossary/token/): Un token est l’unité de texte qu’un modèle de langage lit et écrit : un mot, une partie de mot, un signe de ponctuation ou une suite d’octets produite par le tokenizer du modèle. Les API LLM comptent, limitent et facturent généralement l’usage en tokens. - [Tokenizer (tokeniseur)](https://broadrouter.com/fr/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): Un tokenizer est le composant qui convertit un texte en la suite d’ID entiers de tokens que lit un modèle de langage, et reconvertit les ID générés en texte. Chaque modèle a son propre tokenizer : un même texte peut donc produire un nombre de tokens différent selon le modèle. - [Tokens par seconde (TPS)](https://broadrouter.com/fr/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): Les tokens par seconde (TPS) sont le rythme auquel un modèle de langage génère des tokens de sortie, généralement mesuré pour une requête après son premier token. Ils fixent la vitesse d’affichage d’une réponse lancée. Additionnés sur toutes les requêtes simultanées, ils donnent le débit du système ; l’attente avant le premier token se mesure à part. - [Appel d’outils (tool calling, function calling)](https://broadrouter.com/fr/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): L’appel d’outils, aussi appelé function calling, est une fonctionnalité des API LLM dans laquelle le modèle renvoie une demande structurée d’exécution d’une fonction définie par le développeur, avec des arguments JSON. L’application exécute la fonction et renvoie le résultat ; le modèle n’exécute jamais de code lui-même. - [tool_choice](https://broadrouter.com/fr/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice est un paramètre de requête qui contrôle l’usage des outils. Dans les Chat Completions d’OpenAI, il vaut auto (le modèle décide), required (il doit appeler un ou plusieurs outils), none (aucun outil) ou le nom d’une fonction à forcer. Les options équivalentes chez Anthropic sont auto, any, tool et none. - [Top-p (nucleus sampling)](https://broadrouter.com/fr/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Le top-p, aussi appelé nucleus sampling (échantillonnage par noyau), est une méthode de décodage dans laquelle un modèle de langage tire le token suivant uniquement dans le plus petit ensemble de tokens les plus probables dont la probabilité cumulée atteint un seuil p, par exemple 0,9, et écarte la traîne improbable. - [Normalisation Unicode](https://broadrouter.com/fr/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): La normalisation Unicode convertit un texte dans une forme standard afin que des chaînes d’apparence identique utilisent les mêmes points de code. Les quatre formes sont NFD, NFC, NFKD et NFKC ; les formes K ramènent aussi les caractères de compatibilité, comme les formes de présentation arabes et les ligatures, à des lettres ordinaires. - [Détection d’activité vocale (VAD)](https://broadrouter.com/fr/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): La détection d’activité vocale (VAD, voice activity detection) est une étape de traitement du signal qui classe de courtes trames audio en parole ou non-parole. Les systèmes vocaux s’en servent pour sauter les silences, découper les longs enregistrements aux pauses, détecter la fin d’une prise de parole et éviter de transcrire le bruit de fond. - [Base de données vectorielle (vector database)](https://broadrouter.com/fr/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): Une base de données vectorielle est un système de stockage qui conserve des vecteurs d’embedding, souvent avec leur texte source et des métadonnées, et répond aux requêtes de plus proches voisins : pour un vecteur de requête, elle renvoie les vecteurs stockés les plus similaires, en général selon la distance cosinus, la distance euclidienne ou le produit scalaire. - [Modèle vision-langage (VLM)](https://broadrouter.com/fr/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): Un modèle vision-langage (VLM) est un modèle qui prend en entrée des images et du texte ensemble et produit du texte : il peut décrire des images, répondre à des questions à leur sujet et lire des documents, des graphiques ou des captures d’écran. On parle aussi de modèles image-texte-vers-texte ou multimodaux. - [vLLM](https://broadrouter.com/fr/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM est une bibliothèque open source d’inférence et de serving de LLM, développée à l’origine au Sky Computing Lab de UC Berkeley. Elle gère la mémoire du cache KV avec PagedAttention, regroupe en continu les requêtes entrantes (continuous batching), prend en charge de nombreux formats de quantification et expose un serveur HTTP compatible OpenAI. - [Taille du vocabulaire (vocabulary size)](https://broadrouter.com/fr/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): La taille du vocabulaire est le nombre de tokens distincts dans le tokenizer d’un modèle de langage, chacun avec son propre identifiant et son propre embedding. Un vocabulaire plus grand peut encoder le texte courant en moins de tokens, mais il alourdit les couches d’embedding et de sortie, et n’aide que pour le texte sur lequel il a été entraîné. - [WAV](https://broadrouter.com/fr/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): Le WAV (Waveform Audio File Format, ou WAVE) est un conteneur audio fondé sur le format RIFF de Microsoft et IBM. Il contient en général des échantillons PCM linéaires non compressés, d’où une taille prévisible : fréquence d’échantillonnage fois octets par échantillon fois canaux, par seconde. - [Taux d’erreur sur les mots (WER)](https://broadrouter.com/fr/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): Le taux d’erreur sur les mots (WER) est la mesure standard de précision en reconnaissance vocale : le nombre de substitutions, suppressions et insertions de mots nécessaires pour transformer une transcription en référence correcte, divisé par le nombre de mots de la référence. Plus il est bas, mieux c’est. - [Conservation zéro des données (zero data retention, ZDR)](https://broadrouter.com/fr/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): La conservation zéro des données (zero data retention, ZDR) est une politique selon laquelle un fournisseur d’IA traite prompts et sorties uniquement pour servir une requête, sans conserver ce contenu ensuite, ni dans des logs, ni dans des caches, ni dans des données d’entraînement. Les enregistrements sans contenu, comme le nombre de tokens et les durées, restent généralement conservés. ### Italian (it, Italiano): 294 translated pages - [Articles in Italian](https://broadrouter.com/it/blog/) - [Prezzi delle API LLM a ottobre 2026: 3 aumenti, nessun taglio](https://broadrouter.com/it/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 variazioni sulle pagine prezzi di 5 provider di API LLM su 7 a ottobre 2026: 3 aumenti, nessun taglio, 9 nuovi prezzi e 11 modelli rimossi. Con link all’archivio. - [Costo di un’esecuzione di un agente di coding: passi, contesto, cache](https://broadrouter.com/it/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): Perché un agente di coding costa di più a ogni passo man mano che va avanti, quanto cambia la cache e quando una sessione flat batte il pagamento a token. - [Aggiungere i diacritici all’arabo con un LLM: prompt, controlli, DER](https://broadrouter.com/it/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): Ripristina i diacritici arabi (tashkeel) con un’API LLM: un prompt che non tocca le lettere, un controllo che scarta le modifiche, DER e WER, e quando usare CATT. - [Correzione grammaticale e ortografica dell’arabo con un LLM: guida](https://broadrouter.com/it/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Correggi ortografia e grammatica arabe con un’API LLM: prompt a modifiche minime, controllo anti-riscrittura, modifiche per tipo di errore e punteggio M2. - [Verbali di riunioni in arabo dalle registrazioni: una pipeline API](https://broadrouter.com/it/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Trasforma le registrazioni di riunioni in arabo in verbali con decisioni e azioni: trascrivi, pulisci, riassumi con un LLM in JSON e verifica nomi e date. - [Estrarre entità nominate dall’arabo con un LLM: JSON, date e controlli](https://broadrouter.com/it/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Estrai in JSON nomi, organizzazioni, luoghi e date dell’Egira da testi arabi con un LLM: prompt, controlli che ignorano i diacritici, conversione e punteggio. - [API OCR per l’arabo a confronto: Google, Azure, Textract, Mistral, LLM](https://broadrouter.com/it/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): Quali API OCR supportano ufficialmente l’arabo, stampato e scritto a mano, cosa restituiscono e quanto costano ogni 1.000 pagine, confrontate con gli LLM vision. - [Sottotitoli in arabo (SRT) da un’API speech-to-text: cosa serve](https://broadrouter.com/it/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): Cosa serve a un file SRT, quali API speech-to-text restituiscono timestamp o SRT per l’arabo e una pipeline testata che crea sottotitoli da un’API solo testo. - [Classificare testo arabo con un LLM: sentiment, intenzione, triage](https://broadrouter.com/it/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): Classifica recensioni, messaggi e ticket in arabo con un LLM: design delle etichette, prompt attento al dialetto, output JSON, test con macro F1 e costi a volume. - [Normalizzare il testo arabo per ricerca e RAG: cosa cambiare, cosa no](https://broadrouter.com/it/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Tashkeel, alef, ta marbuta, hamza, lettere persiane, cifre e Unicode: quali regole di normalizzazione dell’arabo aiutano la ricerca e quali no, con Python testato. - [Riassumere testi in arabo con un LLM: prompt, lunghezza e valutazione](https://broadrouter.com/it/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): Riassumi articoli, report e messaggi in dialetto arabo con un’API LLM: prompt, controllo delle parole e dei numeri, e perché il ROUGE di default dà 0. - [Alternative ad Artificial Analysis: dove confrontare gli LLM nel 2026](https://broadrouter.com/it/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Alternative ad Artificial Analysis a confronto: Arena, Epoch AI, Vals AI, classifiche OpenRouter, Hugging Face e dati sui prezzi, per cosa misurano e quando usarle. - [Hardware GPU e serving di BroadNet, come li descrive BroadNet](https://broadrouter.com/it/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): Cosa dice BroadNet dell’hardware di BroadNet Inference: tre livelli di GPU, dalle B300 alle appliance dai clienti, throughput per GPU e cosa ha imparato dai test. - [Cache delle risposte delle API LLM: una cache esatta in Python](https://broadrouter.com/it/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Una cache a corrispondenza esatta per le risposte LLM in Python, testata: cosa mettere nella chiave, cosa non mettere in cache, TTL, privacy e come misurare gli hit. - [Claude Code con modelli open-weight: proxy, endpoint nativi e limiti](https://broadrouter.com/it/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code parla la Messages API di Anthropic. Come farlo girare su modelli open-weight con un proxy LiteLLM o un endpoint nativo, e cosa si rompe lungo la strada. - [Revisione umana nel flusso di document AI: regole e code](https://broadrouter.com/it/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): Progettare la revisione umana in una pipeline di document AI: regole di validazione, cosa passare a una persona, campioni casuali, audit trail e tasso di revisione. - [Legge egiziana sulla protezione dei dati (Legge 151/2020) e API LLM](https://broadrouter.com/it/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Inviare dati personali di egiziani a un’API LLM o speech-to-text all’estero: cosa dicono Legge 151/2020 e Decisione 816/2025 su trasferimenti, licenze e violazioni. - [Prima del fine-tuning: definisci il problema da risolvere](https://broadrouter.com/it/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): Quando prompt, RAG o valutazione risolvono il problema senza fine-tuning, quali dati di addestramento servono e come testare un fine-tune prima di chiederlo. - [JSON non valido da un LLM: cause, riparazione e retry](https://broadrouter.com/it/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): Perché gli LLM restituiscono JSON rotto (code fence, testo extra, tagli a max_tokens, virgole finali) e una funzione Python testata che ripara solo il sicuro. - [Ogni quanto cambiano i prezzi delle API LLM? 3 anni di archivio](https://broadrouter.com/it/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): Abbiamo letto 308 pagine prezzi archiviate di 11 provider API LLM, 2023–2026. L’80% dei modelli non ha mai cambiato prezzo; DeepInfra ha fatto il 75% dei ritocchi. - [Come ridurre i costi delle API LLM: checklist in 10 passi, in ordine](https://broadrouter.com/it/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): Taglia la bolletta delle API LLM nell’ordine giusto: misura, metti in cache il prefisso, accorcia i prompt, limita output e ragionamento, scegli il modello giusto. - [Failover API LLM: passare a un secondo provider nel tuo codice](https://broadrouter.com/it/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): Come fare failover tra provider LLM compatibili con OpenAI: quali errori lo attivano, circuit breaker, mappatura degli ID modello, streaming, costi e privacy. - [Tagli di prezzo delle API LLM 2023–2026: i più grandi, con le fonti](https://broadrouter.com/it/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): I maggiori tagli di prezzo delle API LLM dal 2023 al 2026, con link d’archivio prima e dopo: o3, Codestral, Gemini 1.5, più gli aumenti passati quasi inosservati. - [I prezzi delle API LLM scendono e basta? 52 aumenti e come gestirli](https://broadrouter.com/it/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): I prezzi delle API LLM salgono eccome: 52 aumenti di listino sullo stesso modello nelle pagine prezzi archiviate, 2024–2026, mediana 40%. Fonti e regola di budget. - [Rate limit delle API LLM per provider: RPM, TPM e concorrenza](https://broadrouter.com/it/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): I rate limit di base delle API LLM di 13 provider, dalla loro documentazione: OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks e altri. E cosa sono RPM e TPM. - [SLA delle API LLM: cosa promettono Azure, Bedrock, Google e OpenAI](https://broadrouter.com/it/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): A cosa si impegna per iscritto ogni provider di API LLM: percentuale di uptime, cosa conta come downtime, crediti di servizio e target di latenza, dalle pagine SLA. - [Elaborazione batch con LLM senza Batch API: un runner riprendibile](https://broadrouter.com/it/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Esegui da 10.000 a 1.000.000 di prompt su un’API chat completions senza Batch API: concorrenza, ritmo delle richieste, retry, checkpoint riprendibili e tempi. - [La Safety Map: come 39 LLM gestiscono i dati sensibili](https://broadrouter.com/it/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet ha analizzato 39 LLM di 14 laboratori su credenziali, PII, dati sanitari e finanziari: cosa si trasferisce, cosa no, e le fughe da tool call e ragionamento. - [Testare un LLM sulla sicurezza dei dati sensibili: il metodo BroadNet](https://broadrouter.com/it/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): Il metodo pubblicato da BroadNet per verificare prima del deploy se un LLM fa trapelare credenziali, PII, dati sanitari o finanziari: categorie, leak e miss, 3+ run. - [Migrare dall’API di OpenAI: il codice in Python, Node e cURL](https://broadrouter.com/it/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): Le modifiche al codice per passare da OpenAI a un’API compatibile: client, da Responses a Chat Completions, strumenti, streaming, embedding e uno smoke test. - [I modelli che BroadNet ha valutato e scartato, e perché](https://broadrouter.com/it/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet pubblica i modelli che ha testato e scartato, con i motivi: modelli chat bocciati al test sulle credenziali, modelli di voce, OCR ed embedding battuti. - [Progettare un’esperienza AI in più lingue](https://broadrouter.com/it/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): Progetta un prodotto AI per utenti arabi, inglesi e spagnoli: locale dell’interfaccia e lingua del modello, RTL, testo misto, numeri, date, rilevamento e costi. - [Costo API: modelli open-weight o chiusi? Cosa dicono i prezzi](https://broadrouter.com/it/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): Via API i modelli open-weight costano meno di GPT, Claude o Gemini? Prezzi per token dalle pagine dei provider, differenze tra host e un costo per utente calcolato. - [Storico prezzi API di OpenAI, Anthropic e Google, 2023–2026](https://broadrouter.com/it/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): Prezzi delle API di GPT, Claude e Gemini a confronto dalle pagine prezzi archiviate, 2023–2026: quali linee sono scese, quali salite, ciascuna con link all’archivio. - [Base URL delle API compatibili con OpenAI: 14 provider verificati](https://broadrouter.com/it/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): La base_url di 14 API LLM compatibili con OpenAI, presa dalla documentazione di ogni provider, con gli endpoint elencati e le differenze che bloccano il cambio. - [OpenClaw e Hermes Agent con un modello economico: setup e costi](https://broadrouter.com/it/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): Collega OpenClaw o Hermes Agent a un endpoint compatibile con OpenAI, scopri perché i controlli pianificati bruciano token e come tagliare la bolletta. - [Un endpoint privato parte da un carico di lavoro chiaro](https://broadrouter.com/it/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): Che cosa definire prima di chiedere un endpoint LLM privato: carico di lavoro, token al secondo, concorrenza, contesto, regione, licenza e budget, e come testarlo. - [Prompt injection negli agenti di coding: le difese che reggono](https://broadrouter.com/it/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): Dove la prompt injection raggiunge un agente di coding e le difese a più livelli documentate da vendor e ricerca: sandbox, limiti di rete, allow-list, approvazioni. - [Chunk size e overlap nella RAG: effetti su retrieval e costi](https://broadrouter.com/it/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): Come chunk size e overlap cambiano recall, dimensione dell’indice, costo degli embedding e token inviati all’LLM a ogni query, con risultati pubblicati e calcoli. - [Impostare max_tokens per i modelli di ragionamento: un metodo](https://broadrouter.com/it/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Perché i modelli di ragionamento danno risposte vuote con un max_tokens basso, e come dimensionarlo con percentili misurati, finestra di contesto e retry crescenti. - [Registrare audio nel browser e inviarlo a un’API speech-to-text](https://broadrouter.com/it/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): Registra il microfono con MediaRecorder, converti WebM o l’MP4 di Safari in WAV a 16 kHz e invialo all’API di trascrizione via un piccolo server con la chiave. - [Ridurre i token del prompt: tagli misurati su sistema, tool e dati](https://broadrouter.com/it/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): Risparmi di token misurati sfoltendo prompt di sistema, esempi few-shot, schemi dei tool, formati dei dati, cronologia e diacritici arabi, con tokenizer reali. - [Stesso modello, host diverso: quanto variano i prezzi delle API LLM](https://broadrouter.com/it/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): Su 119 modelli venduti da più host, il prezzo di listino più caro è in mediana 1,76 volte il più economico. Dati dalle pagine prezzi di 41 provider, novembre 2026. - [PDPL saudita ed emiratina e API LLM: checklist per il trasferimento](https://broadrouter.com/it/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Inviare dati personali a un’API LLM da Arabia Saudita o EAU: regole PDPL sui trasferimenti, domande al fornitore, quando serve l’inferenza locale. - [Testo, voce o vision? Parti dal compito](https://broadrouter.com/it/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): Scegli tra chat, input di immagini e trascrizione audio in base al compito, con i modelli e i limiti documentati da BroadNet e ciò che non offre, come il TTS. - [Trascrivere audio lunghi oltre 25 MB: dividere, inviare, ricucire](https://broadrouter.com/it/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): Come trascrivere registrazioni oltre il limite di un’API speech-to-text: riduci il file, dividilo sui silenzi con ffmpeg, invia i pezzi in parallelo e ricuci. - [Stringhe di un’app in arabo con un LLM: placeholder e plurali](https://broadrouter.com/it/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): Traduci in arabo i file JSON i18n con un’API LLM: blocca i placeholder, genera le sei forme plurali, valida la sintassi ICU in CI e isola i nomi inseriti. - [Tradurre audio arabo in testo inglese con un’API: una pipeline testata](https://broadrouter.com/it/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): Trasforma registrazioni arabe in testo inglese con due chiamate API: trascrivi con language=ar, poi traduci con un LLM, con controllo dei numeri, costi e punteggio. - [Live, su richiesta, ricerca: leggere bene un catalogo di modelli](https://broadrouter.com/it/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): Come capire quali modelli di un catalogo puoi chiamare oggi, quali richiedono una richiesta e quali sono solo testati, con il catalogo BroadNet e /v1/models. - [Uno in vLLM: che cosa ha misurato lo studio di BroadNet](https://broadrouter.com/it/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet ha integrato in vLLM l’adapter di self-drafting Uno e l’ha misurato su una H100: 2,58x il throughput con uno stream, 1,74x con 32. Metodo, DOI e limiti. - [Uno su Gemma 4 26B: lo studio di BroadNet sul mixture of experts](https://broadrouter.com/it/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet ha addestrato il primo adapter Uno di self-drafting per un modello mixture of experts, Gemma 4 26B A4B. Che cosa ha misurato, quando conviene e il DOI. - [Budget di token di un agente: dove vanno i token di un ciclo](https://broadrouter.com/it/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): System prompt, schemi dei tool, letture di file, cronologia e ragionamento: come spende token un agente, come misurare ogni parte e cosa tagliare prima. - [Word error rate (WER) nell’ASR in arabo: guida pratica](https://broadrouter.com/it/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): Cosa significa word error rate, come si calcola e perché il WER in arabo è così sensibile a diacritici, normalizzazione e dialetti. Con Python e numeri reali. - [Dialetti arabi e speech-to-text: Golfo, Egitto, Levante, Maghreb](https://broadrouter.com/it/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Perché lo speech-to-text fatica con l’arabo del Golfo, egiziano, levantino e maghrebino, e come creare un test set per dialetto sul tuo audio reale. - [Traduzione arabo–inglese con gli LLM: prompt, glossari, valutazione](https://broadrouter.com/it/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): Traduci dall’arabo all’inglese e viceversa con un’API LLM: prompt che mantengono il tono, glossari per i termini fissi, numeri e nomi, e come valutare la qualità. - [Arabo del Golfo, egiziano e levantino: gli LLM e i dialetti](https://broadrouter.com/it/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): Come se la cavano gli LLM con l’arabo del Golfo, egiziano e levantino: dove sbagliano, come scrivere il prompt, come testarli e cosa riporta BroadNet su Qwen3.8 27B. - [OCR arabo o LLM di visione: leggere documenti e fatture in arabo](https://broadrouter.com/it/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): Quando usare l’OCR classico e quando un LLM di visione per fatture, scontrini e moduli in arabo: trappole, output JSON, costo per pagina e validazione dei dati. - [Prompt engineering in arabo: registro, RTL, cifre e testo misto](https://broadrouter.com/it/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): Prompt per ottenere un buon arabo da un LLM: registro e dialetto, cifre e date, formattazione RTL pulita e gestione del mix tra arabo e inglese. - [RAG in arabo: normalizzazione, chunking e retrieval che funzionano](https://broadrouter.com/it/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): RAG su documenti in arabo: normalizza alef, ya e ta marbuta, dividi in chunk con criterio, scegli embedding multilingue e aggiungi la ricerca per parole chiave. - [API speech-to-text per l’arabo: come sceglierla nel 2026](https://broadrouter.com/it/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): Come scegliere un’API di trascrizione per l’arabo: WER, dialetti, limiti dei file, prezzo per ora di audio e privacy, più un piano di test sulle tue registrazioni. - [Architettura di un chatbot WhatsApp in arabo: dai vocali alle risposte](https://broadrouter.com/it/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): Un’architettura funzionante per un assistente WhatsApp in arabo: webhook, coda, trascrizione dei vocali, risposta LLM, regole di policy e costi, con codice Python. - [Arabizi e LLM: gestire il franco-arabo con numeri come 3 e 7](https://broadrouter.com/it/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): L’Arabizi scrive l’arabo in caratteri latini con cifre come 3 per ع e 7 per ح. Come rilevarlo, convertirlo in caratteri arabi con un LLM e rispondere bene. - [Miglior LLM per Aider: come scegliere il modello e collegarlo](https://broadrouter.com/it/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Come scegliere il miglior LLM per Aider: rispetto del formato di modifica, contesto e costo orario, un test di un’ora sul tuo repo e OPENAI_API_BASE. - [Miglior modello aperto per agenti di coding: tool, contesto e costo](https://broadrouter.com/it/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B o Gemma 4 per il tuo agente di coding? Scegli per tool calling, contesto, ragionamento e costo per attività, con i conti. - [Token di input in cache: come ridurre la bolletta delle API LLM](https://broadrouter.com/it/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): L’input in cache costa una frazione dell’input normale. Come si paga il prompt caching, come leggere cached_tokens in usage e come strutturare i prompt. - [L’API LLM vision più economica: costo per immagine e per pagina](https://broadrouter.com/it/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): Quanto costa un LLM vision per immagine e per pagina: calcoli sui token per Gemma 4, Qwen3.8 27B e DeepSeek V4.1 Flash, e dove DeepInfra costa meno. - [Come scegliere un provider di API LLM: checklist in 20 punti](https://broadrouter.com/it/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): Venti domande da fare prima di scegliere un’API LLM: prezzi, limiti, compatibilità OpenAI, privacy, modelli, supporto e uptime, più punteggi e piano di test. - [Scegliere un’API LLM per l’arabo: cosa testare](https://broadrouter.com/it/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): Come scegliere un LLM per app in arabo: dialetti, qualità di scrittura, traduzione, documenti e immagini, costi, e un test set da creare in un giorno. - [Cline vs Roo Code vs Aider per team con modelli open (2026)](https://broadrouter.com/it/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (ora archiviato, continua come Zoo Code) e Aider a confronto su ciò che conta con i modelli open: tool, contesto, MCP, git e approvazioni. - [Come contare i token prima di inviare una richiesta a un LLM](https://broadrouter.com/it/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): Stima i token del prompt per lingua e modello prima di chiamare un’API LLM: perché i tokenizer cambiano, esempi in cinque lingue e margini per max_tokens. - [10 modi per ridurre i costi LLM degli agenti di coding, con i risparmi](https://broadrouter.com/it/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Dieci modi concreti per abbassare la spesa in token di Cline, Aider, opencode o del tuo agente di coding, ognuno con i conti del risparmio e come misurarlo. - [Alternative all’API DeepSeek: 5 modi per usare i modelli DeepSeek](https://broadrouter.com/it/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): Usare i modelli DeepSeek con l’API ufficiale, un aggregatore, il catalogo di un cloud, GPU tue o BroadNet Inference: pro e contro su prezzo, privacy e limiti. - [Prezzi API DeepSeek V4.1 Flash: per token, input in cache e piani](https://broadrouter.com/it/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): Prezzi API DeepSeek V4.1 Flash per 1M di token (input, in cache, output), piani flat mensili, settimanali e orari e confronto con DeepSeek, DeepInfra e Fireworks. - [Come stimare il costo delle API LLM prima di lanciare una funzione](https://broadrouter.com/it/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): Modello di costo passo per passo per una nuova funzione AI: utenti × azioni × chiamate × token, con cache, ragionamento, retry, rate limit e un esempio mensile. - [Come valutare un LLM prima di cambiare provider](https://broadrouter.com/it/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): Crea un piccolo eval set dal traffico reale, valutalo con controlli nel codice e un LLM judge senza bias, e confronta qualità, costo e latenza prima di migrare. - [Estrarre dati da fatture, scontrini e moduli in JSON con un LLM vision](https://broadrouter.com/it/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): Pipeline di produzione per fatture, scontrini e moduli: preparazione immagini, prompt con schema JSON, validazione Pydantic, controlli sui totali, retry e revisione. - [API LLM a tariffa flat o a consumo per token: i conti del pareggio](https://broadrouter.com/it/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): Quando un piano LLM flat batte il pagamento per token? Quattro bollette mensili d’esempio per DeepSeek V4.1 Flash, il volume di pareggio e come misurare il tuo. - [Credito API LLM gratuito: quanto dura davvero il credito iniziale](https://broadrouter.com/it/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): Cosa compra un credito API LLM gratuito: token, minuti di audio e chiavi orarie con il credito iniziale BroadNet, rispetto ai free tier di Gemini e OpenRouter. - [GDPR e API LLM: 15 domande da fare al tuo provider](https://broadrouter.com/it/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): Checklist GDPR per le API LLM: DPA, sub-responsabili, conservazione, addestramento, trasferimenti extra UE e notifica dei data breach, con gli articoli del GDPR. - [GLM 5.3 o DeepSeek V4.1 Flash: quale usare e quando](https://broadrouter.com/it/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 e DeepSeek V4.1 Flash a confronto su prezzo, contesto, tool, output strutturato, immagini e ragionamento, con costi d’esempio e una regola per scegliere. - [Errore 429 nelle API LLM: backoff, jitter e limitatori](https://broadrouter.com/it/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 o 5xx, backoff esponenziale con full jitter, limitatori di concorrenza e code per le API LLM, con codice Python e Node pronto per la produzione. - [Chiave API LLM oraria: un prezzo fisso all’ora, spiegato](https://broadrouter.com/it/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): La chiave oraria di BroadNet compra un’ora di Gemma 4 VL 26B su un solo stream a prezzo fisso. Cosa include, i limiti, a chi serve e quando conviene. - [Conservazione dati nelle API LLM: OpenAI, Anthropic, Gemini, Mistral](https://broadrouter.com/it/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): Che cosa dicono OpenAI, Anthropic, Gemini API, Mistral e BroadNet su conservazione dei prompt API e addestramento, citati dalle pagine ufficiali (novembre 2026). - [Chiavi API LLM sicure: conservazione, rotazione e IP consentiti](https://broadrouter.com/it/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): Chiavi API LLM fuori da codice e log, una chiave per ogni carico di lavoro, rotazione regolare, solo gli IP dei tuoi server e revoca di una fuga in pochi secondi. - [Prezzi delle API LLM spiegati: input, cache, output e ragionamento](https://broadrouter.com/it/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): Come funzionano i prezzi per token delle API LLM: input, input in cache, output, token di ragionamento, immagini e audio al minuto, con esempi su un listino reale. - [Dati sensibili e API LLM: oscura, minimizza, registra meno](https://broadrouter.com/it/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): Come inviare dati di clienti, sanitari o finanziari a un’API LLM in sicurezza: classifica i campi, usa segnaposto reversibili, zero data retention e log puliti. - [Controllo della spesa API LLM: budget, limiti sulle chiavi e cicli](https://broadrouter.com/it/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): Evita che la bolletta delle API LLM esploda: saldo prepagato, una chiave per app, IP consentiti, log dell’uso, budget nel codice, limiti ai cicli, chiavi flat. - [Log di utilizzo delle API LLM: costo per utente e funzione, senza PII](https://broadrouter.com/it/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): Registra ogni chiamata API LLM con token, costo, latenza e request ID, attribuisci la spesa a utenti e funzioni e oscura le PII prima di scrivere. Codice Python. - [Costo LLM per utente: unit economics di una funzione AI in un SaaS](https://broadrouter.com/it/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): Quanto costa una funzione AI per utente attivo al mese, perché gli utenti intensivi contano più della media e come prezzare i piani SaaS per coprire il modello. - [Prezzi dei modelli open-weight: 6 provider di inferenza a confronto](https://broadrouter.com/it/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): Come OpenRouter, Together AI, Fireworks, Groq, DeepInfra e BroadNet prezzano i modelli open-weight: per token, cache, batch, ore GPU e piani flat, con fonti. - [Latenza degli LLM: TTFT, token al secondo e tempo end-to-end](https://broadrouter.com/it/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): Cosa misurano time to first token, token al secondo e latenza end-to-end, come cronometrarli dal client con il codice e quali interventi spostano quale numero. - [Streaming degli LLM spiegato: chunk SSE, delta, usage e annullamento](https://broadrouter.com/it/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Come funziona lo streaming delle chat completion, come ricostruire la risposta dai delta, leggere usage dall’ultimo chunk e annullare in Python, Node e curl. - [Contesto lungo o RAG: quando 1M di token sostituisce un DB vettoriale](https://broadrouter.com/it/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): Quando una finestra da 1M di token può sostituire un database vettoriale e quando vince ancora la RAG: costo per query, latenza, recall e manutenzione. - [Contesto degli agenti a lungo termine con una finestra da 1M](https://broadrouter.com/it/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): Come mantenere precise ed economiche le sessioni lunghe di un agente con una finestra da 1M di token: storico append-only, cache, quando compattare e cosa tagliare. - [Agenti di lunga durata: timeout, retry e checkpoint che reggono](https://broadrouter.com/it/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): Tenere in vita per ore un agente LLM: timeout per chiamata, streaming per scoprire gli stalli, retry sicuri, side effect idempotenti e un checkpoint a ogni passo. - [Server MCP con modelli compatibili con OpenAI, di qualsiasi provider](https://broadrouter.com/it/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP funziona con qualsiasi LLM: il client trasforma i tool MCP in function call. Come usare MCP con modelli aperti in Cline, Continue, opencode, n8n o in Python. - [Quando conviene una finestra di contesto da 1M di token (e quando no)](https://broadrouter.com/it/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): Cosa sta in 1.048.576 token, quanto costa su DeepSeek V4.1 Flash, come ragionamento e cache cambiano i conti e quando la RAG resta la scelta migliore. - [Quanto costano davvero i sistemi multi-agente: token o sessioni flat](https://broadrouter.com/it/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): I sistemi multi-agente consumano molte volte i token di un solo agente: i conti tra sessioni flat e token, i moltiplicatori nascosti e quando conviene cosa. - [API speech-to-text multilingue: 14 lingue, un solo endpoint](https://broadrouter.com/it/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): Trascrivi inglese, spagnolo, francese, cinese, arabo e altre 9 lingue con un endpoint compatibile OpenAI a prezzo al minuto, e quando usare il modello arabo. - [AI Agent n8n con un LLM economico: setup, modello e costo a esecuzione](https://broadrouter.com/it/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): Workflow con l’AI Agent di n8n su un modello economico compatibile con OpenAI: quale modello per i tool, le impostazioni che tagliano token e i conti su 10.000 run. - [Modelli AI open-weight o chiusi in azienda: costi, controllo, rischi](https://broadrouter.com/it/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): LLM open-weight o chiusi in azienda? Costi, controllo, privacy, divario di qualità e lock-in, con i tre modi di usare i modelli aperti e quando conviene ciascuno. - [Passare a un’API compatibile con OpenAI: la checklist di migrazione](https://broadrouter.com/it/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): Cambiare base_url è la parte facile. Si rompono ID dei modelli, token di ragionamento, usage in streaming, errori, rate limit e chiavi. Una checklist con codice. - [OpenRouter o piano flat per agenti AI: quale costa meno?](https://broadrouter.com/it/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter fattura per token più una commissione sui crediti, un piano flat a sessione. Costi per agenti sempre attivi, commissioni, limiti e quale scegliere. - [Come far lavorare gli agenti di coding di notte, in sicurezza](https://broadrouter.com/it/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): Configurazione pratica per agenti di coding senza supervisione: sandbox, checkpoint git, file di task, limiti di passi, revisione al mattino e sessioni flat. - [Come strutturare i prompt perché il prefisso vada in cache](https://broadrouter.com/it/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): Come organizzare i prompt LLM perché il prefisso in cache resti identico: sistema, schemi dei tool, documenti, cronologia solo in aggiunta e un test che lo prova. - [Token di ragionamento: perché paghi più output di quanto ne vedi](https://broadrouter.com/it/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): Il ragionamento nascosto si paga come output. Come misurarlo, quanto costa per modello e come ridurlo con reasoning_effort, max_tokens e routing. - [Come far girare agenti di coding in parallelo con un’API LLM flat](https://broadrouter.com/it/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Sub-agenti, orchestratori e worker batch su DeepSeek V4.1 Flash illimitato: quante sessioni servono, come dividerle tra chiavi e come limitare la concorrenza. - [LLM in self-hosting o via API: costo reale e punto di pareggio](https://broadrouter.com/it/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): Prezzi di noleggio GPU dalle pagine ufficiali, formula del costo per 1M di token sulla tua GPU, utilizzo necessario per battere un’API e costi operativi dimenticati. - [Prezzi delle API speech-to-text a confronto: costo al minuto (2026)](https://broadrouter.com/it/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): Prezzi al minuto e per 1.000 ore di OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq e BroadNet Transcribe dalle pagine ufficiali, più i costi extra. - [JSON mode o json_schema: output strutturato che regge in produzione](https://broadrouter.com/it/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON mode dà JSON valido; json_schema dà la tua struttura. Come richiederli sui modelli compatibili con OpenAI, validare con Pydantic e correggere in un loop. - [Token per parola in arabo e in inglese: perché l’arabo costa di più](https://broadrouter.com/it/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): Perché l’arabo usa spesso più token LLM dell’inglese a parità di significato, perché i token per parola ingannano, e uno script per misurare il tuo rapporto reale. - [Tool calling con modelli aperti: schemi, chiamate parallele, retry](https://broadrouter.com/it/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): Function calling su modelli aperti compatibili con OpenAI: schemi, chiamate parallele, i cinque modi in cui falliscono e un ciclo Python che valida e riprova. - [Trascrivere le chiamate del call center: pipeline pronta per l’arabo](https://broadrouter.com/it/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): Dividi le chiamate stereo per canale, taglia i turni, trascrivi arabo e inglese, oscura i dati personali e riassumi ogni chiamata con un LLM. Codice Python e costi. - [Modelli aperti negli agenti di coding: Aider, Cline, opencode, Codex](https://broadrouter.com/it/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Collega Aider, Cline, Roo Code, opencode, Continue e Codex CLI a qualsiasi endpoint compatibile con OpenAI. I nomi esatti dei campi e le impostazioni che si rompono. - [Trascrivere i vocali per le aziende: da WhatsApp e Telegram ai ticket](https://broadrouter.com/it/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): Trasforma i messaggi vocali dei clienti da WhatsApp Business e bot Telegram in trascrizioni e ticket: download, trascrizione, campi estratti con un LLM, smistamento. - [Un piano LLM settimanale per progetti brevi: Orbit, ora o mese](https://broadrouter.com/it/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): Quando un piano illimitato DeepSeek V4.1 Flash di una settimana batte la sessione oraria, il piano mensile o il pagamento a token: conti per hackathon e migrazioni. - [Cosa significa davvero «illimitato» in un piano API LLM](https://broadrouter.com/it/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Contatori di token, finestre mobili di 5 ore, tetti settimanali, rate limit e concorrenza. Sette domande per capire se un piano «illimitato» lo è per il tuo agente. - [Alternative all’API Whisper: quando cambiare e come provarle](https://broadrouter.com/it/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): Quando ha senso un’alternativa ospitata all’API Whisper di OpenAI: word error rate sul tuo audio, lingue, prezzo all’ora, limiti e privacy, con codice di test. - [API LLM con zero data retention: una checklist di 10 domande](https://broadrouter.com/it/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): Cosa dovrebbe significare «zero data retention» per un’API LLM: prompt, output, log, addestramento, sub-responsabili e verifiche. 10 domande prima di inviare dati. - [Guides in Italian](https://broadrouter.com/it/guides/) - [Usare l’SDK Python di OpenAI con BroadNet Inference](https://broadrouter.com/it/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): Collega l’SDK Python ufficiale di OpenAI a BroadNet Inference cambiando solo il base_url. Installazione, chiave, prima richiesta e risposta in streaming. - [Usare l’SDK Node.js di OpenAI con BroadNet Inference](https://broadrouter.com/it/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): Collega l’SDK ufficiale di OpenAI per Node.js e TypeScript a BroadNet Inference impostando baseURL. Installazione, prima richiesta, streaming e ID delle richieste. - [Provare BroadNet Inference nel browser, senza account](https://broadrouter.com/it/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): Prova lo speech-to-text in arabo, la lettura di documenti e la chat con Gemma 4 nel playground di BroadNet, senza registrazione né carta, poi passa all’API. - [Prima richiesta a BroadNet Inference con curl](https://broadrouter.com/it/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): Le prime chiamate a BroadNet Inference con curl: una chat completion, una risposta in streaming e una trascrizione audio, più come leggere risposta, usage ed errori. - [Streaming delle chat completions e connessioni interrotte](https://broadrouter.com/it/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): Lo streaming su BroadNet Inference: Server-Sent Events, il marcatore [DONE], l’uso dei token nell’ultimo chunk e cosa fare se lo stream si chiude prima della fine. - [Perché un modello di ragionamento restituisce una risposta vuota](https://broadrouter.com/it/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3 e Qwen3.8 27B ragionano prima di rispondere, e il ragionamento conta in max_tokens. Come evitare le risposte vuote. - [Codici di errore di BroadNet Inference e quando ritentare](https://broadrouter.com/it/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): Cosa significa ogni stato HTTP dell’API di BroadNet Inference, quali ritentare con backoff, quali correggere prima e cosa mandare all’assistenza con l’ID richiesta. - [Limita le chiavi API ai tuoi server con le allowlist di IP](https://broadrouter.com/it/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): Proteggi le chiavi di BroadNet Inference con le allowlist di IP, separa chiavi di inferenza e credenziali del portale, revoca subito una chiave esposta. Con curl. - [Come funzionano i piani API LLM illimitati per gli agenti](https://broadrouter.com/it/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): I piani illimitati DeepSeek V4.1 Flash non hanno contatore di token né limite settimanale. Ecco come funzionano sessioni, agenti in parallelo, chiavi e fatturazione. - [Usare Aider con DeepSeek V4.1 Flash illimitato](https://broadrouter.com/it/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Collega l’agente di coding Aider a BroadNet Inference con l’impostazione compatibile con OpenAI e usalo con un piano flat illimitato per DeepSeek V4.1 Flash. - [Usare LangChain con BroadNet Inference](https://broadrouter.com/it/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): Collega ChatOpenAI di LangChain a BroadNet Inference con base_url e api_key, e scopri cosa LangChain conserva e cosa scarta dei modelli di ragionamento. - [Usare Cline con BroadNet Inference (provider OpenAI Compatible)](https://broadrouter.com/it/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Configura l’estensione Cline per VS Code con BroadNet Inference tramite il provider OpenAI Compatible: campi, ID dei modelli, contesto e soluzioni ai problemi. - [Usare Roo Code con BroadNet Inference (OpenAI Compatible)](https://broadrouter.com/it/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code ha chiuso a maggio 2026. Le impostazioni per le installazioni esistenti e i fork, e dove migrare (Cline o il fork Zoo Code) con BroadNet Inference. - [Usare Continue con BroadNet Inference (config.yaml e apiBase)](https://broadrouter.com/it/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): Aggiungi i modelli di BroadNet Inference all’estensione Continue con provider openai e apiBase in config.yaml, più capability tool_use e lunghezza del contesto. - [Aggiungere BroadNet Inference a opencode come provider personalizzato](https://broadrouter.com/it/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): Aggiungi BroadNet Inference all’agente di coding opencode come provider personalizzato compatibile con OpenAI in opencode.jsonc, con alias, capacità e limiti. - [Collegare Open WebUI a BroadNet Inference](https://broadrouter.com/it/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Aggiungi BroadNet Inference a Open WebUI come connessione compatibile con OpenAI, metti gli ID dei modelli in allowlist e risolvi verifica, risposte vuote e 429. - [Aggiungere BroadNet Inference a LibreChat come custom endpoint](https://broadrouter.com/it/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): Aggiungi BroadNet Inference a LibreChat come custom endpoint compatibile OpenAI in librechat.yaml, con chiave nel .env, elenco dei modelli e title model. - [Usare BroadNet Inference in n8n (credenziale OpenAI, Base URL)](https://broadrouter.com/it/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): Punta la credenziale OpenAI di n8n su BroadNet Inference con il campo Base URL, poi usa il nodo OpenAI Chat Model per ID con Use Responses API disattivato. - [Usare il Vercel AI SDK con BroadNet Inference](https://broadrouter.com/it/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): Chiama BroadNet Inference dal Vercel AI SDK con createOpenAICompatible di @ai-sdk/openai-compatible: generateText, streamText, strumenti e reasoning effort. - [API speech-to-text in arabo, passo dopo passo](https://broadrouter.com/it/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): Trascrivi audio in arabo con BroadNet Transcribe Arabic tramite un endpoint compatibile con OpenAI. Formati, campo language, esempi Python e curl, e il prezzo. - [Instradare LiteLLM verso BroadNet Inference (SDK e proxy)](https://broadrouter.com/it/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): Chiama BroadNet Inference tramite LiteLLM con il prefisso openai/ e api_base, nell’SDK Python o nel config.yaml del proxy, con rate limit allineati. - [Usare l’OpenAI Agents SDK con BroadNet Inference](https://broadrouter.com/it/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Usa l’OpenAI Agents SDK per Python con BroadNet Inference tramite AsyncOpenAI e OpenAIChatCompletionsModel, e disattiva il tracing che richiede una chiave OpenAI. - [Trascrivere i vocali WhatsApp in testo con un’API](https://broadrouter.com/it/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): Converti i vocali WhatsApp .ogg o le registrazioni .m4a del telefono in WAV con ffmpeg e trascrivili in arabo o in altre 13 lingue con una sola chiamata API. - [Trascrivere i vocali WhatsApp in n8n](https://broadrouter.com/it/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): Un workflow n8n che riceve un vocale WhatsApp, lo scarica da Meta, lo converte con ffmpeg e lo trascrive in arabo o in altre 13 lingue. - [Leggere fatture e scontrini in arabo con un modello vision](https://broadrouter.com/it/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Invia la foto di una fattura o di uno scontrino a un modello vision su BroadNet Inference e ricevi totale, data e fornitore in JSON, con domande in arabo o inglese. - [Usare Zed con BroadNet Inference come provider compatibile OpenAI](https://broadrouter.com/it/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Aggiungi BroadNet Inference a Zed come provider compatibile con OpenAI in settings.json, con ID dei modelli, limiti di token, capabilities e variabile della chiave. - [Usare Kilo Code con BroadNet Inference (provider OpenAI Compatible)](https://broadrouter.com/it/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Collega Kilo Code a BroadNet Inference come provider OpenAI Compatible personalizzato, dalla finestra di VS Code o da kilo.json, con ID, limiti e flag dei tool. - [Aggiungere BroadNet Inference a goose come provider personalizzato](https://broadrouter.com/it/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): Collega l’agente AI goose a BroadNet Inference come custom provider OpenAI Compatible, da goose Desktop, con goose configure o con un file JSON. - [Usare Pydantic AI con BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/it/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Esegui agenti Pydantic AI su BroadNet Inference con OpenAIChatModel e OpenAICompatibleProvider, con tool call, output strutturato e correzioni del profilo. - [CrewAI con BroadNet Inference come endpoint OpenAI personalizzato](https://broadrouter.com/it/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): Collega gli agenti CrewAI a BroadNet Inference con LLM(custom_openai=True, base_url=...), poi scegli i modelli per strumenti, contesto lungo e piano flat. - [Usare LlamaIndex con BroadNet Inference tramite OpenAILike](https://broadrouter.com/it/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): Collega LlamaIndex a BroadNet Inference con la classe OpenAILike, impostando bene api_base, context_window, is_chat_model e is_function_calling_model. - [BroadNet Inference in Dify con il plugin OpenAI-API-compatible](https://broadrouter.com/it/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): Collega Dify a BroadNet Inference con il plugin OpenAI-API-compatible: valori esatti dei campi, dimensioni del contesto e impostazioni del function calling. - [Usare Jan con BroadNet Inference come provider compatibile OpenAI](https://broadrouter.com/it/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Aggiungi BroadNet Inference all’app desktop Jan come provider personalizzato compatibile con OpenAI, inserisci gli ID dei modelli e attiva strumenti e visione. - [Collegare SillyTavern a BroadNet Inference (Custom OpenAI-compatible)](https://broadrouter.com/it/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): Collega SillyTavern a BroadNet Inference con la sorgente Custom (OpenAI-compatible): ID dei modelli giusti, post-elaborazione del prompt e limiti di token. - [Usare Spring AI con BroadNet Inference (base-url OpenAI)](https://broadrouter.com/it/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Punta lo starter chat OpenAI di Spring AI su BroadNet Inference con spring.ai.openai.base-url, un ID di modello BroadNet e max-tokens, poi chiamalo da ChatClient. - [Chiamare BroadNet Inference da Go con l’SDK openai-go](https://broadrouter.com/it/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): Usa l’SDK ufficiale openai-go v3 con BroadNet Inference tramite option.WithBaseURL: chat completions, streaming con usage e gestione degli errori in Go. - [Creare agenti LangGraph su BroadNet Inference](https://broadrouter.com/it/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): Esegui uno StateGraph o create_agent di LangGraph su BroadNet Inference con ChatOpenAI, base_url e un ID di modello BroadNet, con strumenti, streaming e memoria. - [Usare pipeline e agenti Haystack con BroadNet Inference](https://broadrouter.com/it/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Punta OpenAIChatGenerator di Haystack su BroadNet Inference con api_base_url, un ID di modello BroadNet e una chiave Secret: pipeline, strumenti, streaming, agenti. - [Usare Semantic Kernel con BroadNet Inference](https://broadrouter.com/it/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Collega Microsoft Semantic Kernel a BroadNet Inference: in Python con un client AsyncOpenAI, in C# con un endpoint personalizzato, per chat, plugin e agenti. - [Usare gli agenti Mastra con BroadNet Inference](https://broadrouter.com/it/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): Esegui agenti Mastra su BroadNet Inference con un oggetto model che imposta id, url e apiKey, compreso il doppio prefisso broadnet/ richiesto dall’ID del modello. - [Output strutturato con Instructor su BroadNet Inference](https://broadrouter.com/it/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Estrai oggetti Pydantic da BroadNet Inference con Instructor, usando from_openai o from_provider con un base_url personalizzato, e scegli il Mode giusto. - [Usare DSPy con BroadNet Inference (imposta model_type="chat")](https://broadrouter.com/it/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): Configura dspy.LM per BroadNet Inference con il prefisso openai/, api_base e model_type="chat", perché DSPy 3.5 manda i modelli OpenAI alla Responses API. - [Valutare i modelli BroadNet con promptfoo](https://broadrouter.com/it/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): Esegui eval con promptfoo e confronti affiancati tra modelli su BroadNet Inference con il provider openai:chat, apiBaseUrl e apiKeyEnvar, più i grader. - [Usare smolagents su BroadNet Inference con OpenAIModel](https://broadrouter.com/it/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Collega smolagents di Hugging Face a BroadNet Inference con OpenAIModel e api_base: CodeAgent su qualsiasi modello, ToolCallingAgent su GLM 5.3 e Qwen3.8 27B. - [Collegare Flowise a BroadNet Inference (OpenAI Custom Model)](https://broadrouter.com/it/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Usa BroadNet Inference nei chatflow di Flowise con il nodo OpenAI Custom Model, un Base Path, una credenziale OpenAI API e il nome del modello BroadNet. - [Collegare Langflow a BroadNet Inference](https://broadrouter.com/it/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Usa BroadNet Inference in Langflow con il campo OpenAI API Base del componente OpenAI e un nome di modello digitato, e scopri perché OpenAI Compatible non lo elenca. - [Glossary entries in Italian](https://broadrouter.com/it/glossary/) - [Agente AI (AI agent)](https://broadrouter.com/it/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): Un agente AI è un sistema in cui un modello linguistico dirige da solo i propri passi verso un obiettivo: decide quali strumenti chiamare, legge i risultati e ripete finché produce una risposta finale o raggiunge un limite. Uno script fisso che chiama un modello in punti prestabiliti di solito si chiama invece workflow. - [Chiave API (API key)](https://broadrouter.com/it/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): Una chiave API è una stringa segreta che identifica e autorizza un client che chiama un’API. Le API LLM di solito la vogliono in un header Authorization: Bearer, e chiunque possieda la chiave può fare richieste fatturate all’account a cui appartiene. - [Diacritici arabi (tashkeel)](https://broadrouter.com/it/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): I diacritici arabi, o tashkeel, sono piccoli segni scritti sopra o sotto le lettere per indicare le vocali brevi, le consonanti doppie e l’assenza di vocale. Sono facoltativi: quasi tutto il testo arabo, notizie comprese, li omette, quindi lettori e software deducono le vocali dal contesto. - [Identificazione del dialetto arabo (dialect identification)](https://broadrouter.com/it/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): L’identificazione del dialetto arabo è il compito di stabilire in quale varietà di arabo è un testo o una registrazione, come egiziano, del Golfo, levantino, maghrebino o arabo standard moderno, a livello di regione, paese o città. Serve a instradare i messaggi, scegliere i prompt e dividere i set di valutazione. - [Normalizzazione del testo arabo (Arabic text normalization)](https://broadrouter.com/it/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): La normalizzazione del testo arabo è il passaggio che riporta a una sola forma le varianti grafiche della stessa parola prima del confronto: toglie diacritici e tatweel, unifica le forme di alef e spesso trasforma ى in ي e ة in ه. Si applica alle chiavi di ricerca e alla valutazione, non al testo che mostri. - [Arabizi](https://broadrouter.com/it/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): L’Arabizi è arabo scritto in modo informale con lettere latine, usando cifre per i suoni arabi senza equivalente latino, come 3 per ع. È diffuso in chat e sui social, è per lo più dialettale, non ha un’ortografia standard e spesso si mescola con inglese o francese. - [AWQ (activation-aware weight quantization)](https://broadrouter.com/it/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (activation-aware weight quantization) è un metodo post-training che comprime i pesi di un LLM a bassa precisione, di solito interi a 4 bit. Usa le statistiche delle attivazioni di un piccolo set di calibrazione per trovare i canali di pesi più importanti e li scala perché perdano meno accuratezza. Non richiede backpropagation. - [Base URL (URL di base)](https://broadrouter.com/it/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): La base URL è l’indirizzo radice fisso di un’API, come https://api.example.com/v1, che un client antepone a ogni percorso di endpoint, come /chat/completions. Negli SDK compatibili con OpenAI, cambiare la base URL, insieme alla chiave API e all’ID del modello, punta lo stesso codice verso un altro provider. - [Batch API](https://broadrouter.com/it/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Una batch API è un endpoint LLM asincrono che accetta molte richieste insieme, di solito come file, le elabora entro una finestra prestabilita invece che subito, e restituisce i risultati tutti insieme. Provider come OpenAI e Anthropic applicano alle richieste batch uno sconto rispetto alle loro API standard. - [BM25](https://broadrouter.com/it/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) è una funzione di ranking per la ricerca per parole chiave che assegna un punteggio a un documento in base a quanto spesso vi compaiono i termini della query, a quanto sono rari nella collezione e alla lunghezza del documento. È la similarità predefinita in Lucene ed Elasticsearch. - [Byte pair encoding (BPE)](https://broadrouter.com/it/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Il byte pair encoding (BPE) è un algoritmo di tokenizzazione in sottoparole che parte da singoli caratteri o byte e unisce ripetutamente la coppia adiacente più frequente in una nuova voce del vocabolario, fino a raggiungere la dimensione di vocabolario prevista. La maggior parte dei tokenizer LLM attuali usa BPE o una sua variante a livello di byte. - [Token di input in cache (cached input)](https://broadrouter.com/it/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): I token di input in cache sono token del prompt che un provider LLM ha elaborato da poco e può riusare da una cache invece di ricalcolarli, di solito un prefisso ripetuto come un system prompt o un documento. I provider li fatturano a una tariffa più bassa dei token di input nuovi. - [Chain-of-thought (CoT)](https://broadrouter.com/it/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Il chain-of-thought (CoT, catena di pensiero) è una tecnica di prompting in cui a un modello linguistico si mostra o si chiede di scrivere i passaggi di ragionamento intermedi prima della risposta finale. Migliora l’accuratezza nei problemi a più passaggi, come aritmetica e logica, e i modelli di reasoning ora lo generano in automatico. - [Character error rate (CER)](https://broadrouter.com/it/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Il character error rate (CER, tasso di errore sui caratteri) misura l’accuratezza di una trascrizione o di un OCR a livello di carattere: sostituzioni, cancellazioni e inserimenti necessari per trasformare l’output nel testo di riferimento, divisi per il numero di caratteri del riferimento. Più è basso meglio è, e 0 è una corrispondenza perfetta. - [Chat Completions API](https://broadrouter.com/it/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): La Chat Completions API è l’endpoint HTTP di OpenAI, POST /v1/chat/completions, che riceve il nome di un modello e una lista di messaggi con un ruolo e restituisce il messaggio successivo del modello. La forma della sua richiesta e risposta è diventata il formato comune che riusano provider e SDK compatibili con OpenAI. - [Chat template](https://broadrouter.com/it/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): Un chat template è la regola di formattazione, di solito scritta in Jinja e distribuita con il tokenizer di un modello, che converte una lista di messaggi di chat con i rispettivi ruoli nell’unica sequenza di token che il modello si aspetta, compresi i suoi token speciali di controllo. Usare il template sbagliato può peggiorare molto le risposte del modello. - [Chunking (RAG)](https://broadrouter.com/it/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Il chunking è il passaggio di una pipeline di retrieval che divide i documenti in pezzi più piccoli, detti chunk, prima di trasformarli in embedding e indicizzarli. La chunk size fissa quanto testo contiene ogni pezzo, e il chunk overlap ripete parte del testo tra chunk vicini, così le idee tagliate su un confine non si perdono. - [Code-switching (commutazione di codice)](https://broadrouter.com/it/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Il code-switching è l’uso di due o più lingue, o varietà di una stessa lingua, all’interno della stessa conversazione o frase, nel parlato o nello scritto. Ne sono esempi gli arabofoni che inseriscono parole inglesi nel dialetto, o che passano dall’arabo standard moderno al loro dialetto. - [Agente di coding AI (AI coding agent)](https://broadrouter.com/it/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): Un agente di coding AI è uno strumento software in cui un modello linguistico lavora su una codebase in un ciclo: legge file, li modifica, esegue comandi come i test, legge l’output e decide il passo successivo finché il compito è finito o si raggiunge un limite. - [Concorrenza (concurrency)](https://broadrouter.com/it/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): La concorrenza è il numero di richieste che un client ha in corso nello stesso momento. Le API LLM spesso la limitano per chiave o per piano, separatamente dalle richieste al minuto, perché ogni richiesta aperta occupa capacità GPU finché la sua risposta non è terminata. - [Context compaction (compattazione del contesto)](https://broadrouter.com/it/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): La context compaction è la pratica di riassumere una conversazione che si avvicina al limite della finestra di contesto e di proseguire la sessione da quel riassunto, così un agente conserva decisioni chiave e stato e scarta la cronologia grezza, come i vecchi output dei tool. - [Context engineering (ingegneria del contesto)](https://broadrouter.com/it/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): Il context engineering è il lavoro di scegliere e mantenere l’insieme di token che un modello linguistico vede a ogni richiesta, tra system prompt, definizioni dei tool, cronologia, documenti recuperati e note, così che il modello riceva il minimo di informazioni ad alto segnale che gli servono per svolgere bene il compito. - [Context rot (degrado del contesto)](https://broadrouter.com/it/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): Il context rot è la perdita graduale di accuratezza e affidabilità che i modelli linguistici mostrano quando l’input si allunga, anche se il compito resta lo stesso e l’input è ben dentro la context window. Il termine è stato reso popolare da un report tecnico di Chroma del 2025 che ha testato 18 modelli. - [Finestra di contesto (context window)](https://broadrouter.com/it/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): La finestra di contesto è il numero massimo di token con cui un modello linguistico può lavorare in una singola richiesta. Comprende prompt, cronologia della conversazione, definizioni degli strumenti, immagini e output generato, token di ragionamento inclusi. Ciò che non ci sta va tagliato, riassunto o diviso su più richieste. - [Residenza dei dati (data residency)](https://broadrouter.com/it/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): La residenza dei dati è un requisito o un impegno per cui i dati vengono conservati ed elaborati in un paese o in una regione specifici. Per un’API LLM riguarda dove vengono elaborati i prompt durante l’inferenza e dove si trova l’eventuale contenuto salvato, e spesso esclude i dati di account e fatturazione. - [Arabo dialettale (dialectal Arabic)](https://broadrouter.com/it/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): L’arabo dialettale è l’arabo parlato ogni giorno in una regione o in una città, come l’arabo egiziano, del Golfo, levantino o maghrebino. I dialetti differiscono dall’arabo standard moderno per suoni, parole e grammatica, non hanno un’ortografia standard e si scrivono sempre più online, il che li rende più difficili per i modelli linguistici. - [Speaker diarization (diarizzazione del parlato)](https://broadrouter.com/it/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): La speaker diarization è il compito di dividere una registrazione audio in segmenti per parlante, rispondendo a "chi ha parlato e quando". Etichetta ogni tratto di parlato con un ID anonimo come SPEAKER_00 e di solito si combina con una trascrizione per mostrare chi ha detto ogni battuta. - [Distillazione (knowledge distillation)](https://broadrouter.com/it/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): La distillazione, o knowledge distillation, è un metodo di addestramento in cui un modello student più piccolo impara a riprodurre gli output di un modello teacher più grande, come le sue distribuzioni di probabilità o le risposte generate, così da conservare buona parte della qualità del teacher a un costo di esecuzione inferiore. - [Accordo sul trattamento dei dati (DPA)](https://broadrouter.com/it/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): Un accordo sul trattamento dei dati (DPA) è il contratto tra titolare e responsabile del trattamento che regola come il responsabile gestisce i dati personali. L’articolo 28, paragrafo 3, del GDPR lo impone e ne elenca i contenuti: istruzioni documentate, riservatezza, sicurezza, regole sui sub-responsabili, cancellazione o restituzione dei dati e audit. - [Embedding (embeddings)](https://broadrouter.com/it/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): Gli embedding sono liste di numeri in virgola mobile (vettori) che un modello produce per rappresentare un testo, un’immagine o un altro input. Input con significato simile ottengono vettori vicini tra loro, quindi la distanza tra vettori serve a ordinare i risultati di ricerca, raggruppare documenti o trovare elementi correlati. - [Eval (valutazione degli LLM)](https://broadrouter.com/it/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Una eval (abbreviazione di evaluation, valutazione) è un test ripetibile di un modello linguistico o di un prompt: un insieme fisso di input, una descrizione di come deve essere un buon output e un valutatore (codice, una persona o un altro modello) che assegna un punteggio a ogni output, così da poter confrontare le versioni. - [Backoff esponenziale (exponential backoff)](https://broadrouter.com/it/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): Il backoff esponenziale è una strategia di retry in cui un client aspetta sempre di più dopo ogni tentativo fallito, di solito raddoppiando il ritardo fino a un tetto, prima di riprovare. Aggiungere un jitter casuale a ogni attesa distribuisce i retry di molti client, così non colpiscono il server nello stesso istante. - [Few-shot prompting (prompt con esempi)](https://broadrouter.com/it/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): Il few-shot prompting è una tecnica in cui il prompt include un piccolo numero di esempi svolti di input e output prima dell’input reale, così il modello linguistico ne ricava compito e formato. I pesi del modello non cambiano; lo zero-shot non usa esempi e il one-shot ne usa uno. - [Fine-tuning (addestramento di un LLM)](https://broadrouter.com/it/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): Il fine-tuning è il processo con cui si continua ad addestrare un modello linguistico già pre-addestrato su un dataset più piccolo e specifico, come conversazioni di esempio, così che i suoi pesi cambino e svolga quel compito o quello stile in modo più affidabile, senza istruzioni lunghe o esempi in ogni prompt. - [finish_reason](https://broadrouter.com/it/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason è un campo di ogni choice nella risposta di una chat completion che dice perché il modello ha smesso di generare. Nel formato OpenAI i valori sono stop, length, tool_calls, content_filter e il deprecato function_call. La Messages API di Anthropic riporta la stessa informazione come stop_reason. - [Piano LLM flat (flat-rate)](https://broadrouter.com/it/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): Un piano LLM flat applica un prezzo fisso per un periodo, per esempio un mese, per accedere all’API di un modello linguistico invece di fatturare per token. Di solito questi piani limitano qualcos’altro, come concorrenza, velocità, modelli o finestre di utilizzo, quindi i limiti contano quanto il prezzo. - [FP8 (virgola mobile a 8 bit)](https://broadrouter.com/it/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 è un formato numerico in virgola mobile a 8 bit usato nel deep learning, definito in due codifiche: E4M3, con 4 bit di esponente e 3 di mantissa, ed E5M2, con 5 bit di esponente e 2 di mantissa. Dimezza la memoria rispetto ai formati a 16 bit e gira in modo nativo sulle GPU recenti. - [Frequency e presence penalty](https://broadrouter.com/it/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): Frequency e presence penalty sono impostazioni di campionamento che abbassano il punteggio dei token già presenti nell’output. La presence penalty si applica appena un token è comparso; la frequency penalty cresce a ogni ripetizione. Nell’API di OpenAI entrambe vanno da -2.0 a 2.0, con 0 come predefinito. - [GDPR (Regolamento generale sulla protezione dei dati)](https://broadrouter.com/it/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): Il GDPR (Regolamento generale sulla protezione dei dati, Regolamento (UE) 2016/679) è la legge UE sul trattamento dei dati personali. Si applica dal 25 maggio 2018 a organizzazioni stabilite nell’UE e ad alcune extra UE. Per le API LLM regola i contratti con i responsabili del trattamento, i trasferimenti internazionali e la minimizzazione dei dati. - [GGUF (formato dei modelli llama.cpp)](https://broadrouter.com/it/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF è un formato di file binario per salvare modelli linguistici da usare per l’inferenza con runtime basati su GGML, come llama.cpp. Un solo file GGUF contiene i pesi, spesso quantizzati, e i metadati necessari a caricare il modello, compresi tokenizer e dettagli dell’architettura, e può essere mappato in memoria per un caricamento rapido. - [Ora GPU (GPU-hour)](https://broadrouter.com/it/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): Un’ora GPU (GPU-hour) è un processore grafico usato per un’ora. I provider cloud fatturano le istanze GPU a noleggio a ore o a secondi, e i ricercatori usano le ore GPU per indicare quanto calcolo ha richiesto un lavoro. Paghi il tempo della macchina, che sia occupata o no. - [Grounding (ancoraggio alle fonti)](https://broadrouter.com/it/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): Il grounding è la pratica di collegare l’output di un modello linguistico a fonti di informazione specifiche e verificabili, come documenti forniti, risultati di ricerca o record di database, chiedendogli di rispondere a partire da quelle. Le risposte ancorate si possono ricondurre a una fonte e contengono meno spesso fatti inventati. - [Guardrail (controlli di sicurezza per LLM)](https://broadrouter.com/it/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): I guardrail sono controlli che un’applicazione esegue attorno alla chiamata a un modello linguistico per tenerla entro limiti stabiliti. I guardrail di input esaminano ciò che entra nel modello, quelli di output controllano ciò che esce e quelli sui tool validano le azioni prima che vengano eseguite. Un controllo fallito blocca, riscrive o segnala la richiesta. - [Allucinazione (hallucination)](https://broadrouter.com/it/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): Un’allucinazione è un output di un modello linguistico che sembra plausibile ma è falso, inventato o non supportato dall’input ricevuto: una citazione inesistente, un numero sbagliato, un’API che non esiste. Nasce dal fatto che il modello predice il testo più probabile invece di verificare i fatti. - [Hamza (lettera araba)](https://broadrouter.com/it/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): La hamza (ء) è la lettera araba che indica il colpo di glottide. Si scrive da sola o su un supporto: alef (أ إ), waw (ؤ) o yeh (ئ), e آ unisce alef e madda. Dato che il testo digitato la usa in modo incoerente, motori di ricerca e valutazione del parlato spesso riducono le forme con alef a ا. - [Ricerca ibrida (hybrid search)](https://broadrouter.com/it/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): La ricerca ibrida è un metodo di retrieval che esegue sulla stessa query sia una ricerca per parole chiave, come BM25, sia una ricerca vettoriale sugli embedding, poi fonde le due classifiche in una sola, spesso con la reciprocal rank fusion (RRF) o con una somma pesata di punteggi normalizzati. - [Chiave di idempotenza (idempotency key)](https://broadrouter.com/it/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Una chiave di idempotenza è un valore univoco, di solito un ID casuale come un UUID, che il client allega a una richiesta che modifica lo stato, come un pagamento o un acquisto. Se la stessa richiesta arriva di nuovo con la stessa chiave, il server restituisce il risultato originale invece di eseguire l’azione due volte. - [Token immagine (image tokens)](https://broadrouter.com/it/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): I token immagine sono i token di input con cui un modello vision-language rappresenta un’immagine nel prompt. I provider li calcolano dalle dimensioni dell’immagine, di solito dividendola in patch o tile di misura fissa dopo il ridimensionamento, e li fatturano come token di input testuali: le immagini più grandi costano di più. - [Provider di inferenza (inference provider)](https://broadrouter.com/it/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): Un provider di inferenza è un’azienda che fa girare modelli di machine learning, spesso LLM open-weight, su GPU di proprietà, in leasing o a noleggio, e ne vende l’accesso tramite un’API, di solito a prezzo per token, per richiesta o per unità di tempo. Si occupa di serving, scalabilità e hardware. - [Token di input e di output](https://broadrouter.com/it/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): I token di input sono quelli che un modello linguistico legge in una richiesta: system prompt, cronologia, definizioni degli strumenti e immagini. I token di output sono quelli che genera in risposta, ragionamento compreso. Le API LLM di solito li prezzano separatamente, e l’output costa di più. - [INT4 (quantizzazione a 4 bit)](https://broadrouter.com/it/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 è un formato numerico intero a 4 bit che può rappresentare 16 livelli. Nell’inferenza degli LLM indica di solito la quantizzazione a 4 bit dei soli pesi (W4A16): i pesi sono salvati in INT4 con fattori di scala ed espansi a 16 bit durante il calcolo, usando circa un quarto della memoria dei pesi a 16 bit. - [IP allow list (whitelist di IP)](https://broadrouter.com/it/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): Una IP allow list è un controllo di sicurezza che accetta richieste solo da indirizzi IP o intervalli CIDR elencati e rifiuta tutti gli altri. Su un’API limita i danni di una chiave trapelata, perché la chiave non funziona dalle reti che non sono nella lista. - [Jailbreak (LLM)](https://broadrouter.com/it/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): Un jailbreak è un input costruito per far ignorare a un modello linguistico il suo addestramento di sicurezza o le regole fissate dall’applicazione che lo usa, così che produca contenuti o compia azioni che normalmente rifiuterebbe. OWASP considera il jailbreaking una forma di prompt injection. - [JSON mode (modalità JSON)](https://broadrouter.com/it/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): La JSON mode è un’impostazione delle API LLM, response_format {"type": "json_object"} nelle API in stile OpenAI, che rende la risposta del modello JSON valido e analizzabile. A differenza degli structured output con un JSON Schema, non garantisce quali chiavi, tipi o valori contenga l’oggetto. - [JSON Schema](https://broadrouter.com/it/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema è un vocabolario, scritto in JSON, per descrivere e validare la struttura dei dati JSON: tipi, campi obbligatori, valori ammessi e annidamento. La versione attuale è la 2020-12. Le API LLM lo usano per definire i parametri dei tool e la forma esatta che un output strutturato deve seguire. - [KV cache (cache chiave-valore)](https://broadrouter.com/it/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): La KV cache è la memoria in cui un modello linguistico transformer conserva le chiavi e i valori di attenzione calcolati per ogni token già elaborato, così ogni nuovo token li riusa invece di ricalcolare l’intera sequenza. Cresce con la lunghezza del contesto e spesso limita quante richieste una GPU può servire. - [Latenza (latency)](https://broadrouter.com/it/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): La latenza, per un’API di modelli linguistici, è il tempo che passa dall’invio di una richiesta alla ricezione dell’ultimo token di output. È la somma del time to first token e del tempo di generazione, e di solito si riporta in percentili come p50 e p95 invece che come media. - [LLM-as-a-judge (LLM come giudice)](https://broadrouter.com/it/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge è un metodo di valutazione in cui un modello linguistico giudica gli output di altri modelli, dando un punteggio a una singola risposta secondo una rubrica oppure scegliendo la migliore tra due. Approssima il giudizio umano a un costo molto più basso, ma ha bias noti, come preferire la prima risposta o quella più lunga. - [Benchmark LLM](https://broadrouter.com/it/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): Un benchmark LLM è un insieme fisso di compiti, di solito pubblico, con un metodo di punteggio, usato per misurare e confrontare i modelli linguistici su una capacità, come domande di conoscenza, programmazione o riconoscimento vocale. Il punteggio mostra la capacità relativa su quel test, non come andrà un modello sul tuo carico di lavoro. - [LLM router (router per LLM)](https://broadrouter.com/it/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): Un LLM router è un software o un servizio che si mette tra un’applicazione e una o più API di modelli linguistici e decide, per ogni richiesta, quale modello o provider la gestisce, in base a regole come costo, latenza, disponibilità o tipo di compito. - [Logprobs (log-probabilità)](https://broadrouter.com/it/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): I logprobs sono le log-probabilità che un modello linguistico ha assegnato ai token che ha generato. Un’API che li restituisce mostra, per ogni token di output, quanto il modello lo riteneva probabile e, facoltativamente, i token alternativi più probabili in quella posizione. - [LoRA (low-rank adaptation)](https://broadrouter.com/it/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation, adattamento a basso rango) è un metodo di fine-tuning efficiente nei parametri che congela i pesi di un modello pre-addestrato e addestra una coppia di piccole matrici a basso rango per i layer scelti. L’aggiornamento appreso può essere salvato come piccolo adapter o fuso nei pesi, senza aggiungere latenza in inferenza. - [Lost in the middle (perso nel mezzo)](https://broadrouter.com/it/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle è la tendenza osservata dei modelli linguistici a usare meglio le informazioni rilevanti quando compaiono vicino all’inizio o alla fine di un input lungo, e peggio quando stanno nel mezzo. Il nome viene da uno studio del 2023 di Liu et al. - [max_tokens](https://broadrouter.com/it/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens è un parametro delle API LLM che fissa il numero massimo di token che un modello può generare in una risposta. Quando la generazione raggiunge il limite, si ferma e la risposta viene segnata come troncata, per esempio con finish_reason impostato a "length". - [MCP (Model Context Protocol)](https://broadrouter.com/it/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): L’MCP (Model Context Protocol) è un protocollo aperto, basato su JSON-RPC 2.0, che standardizza il modo in cui le applicazioni LLM si collegano a strumenti e dati esterni. Un server MCP espone strumenti, risorse e prompt; un client MCP dentro l’applicazione host li scopre e inoltra al server le chiamate agli strumenti fatte dal modello. - [Mixture of experts (MoE)](https://broadrouter.com/it/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): La mixture of experts (MoE) è un’architettura di rete neurale in cui un layer contiene molte sotto-reti esperte e un router appreso manda ogni token solo ad alcune di esse. Un modello linguistico MoE ha un numero totale di parametri elevato, ma per ogni token ne usa un piccolo sottoinsieme attivo. - [Deprecazione dei modelli (model deprecation)](https://broadrouter.com/it/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): La deprecazione di un modello avviene quando un provider di API annuncia che un modello non sarà più offerto, di solito indicando una data di ritiro o di spegnimento e un sostituto consigliato. Dopo quella data le richieste al modello falliscono, quindi le applicazioni devono passare prima a un altro ID di modello. - [Fallback del modello (model fallback)](https://broadrouter.com/it/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): Il fallback del modello è uno schema di affidabilità in cui una richiesta LLM fallita viene ritentata su un altro modello o provider di una lista ordinata. Può scattare per un errore, un rate limit, un errore di lunghezza del contesto o un rifiuto per policy sui contenuti. Gira nel client o in un gateway davanti alle API. - [Licenza del modello (model license)](https://broadrouter.com/it/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): La licenza di un modello è l’insieme delle condizioni legali con cui pesi, codice e output di un modello possono essere usati, modificati e condivisi. I modelli open-weight usano licenze software standard come Apache 2.0 o MIT, oppure licenze custom che possono aggiungere condizioni, per esempio limiti per i grandi servizi commerciali. - [Arabo standard moderno (MSA)](https://broadrouter.com/it/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): L’arabo standard moderno (MSA) è la forma standardizzata dell’arabo usata nella scrittura, nei notiziari, nell’istruzione e nei discorsi formali in tutto il mondo arabo. Non è la lingua madre quotidiana di nessuno: le persone parlano dialetti regionali, come l’egiziano o l’arabo del Golfo, e passano all’MSA nei contesti formali. - [Sistema multi-agente (multi-agent system)](https://broadrouter.com/it/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): Un sistema multi-agente è un’applicazione LLM in cui più agenti, ognuno con il proprio prompt, i propri strumenti e il proprio contesto, lavorano su parti di un unico compito. Uno schema comune è orchestrator-worker: un agente principale pianifica, avvia agenti worker in parallelo e combina i loro risultati. - [AI multimodale (multimodal AI)](https://broadrouter.com/it/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): L’AI multimodale è un modello o un sistema che lavora con più di un tipo di dato, come testo, immagini, video e audio. Un modello multimodale può leggere insieme una foto e una domanda, oppure ricevere parlato e rispondere in testo. Il tipo più comune è il modello vision-language. - [Riconoscimento di entità nominate (NER)](https://broadrouter.com/it/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): La named entity recognition (NER) è il compito di trovare i nomi in un testo ed etichettare ciascuno con un tipo, come persona, organizzazione o luogo. I sistemi classici etichettano ogni token; agli LLM si chiede spesso invece di restituire una lista di entità in JSON, che poi si verifica sul testo originale. - [Riconoscimento ottico dei caratteri (OCR)](https://broadrouter.com/it/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): L’OCR (riconoscimento ottico dei caratteri) è la tecnologia che individua il testo stampato o scritto a mano in un’immagine, una scansione o un PDF e lo converte in caratteri leggibili dalle macchine. L’OCR classico restituisce il testo e la sua posizione nella pagina; i modelli vision-language possono leggere il testo e rispondere direttamente a domande su di esso. - [Ogg Opus](https://broadrouter.com/it/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus è audio codificato con il codec Opus (RFC 6716) e salvato in un container Ogg (RFC 7845). È pensato per il parlato e l’audio interattivo a bitrate bassi, ed è il formato che WhatsApp usa per i messaggi vocali, di solito come file .ogg o .opus. - [Modello open-weight (a pesi aperti)](https://broadrouter.com/it/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): Un modello open-weight è un modello di machine learning i cui pesi addestrati sono pubblicati perché chiunque possa scaricarli ed eseguirli, di solito con una licenza che pone condizioni. Si distingue dall’AI completamente open source, che rilascia anche il codice di addestramento e informazioni dettagliate sui dati, così altri possono studiare e riprodurre il modello. - [API compatibile con OpenAI](https://broadrouter.com/it/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): Un’API compatibile con OpenAI è un’API HTTP che accetta gli stessi formati di richiesta e di risposta dell’API di OpenAI, come POST /v1/chat/completions, così SDK e strumenti di OpenAI possono usarla cambiando solo base URL, chiave API e nome del modello. - [Chiamate parallele a strumenti (parallel tool calls)](https://broadrouter.com/it/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Le chiamate parallele a strumenti (parallel tool calls) si hanno quando un modello restituisce più di una chiamata a strumenti in un’unica risposta, per esempio tre letture di file insieme. Nel formato OpenAI arrivano come più voci in tool_calls, e parallel_tool_calls impostato a false limita il modello a zero o una chiamata per turno. - [Pagamento a consumo (pay-as-you-go)](https://broadrouter.com/it/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): Il pagamento a consumo è un modello di prezzo in cui paghi solo ciò che usi, misurato per unità come token, richieste o minuti di audio, senza abbonamento né impegno minimo. La maggior parte delle API LLM fattura così, da un saldo prepagato o con una fattura periodica. - [PDPL (leggi privacy di Arabia Saudita ed EAU)](https://broadrouter.com/it/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) è il nome comune della legge saudita sui dati personali, emanata con il Regio Decreto M/19 e applicata dalla SDAIA, e del Decreto-legge federale n. 45 del 2021 degli EAU. Entrambe regolano il trattamento dei dati personali e pongono condizioni ai trasferimenti all’estero, compresi i prompt inviati ad API LLM straniere. - [PII (informazioni personali identificabili)](https://broadrouter.com/it/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): I PII (personally identifiable information) sono informazioni che permettono di distinguere o rintracciare l’identità di una persona, come nome, numero identificativo o dato biometrico, più altre informazioni collegate o collegabili a essa. Il GDPR usa il termine più ampio dati personali. Con gli LLM, l’oscuramento dei PII li toglie dai prompt prima che escano dai tuoi sistemi. - [Prefill e decode](https://broadrouter.com/it/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill e decode sono le due fasi dell’inferenza di un LLM. Nel prefill il modello elabora tutto il prompt in parallelo e riempie la sua KV cache; nel decode genera la risposta un token alla volta, riusando quella cache. Il prefill determina soprattutto il tempo al primo token; il decode la velocità di output. - [Credito prepagato (prepaid credit)](https://broadrouter.com/it/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): Il credito prepagato è denaro aggiunto a un account API prima dell’uso. Ogni richiesta scala il proprio costo dal saldo e, quando il saldo si esaurisce, le richieste a consumo si fermano finché non ricarichi. Limita la spesa a quanto hai versato. - [Prompt engineering (ingegneria dei prompt)](https://broadrouter.com/it/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): Il prompt engineering è il processo con cui si scrivono, strutturano e testano le istruzioni date a un modello linguistico, perché produca in modo costante un output che rispetta i tuoi requisiti. Comprende messaggi di ruolo, istruzioni chiare, esempi, formattazione e la misura dei risultati su casi di test. - [Prompt injection](https://broadrouter.com/it/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): La prompt injection è un attacco in cui un testo fornito a un modello linguistico, da un utente oppure nascosto in una pagina web, un’email, un file o il risultato di uno strumento, cambia il comportamento del modello contro le intenzioni dello sviluppatore. OWASP la mette al primo posto (LLM01:2025) nella sua Top 10 per le applicazioni LLM. - [Quantizzazione (quantization)](https://broadrouter.com/it/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): La quantizzazione è il processo con cui si memorizzano i numeri di una rete neurale, soprattutto i pesi, in un formato a precisione ridotta come 8 o 4 bit invece di 16 bit, così il modello usa meno memoria e spesso gira più veloce. I metodi validi mantengono piccola la perdita di accuratezza. - [RAG (retrieval-augmented generation)](https://broadrouter.com/it/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): La RAG (retrieval-augmented generation, generazione aumentata dal recupero) è uno schema in cui un’applicazione prima recupera da un archivio esterno, come un indice di ricerca o vettoriale, i passaggi pertinenti a una domanda, poi li aggiunge al prompt, così un modello linguistico può rispondere a partire da quel testo e non solo dalla memoria. - [Rate limit (limite di frequenza)](https://broadrouter.com/it/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): Un rate limit è un tetto che un’API pone a quanto un client può usarla in un intervallo di tempo, per esempio richieste al minuto o token al minuto. Le richieste oltre il limite vengono rifiutate, di solito con HTTP 429 Too Many Requests, finché l’uso non torna sotto la soglia. - [ReAct (reason and act)](https://broadrouter.com/it/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (reason and act) è un pattern di prompting, introdotto da Yao et al. nel 2022, in cui un modello linguistico alterna tracce di ragionamento e azioni come ricerche o chiamate a strumenti. Legge ogni osservazione, aggiorna il suo piano e ripete finché può dare una risposta finale. - [Token di ragionamento (reasoning tokens)](https://broadrouter.com/it/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): I token di ragionamento sono token che un modello di ragionamento genera per elaborare un problema prima di scrivere la risposta visibile. Di solito sono fatturati come token di output, contano nel limite di output e nella finestra di contesto, e l’API può nasconderli o riassumerli. - [Timeout della richiesta (request timeout)](https://broadrouter.com/it/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): Il timeout di una richiesta è il tempo massimo che un client HTTP aspetta un server prima di abbandonare la richiesta. I client di solito lo dividono in un timeout di connessione, per aprire la connessione, e un timeout di lettura, per l’attesa dei dati. Le chiamate LLM richiedono timeout di lettura generosi perché le risposte lunghe richiedono tempo. - [Reranking (riordinamento dei risultati)](https://broadrouter.com/it/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Il reranking è una seconda fase di retrieval in cui un modello più accurato ma più lento, di solito un cross-encoder, assegna un punteggio a ogni coppia query-documento candidato di una prima ricerca e le riordina, così solo i pochi documenti più pertinenti arrivano al modello linguistico. - [Testo da destra a sinistra (RTL, right-to-left)](https://broadrouter.com/it/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): Il testo da destra a sinistra (RTL) è una scrittura, come l’arabo o l’ebraico, visualizzata da destra verso sinistra. I computer lo memorizzano in ordine logico di lettura, e l’algoritmo bidirezionale Unicode decide l’ordine di visualizzazione quando si mescola con testo da sinistra a destra, cifre, nomi di modelli o codice. - [Frequenza di campionamento (sample rate)](https://broadrouter.com/it/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): La frequenza di campionamento (sample rate) è il numero di campioni audio acquisiti al secondo, misurato in hertz (Hz). I modelli di riconoscimento vocale si aspettano di solito audio mono a 16.000 Hz (16 kHz). Insieme a profondità di bit e numero di canali, determina quanto pesa al minuto un file audio non compresso. - [Parametro seed](https://broadrouter.com/it/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): Il parametro seed è un intero che invii con una richiesta perché l’API provi a campionare gli stessi token quando la stessa richiesta viene ripetuta. OpenAI lo descrive come best effort: il determinismo non è garantito, e il campo system_fingerprint indica quando la configurazione del backend è cambiata. - [Inferenza serverless (serverless inference)](https://broadrouter.com/it/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): L’inferenza serverless è un modo di eseguire modelli di machine learning in cui la piattaforma avvia, scala e ferma automaticamente le risorse di calcolo dietro un endpoint, spesso fino a zero quando è inattivo, e fattura l’uso invece delle macchine riservate. Il compromesso è un possibile ritardo da cold start. - [SLA (accordo sul livello di servizio)](https://broadrouter.com/it/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): Uno SLA (service level agreement, accordo sul livello di servizio) è un impegno contrattuale di un fornitore su un livello di servizio misurabile, di solito una percentuale di uptime mensile, con un rimedio definito, di solito crediti di servizio, se non viene rispettato. Lo SLA stabilisce cosa conta come downtime, cosa è escluso e come fare reclamo. - [Speculative decoding (decodifica speculativa)](https://broadrouter.com/it/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): Lo speculative decoding (decodifica speculativa) è una tecnica di inferenza in cui un modello draft piccolo e veloce propone diversi token successivi e il modello target, più grande, li verifica in un solo forward pass, tenendo quelli accettati. Accelera la generazione senza cambiare la distribuzione dell’output. - [Speech-to-text (riconoscimento vocale)](https://broadrouter.com/it/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): Lo speech-to-text, detto anche riconoscimento automatico del parlato (ASR), è la tecnologia che converte l’audio parlato in testo scritto. Un’API speech-to-text riceve un file o uno stream audio e restituisce una trascrizione, di solito a prezzo per minuto di audio. L’accuratezza si misura con il word error rate e varia con accento, dialetto e qualità della registrazione. - [Limite di spesa (spend limit)](https://broadrouter.com/it/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): Un limite di spesa è un tetto a quanto un account, un progetto o una chiave API può spendere in un periodo, di solito un mese. Quando l’uso lo raggiunge, il provider rifiuta le richieste successive, spesso con un errore HTTP 429 o 400, finché il periodo non si azzera o il limite non viene alzato. - [Sottotitoli SRT (file .srt)](https://broadrouter.com/it/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) è un formato di sottotitoli in testo semplice. Ogni sottotitolo è composto da un numero progressivo, un tempo di inizio e di fine scritto come hh:mm:ss,mmm --> hh:mm:ss,mmm, una o più righe di testo e una riga vuota. WebVTT è il formato web simile usato dai video HTML. - [Server-sent events (SSE)](https://broadrouter.com/it/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): I server-sent events (SSE) sono uno standard web con cui un server invia a un client un flusso di eventi testuali su un’unica risposta HTTP di lunga durata, con content type text/event-stream. Ogni evento è composto da una o più righe di campo, come data:, ed è chiuso da una riga vuota. - [Sequenze di stop (stop sequences)](https://broadrouter.com/it/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): Le sequenze di stop sono stringhe che invii con una richiesta e che fanno smettere il modello di generare appena ne produce una. OpenAI chiama il campo stop e accetta fino a 4 sequenze; Anthropic lo chiama stop_sequences. Di norma il testo trovato non compare nell’output restituito. - [Streaming](https://broadrouter.com/it/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): Lo streaming è un modo di restituire la risposta di un modello linguistico in piccoli pezzi mentre viene ancora generata, invece che in un’unica risposta alla fine. Le API compatibili con OpenAI trasmettono in streaming tramite server-sent event quando la richiesta imposta stream a true. - [Output strutturato (structured output)](https://broadrouter.com/it/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): L’output strutturato è una funzionalità delle API LLM che vincola la risposta di un modello a un formato indicato dallo sviluppatore, di solito un JSON Schema, così il codice può interpretarla in modo affidabile. È più rigoroso della JSON mode, che garantisce un JSON valido ma non una struttura precisa. - [Subagent (sub-agente)](https://broadrouter.com/it/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): Un subagent è un agente separato che un agente principale avvia per gestire un compito specifico. Lavora con una propria finestra di contesto, un proprio system prompt e propri strumenti, svolge il lavoro e restituisce all’agente principale solo un risultato o un riepilogo, mantenendo piccola la conversazione principale. - [Sub-responsabile del trattamento (subprocessor)](https://broadrouter.com/it/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): Un sub-responsabile del trattamento (subprocessor) è un terzo a cui un responsabile affida attività di trattamento di dati personali per conto del titolare. Per un’API LLM può includere host GPU, provider di inferenza partner e capacità di backup. Secondo l’articolo 28 del GDPR serve l’autorizzazione del titolare e gli stessi obblighi di protezione dei dati. - [System prompt (prompt di sistema)](https://broadrouter.com/it/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): Il system prompt è il testo di istruzioni che un’applicazione invia a un modello linguistico prima della conversazione, di solito come messaggio con ruolo system o developer. Stabilisce ruolo, regole, tono e formato di output del modello, che dovrebbe seguirlo anche quando i messaggi dell’utente lo contraddicono. - [Tā’ marbūṭa (taa marbuta, ة)](https://broadrouter.com/it/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): La taa marbuta (ة, U+0629) è una lettera araba che compare solo alla fine di una parola e spesso segna sostantivi e aggettivi femminili. Somiglia a una heh (ه, U+0647) con due punti, quindi nel testo digitato le due si confondono, e molti normalizzatori per la ricerca mappano ة su ه. - [Temperatura (temperature)](https://broadrouter.com/it/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): La temperatura è un parametro di campionamento che regola quanto è casuale l’output di un modello linguistico. Il modello divide i punteggi dei token per la temperatura prima di trasformarli in probabilità: con valori bassi domina il token più probabile, con valori alti la distribuzione si appiattisce e i token meno probabili vengono scelti più spesso. - [Text-to-speech (TTS, sintesi vocale)](https://broadrouter.com/it/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): Il text-to-speech (TTS), o sintesi vocale, converte il testo scritto in audio parlato. Un’API TTS riceve un testo, una voce e un formato di output come MP3 o WAV, e restituisce il parlato generato. È l’opposto dello speech-to-text, che trasforma l’audio in testo. - [Throughput](https://broadrouter.com/it/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): Il throughput, nel serving di LLM, è la quantità totale di lavoro che un sistema completa per unità di tempo su tutte le richieste concorrenti, di solito misurata in token di output al secondo o richieste al secondo. È diverso dalla velocità per stream, cioè quanto rapidamente una singola richiesta riceve i suoi token. - [Time to first token (TTFT)](https://broadrouter.com/it/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): Il time to first token (TTFT) è il tempo che passa dall’invio di una richiesta a un modello linguistico all’arrivo del primo token di output. Comprende il tempo di rete, la coda sul server e il prefill, la fase in cui il modello legge l’intero prompt, quindi cresce con la lunghezza del prompt e con il carico. - [Timestamp (speech-to-text)](https://broadrouter.com/it/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): I timestamp nello speech-to-text sono tempi di inizio e fine, in secondi, associati a parti di una trascrizione. I timestamp di segmento coprono una frase o un periodo; quelli per parola coprono ogni parola. Servono per sottotitoli, ricerca dentro le registrazioni e allineamento del testo con i parlanti. - [Token](https://broadrouter.com/it/glossary/token/) (translation of https://broadrouter.com/glossary/token/): Un token è l’unità di testo che un modello linguistico legge e scrive: una parola, una parte di parola, un segno di punteggiatura o una sequenza di byte prodotta dal tokenizer del modello. Le API LLM contano, limitano e di solito fatturano l’uso in token. - [Tokenizer](https://broadrouter.com/it/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): Un tokenizer è il componente che converte il testo nella sequenza di ID interi di token che un modello linguistico legge, e riconverte in testo gli ID generati. Ogni modello ha il proprio tokenizer, quindi lo stesso testo può produrre un numero di token diverso su modelli diversi. - [Token al secondo (tokens per second)](https://broadrouter.com/it/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): I token al secondo (TPS) sono la velocità con cui un modello linguistico genera token di output, di solito misurata per una richiesta dopo l’arrivo del primo token. Stabiliscono quanto scorre veloce una risposta una volta partita. Sommati su tutte le richieste concorrenti diventano throughput di sistema; l’attesa del primo token si misura a parte. - [Tool calling (chiamata di strumenti)](https://broadrouter.com/it/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): Il tool calling, detto anche function calling, è una funzionalità delle API LLM in cui il modello restituisce una richiesta strutturata di eseguire una funzione definita dallo sviluppatore, con argomenti JSON. L’applicazione esegue la funzione e rimanda il risultato; il modello non esegue mai codice da solo. - [tool_choice](https://broadrouter.com/it/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice è un parametro della richiesta che controlla l’uso degli strumenti. In Chat Completions di OpenAI può valere auto (decide il modello), required (deve chiamare uno o più strumenti), none (nessuno strumento) o una funzione specifica da forzare. Le opzioni equivalenti di Anthropic sono auto, any, tool e none. - [Top-p (nucleus sampling)](https://broadrouter.com/it/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Il top-p, detto anche nucleus sampling, è un metodo di decodifica in cui un modello linguistico sceglie il token successivo solo dal più piccolo insieme di token più probabili la cui probabilità complessiva raggiunge una soglia p, per esempio 0,9, scartando la coda improbabile. - [Normalizzazione Unicode (Unicode normalization)](https://broadrouter.com/it/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): La normalizzazione Unicode converte il testo in una forma standard, così che stringhe che appaiono uguali usino gli stessi code point. Le quattro forme sono NFD, NFC, NFKD e NFKC; le forme K riducono anche i caratteri di compatibilità, come le forme di presentazione arabe e le legature, a lettere ordinarie. - [Voice activity detection (VAD)](https://broadrouter.com/it/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): La voice activity detection (VAD, rilevamento dell’attività vocale) è un passaggio di elaborazione del segnale che classifica brevi frame audio come parlato o non parlato. I sistemi vocali la usano per saltare i silenzi, dividere registrazioni lunghe nelle pause, capire quando chi parla ha finito ed evitare di trascrivere il rumore di fondo. - [Database vettoriale (vector database)](https://broadrouter.com/it/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): Un database vettoriale è un archivio dati che conserva vettori di embedding, spesso con il testo di origine e i metadati, e risponde a query di nearest neighbour: dato un vettore di query, restituisce i vettori memorizzati più simili, di solito per distanza coseno, distanza euclidea o prodotto scalare. - [Modello vision-language (VLM)](https://broadrouter.com/it/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): Un modello vision-language (VLM) è un modello che riceve in input immagini e testo insieme e produce testo, quindi può descrivere foto, rispondere a domande su di esse e leggere documenti, grafici o screenshot. I VLM sono chiamati anche modelli image-text-to-text o multimodali. - [vLLM](https://broadrouter.com/it/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM è una libreria open source per l’inferenza e il serving di LLM, sviluppata inizialmente allo Sky Computing Lab di UC Berkeley. Gestisce la memoria della KV cache con PagedAttention, raggruppa in batch continuo le richieste in arrivo, supporta molti formati di quantizzazione ed espone un server HTTP compatibile OpenAI. - [Dimensione del vocabolario (vocabulary size)](https://broadrouter.com/it/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): La dimensione del vocabolario è il numero di token distinti nel tokenizer di un modello linguistico, ognuno con il proprio ID e il proprio embedding. Un vocabolario più grande può codificare il testo comune in meno token, ma ingrandisce i livelli di embedding e di output, e aiuta solo sul testo su cui è stato addestrato. - [WAV (file audio WAV)](https://broadrouter.com/it/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format, o WAVE) è un contenitore audio basato sul formato RIFF di Microsoft e IBM. Di solito contiene campioni PCM lineari non compressi, quindi la sua dimensione è prevedibile: frequenza di campionamento per byte per campione per canali, al secondo. - [Word error rate (WER)](https://broadrouter.com/it/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): Il word error rate (WER, tasso di errore sulle parole) è la misura standard di accuratezza del riconoscimento vocale: il numero di sostituzioni, cancellazioni e inserimenti di parole necessari per trasformare una trascrizione nel riferimento corretto, diviso per il numero di parole del riferimento. Più è basso, meglio è. - [Zero data retention (nessuna conservazione dei dati)](https://broadrouter.com/it/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): La zero data retention (ZDR) è una policy di gestione dei dati in cui un provider AI elabora prompt e output solo per servire una richiesta e poi non ne conserva il contenuto, né in log né in cache né in dati di addestramento. Di solito restano registri senza contenuti, come conteggi di token e tempi. ### Chinese (Simplified) (zh, 简体中文): 294 translated pages - [Articles in Chinese (Simplified)](https://broadrouter.com/zh/blog/) - [2026 年 10 月大模型 API 价格变动:3 次涨价,没有降价](https://broadrouter.com/zh/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 2026 年 10 月,7 家大模型 API 服务商中有 5 家的价格页出现 23 处变动:3 次涨价、没有降价,9 个模型新定价,11 个被移除。每一项都附存档链接。 - [AI 编程智能体跑一次要花多少钱:步数、上下文增长与缓存](https://broadrouter.com/zh/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): 为什么编程智能体运行越久、每一步越贵,缓存如何改变总成本,以及固定价格会话何时比按 token 计费更划算。附 token 计算过程。 - [用大模型给阿拉伯语加注音符号:提示词、校验与 DER](https://broadrouter.com/zh/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): 用大模型 API 恢复阿拉伯语注音符号(tashkeel):不改动任何字母的提示词、拒收改字回复的校验、DER 与 WER 评分,以及何时改用 CATT。 - [用大模型做阿拉伯语语法和拼写纠错:实操指南](https://broadrouter.com/zh/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): 用大模型 API 纠正阿拉伯语拼写和语法:最小改动提示词、防止改写的校验、按错误类型标注的词级修改,以及 M2 式评分。 - [用 API 把阿拉伯语会议录音生成会议纪要](https://broadrouter.com/zh/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): 把阿拉伯语会议录音变成包含决策和待办事项的会议纪要:转写、清理、用大模型总结成 JSON,再核对人名和日期。 - [用大模型做阿拉伯语命名实体抽取:JSON、日期与校验](https://broadrouter.com/zh/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): 用大模型把阿拉伯语人名、机构、地点和伊斯兰历日期抽取为 JSON:提示词、忽略变音符号的片段校验、日期换算和评分脚本。 - [阿拉伯语 OCR API 对比:Google、Azure、Textract、Mistral 与大模型](https://broadrouter.com/zh/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): 哪些 OCR API 官方支持阿拉伯语印刷体和手写体,各自返回什么、每 1,000 页多少钱,以及与视觉大模型 API 的对比。 - [用语音转文字 API 生成阿拉伯语 SRT 字幕:你需要知道什么](https://broadrouter.com/zh/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): SRT 文件需要什么、哪些语音转文字 API 能为阿拉伯语返回时间戳或 SRT,以及一条用纯文本 API 生成字幕、经过测试的流水线。 - [用大模型做阿拉伯语文本分类:情感、意图与工单分派](https://broadrouter.com/zh/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): 用大模型 API 对阿拉伯语评论、消息和客服工单分类:标签设计、兼顾方言的提示词、JSON 输出、macro F1 测试,以及大规模下的成本。 - [面向搜索和 RAG 的阿拉伯语文本规范化:该改什么,不该改什么](https://broadrouter.com/zh/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Tashkeel、alef 变体、ta marbuta、hamza 载体、波斯字母、数字与 Unicode:哪些阿拉伯语规范化规则有助于搜索、哪些有害,附经过测试的 Python 代码。 - [用大模型做阿拉伯语文本摘要:提示词、长度控制与评分](https://broadrouter.com/zh/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): 用大模型 API 摘要阿拉伯语文章、报告和方言消息:提示词写法、字数控制、数字核对,以及默认 ROUGE 为什么给阿拉伯语打 0 分。 - [Artificial Analysis 替代品:2026 年在哪里比较大模型](https://broadrouter.com/zh/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Artificial Analysis 替代品对比:Arena、Epoch AI、Vals AI、OpenRouter 排行、Hugging Face 和价格数据,各自衡量什么、什么时候用。 - [BroadNet 的 GPU 硬件与推理部署方式(据 BroadNet 自述)](https://broadrouter.com/zh/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): BroadNet 对 BroadNet Inference 背后硬件的介绍:从 B300 到客户现场一体机的三层 GPU、每张 GPU 的吞吐量,以及它从测试中得出的经验。 - [缓存大模型 API 响应以降低成本:用 Python 实现精确匹配缓存](https://broadrouter.com/zh/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): 一个经过测试的 Python 大模型 API 响应精确匹配缓存:缓存键里放什么、哪些绝不能缓存、TTL、隐私,以及动手之前如何测出命中率。 - [Claude Code 接入开源模型:代理、原生端点与限制](https://broadrouter.com/zh/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code 使用的是 Anthropic Messages API。怎样通过 LiteLLM 代理或原生端点让它跑在开源模型上,以及途中会出什么问题。 - [把人工复核纳入文档 AI 工作流](https://broadrouter.com/zh/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): 如何设计文档 AI 流水线中的人工复核环节:校验规则、哪些文档交给人看、随机抽样审计、审计记录和复核率。 - [埃及数据保护法(2020 年第 151 号法律)与大模型 API](https://broadrouter.com/zh/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): 把埃及人的个人数据发送到境外的大模型或语音转文字 API:2020 年第 151 号法律和 2025 年第 816 号决定对传输、许可和数据泄露有何规定。 - [微调之前:先定义要解决的问题](https://broadrouter.com/zh/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): 什么时候提示词、RAG 或评测就能解决问题而无需微调,需要什么样的训练数据,以及在申请微调之前如何测试它。 - [大模型返回的 JSON 不合法怎么办:原因、修复与重试](https://broadrouter.com/zh/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): 大模型为什么会返回坏掉的 JSON(代码块围栏、多余文字、max_tokens 截断、尾随逗号),以及一个经过测试的 Python 函数:能安全修复的就修,其余一律拒绝。 - [大模型 API 价格多久变一次?三年存档价格页数据](https://broadrouter.com/zh/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): 我们读取了 11 家大模型 API 服务商 2023–2026 年的 308 份存档价格页:80% 的模型从未改价,DeepInfra 一家贡献了 75% 的标价变动。 - [如何降低大模型 API 费用:按顺序执行的 10 步清单](https://broadrouter.com/zh/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): 按正确顺序压缩大模型 API 账单:先度量,再缓存前缀、精简提示词、限制输出和思考、选对模型、批处理,并弄清固定价格套餐何时更划算。 - [大模型 API 故障转移:在自己的代码里切换到备用服务商](https://broadrouter.com/zh/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): 如何在 OpenAI 兼容的大模型服务商之间做故障转移:哪些错误该切换、熔断器、模型 ID 映射、流式输出,以及费用和隐私条款是否一致。 - [大模型 API 降价记录 2023–2026:最大降幅与出处](https://broadrouter.com/zh/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): 2023–2026 年大模型 API 的最大降价,每条都附 Internet Archive 变更前后链接:o3、Codestral、Gemini 1.5,以及少有人注意到的涨价。 - [大模型 API 只会降价吗?52 次涨价记录与预算方法](https://broadrouter.com/zh/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): 大模型 API 也会涨价。我们在 2024–2026 年的存档价格页中找到同一模型 52 次标价上涨,涨幅中位数 40%。附全部来源和一条预算规则。 - [各家大模型 API 速率限制对比:RPM、TPM 与并发](https://broadrouter.com/zh/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): 从官方文档整理 13 家大模型 API 入门档位的速率限制:OpenAI、Anthropic、Gemini、Groq、DeepSeek、Fireworks 等,并解释 RPM 和 TPM 是什么。 - [大模型 API SLA 对比:Azure、Bedrock、Google 和 OpenAI 承诺了什么](https://broadrouter.com/zh/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): 各家大模型 API 服务商白纸黑字承诺了什么:可用性百分比、什么算宕机、服务抵扣和延迟目标,全部来自各家的 SLA 页面。 - [没有 Batch API 也能批量调用大模型:可断点续跑的执行器](https://broadrouter.com/zh/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): 不用 Batch API,也能用 chat completions 接口跑完 1 万到 100 万条提示词:并发上限、速率控制、重试、可续跑的检查点,以及耗时怎么算。 - [Safety Map:39 个大模型如何处理敏感数据](https://broadrouter.com/zh/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet 评测了来自 14 家实验室的 39 个大模型对凭据、个人信息、健康和金融数据的处理:哪些安全能力可以迁移、哪些不能,以及工具调用和思考过程中的泄露。 - [如何测试大模型的敏感数据安全:BroadNet 的方法](https://broadrouter.com/zh/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): BroadNet 公开的评测方法:在部署前测试大模型是否会泄露凭据、个人信息、健康或金融数据。包括分类测试、泄露与漏检,以及至少 3 次运行。 - [从 OpenAI API 迁移:Python、Node 和 cURL 代码实战](https://broadrouter.com/zh/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): 从 OpenAI 迁移到 OpenAI 兼容 API 需要改哪些代码:客户端配置、从 Responses 改写为 Chat Completions、工具调用、流式输出、embeddings 和冒烟测试。 - [BroadNet 评估后淘汰了哪些模型,原因是什么](https://broadrouter.com/zh/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet 公开了它测试后淘汰的模型及原因:未通过凭据安全测试的对话模型,以及落选的语音、OCR 和向量模型。 - [如何为多语言用户设计 AI 产品体验](https://broadrouter.com/zh/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): 为阿拉伯语、英语和西班牙语用户设计 AI 产品:界面语言与模型语言分开、RTL 与混排文本、数字和日期、语言检测以及成本。 - [开源与闭源模型 API 成本对比:价格数据怎么说](https://broadrouter.com/zh/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): 通过 API 使用开放权重模型,真的比 GPT、Claude、Gemini 便宜吗?基于各服务商官方页面的 token 价格、托管方价差,以及单用户成本实例。 - [OpenAI、Anthropic 与 Google API 价格历史(2023–2026)](https://broadrouter.com/zh/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): 基于 2023–2026 年存档价格页,并排对比 GPT、Claude 和 Gemini API 价格:哪些产品线变便宜了,哪些变贵了,每个价格都附存档链接。 - [OpenAI 兼容 API 的 base_url 汇总:14 家服务商逐一核对](https://broadrouter.com/zh/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): 14 家 OpenAI 兼容大模型 API 的 base_url,均摘自各服务商官方文档,附上它们列出的端点,以及会让“直接替换”失败的差异。 - [OpenClaw 和 Hermes Agent 接便宜模型:配置与 token 账](https://broadrouter.com/zh/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): 把 OpenClaw 或 Hermes Agent 接到 OpenAI 兼容端点,弄清定时检查为什么烧 token,并用轻量上下文和更便宜的模型压低账单。 - [申请私有端点,先把工作负载说清楚](https://broadrouter.com/zh/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): 申请大模型私有端点前要明确什么:工作负载、每秒 token 数、并发、上下文、区域、许可证和预算,以及拿到端点后如何测试。 - [编程智能体的提示词注入:真正管用的防御措施](https://broadrouter.com/zh/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): 提示词注入从哪里进入编程智能体,以及厂商和研究者记录的分层防御:沙箱、出站网络限制、命令白名单和人工审批。 - [RAG 分块大小与重叠:对检索效果和成本的影响](https://broadrouter.com/zh/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): 分块大小和重叠如何改变 RAG 的召回率、索引规模、向量化成本,以及每次查询发给大模型的 token 数。附公开测试结果和具体计算。 - [思考模型的 max_tokens 怎么设:一套预算方法](https://broadrouter.com/zh/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): 为什么 max_tokens 设得低时思考模型会返回空回复,以及怎样根据实测分位数、上下文窗口和逐级加倍的重试来确定预算。 - [浏览器录音并发送到语音转文字 API](https://broadrouter.com/zh/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): 用 MediaRecorder 录制麦克风,把 WebM 或 Safari 的 MP4 转成 16 kHz WAV,再通过一个保管密钥的小服务端发送到语音转文字 API。 - [减少提示词 token:系统提示词、工具和数据的实测精简](https://broadrouter.com/zh/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): 用真实分词器实测:精简系统提示词、few-shot 示例、工具 schema、数据格式、对话历史和阿拉伯语变音符号,各能省下多少 token。 - [同一模型、不同服务商:大模型 API 价格差多少](https://broadrouter.com/zh/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): 在由多家托管方销售的 119 个模型中,最贵标价是最便宜标价的中位数 1.76 倍。数据来自 41 家服务商自己的价格页,2026 年 11 月。 - [沙特与阿联酋 PDPL 和大模型 API:数据跨境检查清单](https://broadrouter.com/zh/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): 从沙特阿拉伯或阿联酋把个人数据发送到大模型 API:PDPL 的跨境传输规则怎么说、该问服务商哪些问题,以及什么时候需要境内推理。 - [文本、语音还是视觉?先从任务出发](https://broadrouter.com/zh/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): 按任务选择文本对话、图片输入或音频转写,附 BroadNet 文档列出的模型与限制,以及它不提供的功能,例如 TTS。 - [转写超过 25 MB 的长音频:切分、发送、拼接](https://broadrouter.com/zh/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): 录音超过语音转文字 API 的文件限制怎么办:先压缩文件,用 ffmpeg 在静音处切分,并行发送各段,再按顺序拼接文本。 - [用大模型把应用字符串翻译成阿拉伯语:占位符与复数](https://broadrouter.com/zh/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): 用大模型 API 把 i18n JSON 文件翻译成阿拉伯语:冻结占位符,生成全部六种阿拉伯语复数形式,在 CI 中校验 ICU 语法,并隔离插入的名称。 - [用 API 把阿拉伯语音频翻译成英文文本:一条经过测试的流水线](https://broadrouter.com/zh/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): 两次 API 调用把阿拉伯语录音变成英文文本:先用 language=ar 转写,再用大模型翻译,附数字核对、分段、成本与评分。 - [已上线、按需提供、仅研究:看懂模型目录](https://broadrouter.com/zh/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): 以 BroadNet 的模型目录和 /v1/models 为例,教你分辨目录里哪些模型今天就能调用、哪些需要咨询、哪些只是测试过。 - [vLLM 中的 Uno:BroadNet 部署研究测了什么](https://broadrouter.com/zh/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet 把 Uno 自草稿适配器集成进 vLLM,并在 H100 上测量:单流吞吐量是普通 vLLM 的 2.58 倍,32 流时 1.74 倍。本文介绍方法、DOI 和局限。 - [Gemma 4 26B 上的 Uno:BroadNet 的混合专家(MoE)研究](https://broadrouter.com/zh/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet 首次为混合专家模型 Gemma 4 26B A4B 训练了 Uno 自草稿适配器。本文介绍它测了什么、自草稿何时划算,以及论文 DOI。 - [智能体 token 预算:智能体循环的 token 都花在哪里](https://broadrouter.com/zh/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): 系统提示词、工具 schema、文件读取、历史记录和思考:智能体循环如何消耗 token,如何用用量数据测量每一部分,以及先裁剪什么。 - [阿拉伯语语音识别的词错误率(WER)详解](https://broadrouter.com/zh/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): 词错误率是什么意思、怎么计算,以及为什么阿拉伯语 WER 对变音符号、拼写归一化和方言如此敏感。附 Python 代码和真实数据。 - [阿拉伯语方言语音识别:海湾、埃及、黎凡特、马格里布](https://broadrouter.com/zh/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): 语音转文字为什么难以处理海湾、埃及、黎凡特和马格里布阿拉伯语,以及如何搭建方言测试集,看清服务商在你的音频上表现如何。 - [用大模型做阿拉伯语与英语互译:提示词、术语表与质量评估](https://broadrouter.com/zh/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): 用大模型 API 做阿拉伯语与英语互译:保持语气的提示词、固定术语的术语表、数字与人名的处理,以及如何评估翻译质量。 - [海湾、埃及和黎凡特阿拉伯语:大模型如何处理方言](https://broadrouter.com/zh/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): 大模型如何应对海湾、埃及和黎凡特阿拉伯语:在哪里出错、怎样针对方言写提示词、怎样测试,以及 BroadNet 对 Qwen3.8 27B 的报告。 - [阿拉伯语 OCR 与视觉大模型:识别阿拉伯语文档和发票](https://broadrouter.com/zh/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): 阿拉伯语发票、收据和表单,什么时候用传统 OCR,什么时候用视觉大模型:准确率陷阱、JSON 输出、每页成本和校验流水线。 - [阿拉伯语提示词工程:语体、RTL、数字与中英混排](https://broadrouter.com/zh/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): 写出能让大模型输出优质阿拉伯语的提示词:选定语体和方言,控制数字和日期格式,保持 RTL 排版整洁,处理阿拉伯语与英语混排。 - [阿拉伯语 RAG:真正有效的规范化、分块与检索](https://broadrouter.com/zh/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): 针对阿拉伯语文档的检索增强生成:规范化 alef、ya 和 ta marbuta,合理分块,选择多语言嵌入模型,再加上关键词检索。 - [阿拉伯语语音转文字 API:2026 年怎么选](https://broadrouter.com/zh/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): 如何挑选阿拉伯语转写 API:词错误率、方言、文件限制、每音频小时价格和隐私,外加一套用你自己录音测试的方案。 - [阿拉伯语 WhatsApp 聊天机器人架构:从语音消息到自动回复](https://broadrouter.com/zh/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): 一套可落地的阿拉伯语 WhatsApp 助手架构:webhook、队列、语音消息转写、大模型回复、平台政策规则与成本,附可改造的 Python 代码。 - [Arabizi 与大模型:处理用 3、7 等数字书写的拉丁化阿拉伯语](https://broadrouter.com/zh/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): Arabizi 用拉丁字母书写阿拉伯语,用 3 表示 ع、7 表示 ح。如何识别它、用大模型把它转换成阿拉伯字母,以及如何正确回复。 - [Aider 用什么模型最好:如何选择并接入大模型](https://broadrouter.com/zh/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): 如何为 Aider 选出最合适的大模型:编辑格式的遵循度、上下文大小和每小时成本,在自己仓库上做一小时测试,以及准确的 OPENAI_API_BASE 配置。 - [编程智能体最佳开源模型:按工具、上下文和成本来选](https://broadrouter.com/zh/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): 编程智能体该用 GLM 5.3、DeepSeek V4.1 Flash、Qwen3.8 27B 还是 Gemma 4?按工具调用、上下文大小、思考和每个任务的成本来选,附计算过程。 - [缓存输入 token 详解:如何降低大模型 API 账单](https://broadrouter.com/zh/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): 缓存输入的价格只是普通输入的一小部分。本文讲清提示词缓存怎么计费、如何读取 usage 中的 cached_tokens,以及如何组织提示词来命中缓存。 - [最便宜的视觉大模型 API:每张图片和每页文档的成本算法](https://broadrouter.com/zh/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): 视觉大模型调用每张图片、每页文档要花多少钱:Gemma 4、Qwen3.8 27B 和 DeepSeek V4.1 Flash 的 token 算法,以及 DeepInfra 在哪些方面更便宜。 - [如何选择大模型 API 服务商:20 项检查清单](https://broadrouter.com/zh/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): 选大模型 API 之前要问的 20 个问题:价格、限额、OpenAI 兼容性、隐私、模型、支持与可用性,另附评分表和测试计划。 - [如何选择阿拉伯语大模型 API:该测试什么](https://broadrouter.com/zh/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): 为阿拉伯语应用挑选大模型:方言理解、写作质量、翻译、阿拉伯语文档和图像、成本,外加一个一天就能搭好的测试集。 - [Cline vs Roo Code vs Aider:用开源模型的团队怎么选(2026)](https://broadrouter.com/zh/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): 对比 Cline、Roo Code(已归档,由 Zoo Code 延续)和 Aider 在开源模型下最关键的几点:工具调用、上下文用量、MCP、git 和审批。 - [发送大模型请求前,怎么计算 token 数?](https://broadrouter.com/zh/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): 调用大模型 API 前按语言和模型估算提示词 token 数:分词器为何不同、五种语言的实测数据,以及 max_tokens 该留多少余量。 - [降低编程智能体大模型成本的 10 个方法(附节省估算)](https://broadrouter.com/zh/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): 降低 Cline、Aider、opencode 或自研编程智能体 token 开销的十个具体方法,每个都附上节省额的计算过程和衡量方式。 - [DeepSeek API 替代方案:调用 DeepSeek 模型的 5 种途径](https://broadrouter.com/zh/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): 通过官方 API、聚合平台、云厂商模型目录、自有 GPU 或 BroadNet Inference 调用 DeepSeek 模型:价格、隐私和限制各有什么取舍。 - [DeepSeek V4.1 Flash API 价格:按 token、缓存与套餐计费](https://broadrouter.com/zh/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): DeepSeek V4.1 Flash API 每百万输入、缓存输入和输出 token 的价格,按月、按周、按小时的固定价格套餐,以及与 DeepSeek 官方、DeepInfra、Fireworks 的价格对比。 - [上线新功能前,怎样估算大模型 API 费用](https://broadrouter.com/zh/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): 为新的 AI 功能逐步建立成本模型:用户数 × 操作数 × 调用数 × token 数,计入缓存、思考、重试和速率限制,并附一张月度账单示例表。 - [切换服务商之前,如何评估大模型](https://broadrouter.com/zh/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): 用真实流量构建一个小型评估集,先用代码检查、再用去偏差的 LLM 评审打分,在迁移前比较质量、成本和延迟。 - [用视觉大模型把发票、收据和表单数据提取为 JSON](https://broadrouter.com/zh/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): 一条面向发票、收据和表单的生产级流水线:图像预处理、JSON 字段提示词、Pydantic 校验、金额核对、重试和人工复核队列。 - [大模型 API 包月还是按 token 付费:盈亏平衡怎么算](https://broadrouter.com/zh/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): 固定价格的大模型套餐什么时候比按 token 付费更划算?DeepSeek V4.1 Flash 的四份月度账单实例、盈亏平衡用量,以及怎样测量你自己的用量。 - [大模型 API 免费额度:新用户初始额度到底能用多少](https://broadrouter.com/zh/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): 大模型 API 的免费额度实际能买到什么:BroadNet 初始额度可换算成多少 token、音频分钟和按小时密钥,并与 Gemini 和 OpenRouter 的免费层对比。 - [GDPR 与大模型 API:要问服务商的 15 个问题](https://broadrouter.com/zh/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): 面向大模型 API 的 GDPR 核对清单:数据处理协议(DPA)、子处理方、数据保留、训练、跨境传输和数据泄露通知,每个问题都附上对应的 GDPR 条款。 - [GLM 5.3 vs DeepSeek V4.1 Flash:该用哪个,什么时候用](https://broadrouter.com/zh/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): 从价格、上下文、工具调用、结构化输出、图像和思考几方面对比 GLM 5.3 与 DeepSeek V4.1 Flash,附实际工作负载的费用演算,以及一条简单的选型规则。 - [大模型 API 报 429 错误怎么办:退避、抖动与并发限制](https://broadrouter.com/zh/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 和 5xx 有什么区别?讲解大模型 API 的指数退避加全抖动、并发限制器和队列,附可直接用于生产环境的 Python 和 Node 代码。 - [按小时计费的大模型 API 密钥:每小时固定价格详解](https://broadrouter.com/zh/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): BroadNet 的按小时密钥以固定价格买下 Gemma 4 VL 26B 单路流一小时的使用权。它包含什么、有哪些限制、适合谁,以及什么时候划算。 - [大模型 API 数据保留政策对比:OpenAI、Anthropic、Gemini、Mistral](https://broadrouter.com/zh/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): OpenAI、Anthropic、Gemini API、Mistral 和 BroadNet 官方如何说明 API 提示词的存储期限、是否用于训练?引自它们 2026 年 11 月的页面原文。 - [大模型 API 密钥安全:存储、轮换与 IP 白名单](https://broadrouter.com/zh/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): 不把大模型 API 密钥写进代码和日志,每个工作负载一个密钥,定期轮换,把密钥锁定到服务器 IP,泄露时几秒内吊销。 - [大模型 API 怎么计费:输入、缓存、输出与思考 token 详解](https://broadrouter.com/zh/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): 大模型 API 按 token 计费是怎么算的?逐项讲清输入、缓存输入、输出、思考 token、图像和按分钟计费的音频,并用一份公开价目表做实例计算。 - [把敏感数据发给大模型 API:脱敏、最小化、少记日志](https://broadrouter.com/zh/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): 如何安全地把客户、健康或金融数据发给大模型 API:给字段分类、用可还原的占位符脱敏、选择零数据保留,并保持日志干净。 - [大模型 API 费用控制:预算、密钥限制与死循环防护](https://broadrouter.com/zh/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): 别让大模型 API 账单失控:预付余额、每个应用一个密钥、IP 白名单、用量日志、代码内预算、循环上限,以及固定价格密钥。 - [大模型 API 用量日志:按用户和功能算成本,脱敏 PII](https://broadrouter.com/zh/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): 为每次大模型 API 调用记录 token、成本、延迟和请求 ID,把花费归到用户和功能上,并在写入前脱敏个人信息。附 Python 代码。 - [大模型单用户成本:SaaS AI 功能的单位经济](https://broadrouter.com/zh/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): 算清一个 AI 功能每个活跃用户每月要花多少钱,为什么重度用户比平均值更重要,以及怎样给 SaaS 套餐定价,让模型账单始终有覆盖。 - [推理服务商怎样给开源模型定价:6 家对比](https://broadrouter.com/zh/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): OpenRouter、Together AI、Fireworks、Groq、DeepInfra 和 BroadNet 如何给开源模型定价:按 token、缓存、批处理、GPU 小时和固定价格套餐,附官方价格出处。 - [大模型延迟:首 token 延迟、每秒 token 数与端到端耗时](https://broadrouter.com/zh/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): 首 token 延迟(TTFT)、每秒 token 数和端到端延迟各自衡量什么,怎样用代码在客户端测量,以及哪种优化能改善哪个数字。 - [大模型流式输出详解:SSE 分块、delta、用量与取消](https://broadrouter.com/zh/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): chat completion 的流式输出在传输层是什么样子,如何从 delta 拼出完整回答、从最后一个分块读取用量,以及如何在 Python、Node 和 curl 中干净地取消。 - [长上下文还是 RAG:百万 token 什么时候能取代向量数据库](https://broadrouter.com/zh/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): 百万 token 上下文窗口什么时候能取代向量数据库,什么时候 RAG 依然更好?从单次查询成本、延迟、召回率和维护成本四个方面对比。 - [长时间运行的智能体如何管理上下文:用好 1M 窗口](https://broadrouter.com/zh/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): 用 1M token 窗口让长时间运行的智能体既准确又省钱:只追加的历史、对缓存友好的提示词、何时压缩上下文,以及优先删减哪些内容。 - [长时间运行的智能体:可靠的超时、重试与检查点](https://broadrouter.com/zh/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): 让大模型智能体连续运行数小时:为每次调用设置超时、用流式输出发现卡顿、安全重试、让工具副作用幂等,并在每一步后保存检查点。 - [用 OpenAI 兼容模型接入 MCP 服务器(任意服务商)](https://broadrouter.com/zh/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP 适用于任何大模型:由客户端把 MCP 工具转换成函数调用。讲解如何在 Cline、Continue、opencode、n8n 或你自己的 Python 代码中让开源模型使用 MCP。 - [百万 token 上下文窗口什么时候值得用(什么时候不值得)](https://broadrouter.com/zh/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): 1,048,576 个 token 能放下什么,在 DeepSeek V4.1 Flash 上要花多少钱,思考和缓存如何改变这笔账,以及什么时候 RAG 仍是更好的选择。 - [多智能体方案到底要花多少钱:按 token 计费 vs 固定价格会话](https://broadrouter.com/zh/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): 多智能体系统消耗的 token 是单个智能体的数倍。固定价格套餐与按 token 计费的会话算法、隐藏的放大因素,以及各自何时更划算。 - [多语言语音转文字 API:14 种语言,一个端点](https://broadrouter.com/zh/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): 通过一个按分钟计费的 OpenAI 兼容端点,转写英语、西班牙语、法语、中文、阿拉伯语等 14 种语言;以及什么时候该用阿拉伯语专用模型。 - [n8n AI Agent 接入低价大模型:配置、模型选择与单次运行成本](https://broadrouter.com/zh/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): 用低成本的 OpenAI 兼容模型构建 n8n AI Agent 工作流:工具调用该选哪个模型、能减少 token 的设置,以及每 10,000 次运行的成本算法。 - [企业该选开放权重还是闭源大模型:成本、控制与风险](https://broadrouter.com/zh/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): 企业该用开放权重还是闭源大模型?比较成本、控制权、隐私、质量差距和供应商锁定,以及运行开放模型的三种方式和各自适用的场景。 - [迁移到 OpenAI 兼容 API:切换服务商检查清单](https://broadrouter.com/zh/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): 改 base_url 是最简单的一步。真正容易出问题的是模型 ID、思考 token、流式输出的用量统计、错误处理、速率限制和密钥。一份附代码的检查清单。 - [AI 智能体用 OpenRouter 还是固定价格套餐:哪个更省钱?](https://broadrouter.com/zh/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter 按 token 计费,购买 credits 另收手续费;固定价格套餐按会话计费。全天运行的智能体怎么算账、OpenRouter 的费用与限制,以及各自适合谁。 - [如何安全地让编程智能体通宵无人值守运行](https://broadrouter.com/zh/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): 无人值守运行编程智能体的实用配置:沙箱、git 检查点、任务文件、步数上限、早晨审查,以及用固定价格的套餐会话跑通宵。 - [如何组织提示词,让前缀被缓存](https://broadrouter.com/zh/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): 如何安排大模型提示词的布局,让被缓存的前缀保持一致:系统提示词、工具 schema、文档、只追加的历史,外加一个能证明缓存生效的测试。 - [思考 token:为什么你付费的输出比看到的多](https://broadrouter.com/zh/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): 思考(推理)模型把隐藏的思考过程按输出 token 计费。怎样测量、各模型要花多少钱,以及怎样用 reasoning_effort、max_tokens 和模型路由来降低费用。 - [如何在固定价格大模型 API 上运行并行编程智能体](https://broadrouter.com/zh/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): 在不限量的 DeepSeek V4.1 Flash 上运行子智能体、编排器和批处理 worker:需要多少个套餐会话,如何分配到多个密钥,以及如何限制并发重叠。 - [自建部署大模型还是用 API:真实成本与盈亏平衡点](https://broadrouter.com/zh/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): 来自官方页面的 GPU 租用价格、自有 GPU 每百万 token 成本的计算公式、要比 API 便宜所需的利用率,以及常被忽略的运维成本。 - [语音转文字 API 价格对比:2026 年每分钟多少钱](https://broadrouter.com/zh/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): OpenAI、Google、AWS、Deepgram、AssemblyAI、Speechmatics、Groq 和 BroadNet Transcribe 的每分钟及每 1,000 小时价格,均来自官方页面,并列出附加功能费用。 - [JSON 模式与 json_schema:靠得住的结构化输出](https://broadrouter.com/zh/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON 模式保证合法 JSON,json_schema 保证你要的结构。讲解如何在 OpenAI 兼容模型上分别请求、用 Pydantic 校验,并用循环修复错误输出。 - [阿拉伯语与英语每词 token 数:为什么阿拉伯语更贵](https://broadrouter.com/zh/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): 为什么同样的意思,阿拉伯语往往比英语消耗更多大模型 token;为什么“每词 token 数”有误导性;以及一个从 usage 字段测量你自己比例的脚本。 - [开源模型的工具调用:schema、并行调用与重试](https://broadrouter.com/zh/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): OpenAI 兼容开源模型上的函数调用:schema 编写要点、并行工具调用、工具调用失败的五种情况,以及一个会校验并重试的 Python 循环。 - [呼叫中心录音怎么转写:一条支持阿拉伯语的流水线](https://broadrouter.com/zh/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): 按声道拆分立体声通话、切分说话轮次、转写阿拉伯语和英语、脱敏个人信息,再用大模型总结每通电话。附可运行的 Python 流水线和成本。 - [编程智能体接入开源模型:Aider、Cline、opencode、Codex](https://broadrouter.com/zh/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): 把 Aider、Cline、Roo Code、opencode、Continue 和 Codex CLI 接入任意 OpenAI 兼容端点:每个工具的准确字段名,以及最容易出错的设置。 - [企业语音消息转文字:把 WhatsApp 和 Telegram 语音变成工单](https://broadrouter.com/zh/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): 把 WhatsApp Business 和 Telegram 机器人收到的客户语音消息转成文字和客服工单:下载、转写、用大模型提取字段、分派。 - [短期项目的大模型周套餐:Orbit、按小时还是包月](https://broadrouter.com/zh/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): 一周不限量的 DeepSeek V4.1 Flash 套餐,什么时候比按小时会话、包月套餐或按 token 付费更划算?附黑客松、冲刺和代码迁移的计算方法。 - [大模型 API 套餐的“不限量”,到底意味着什么](https://broadrouter.com/zh/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): token 计量、5 小时滚动窗口、每周上限、速率限制和并发限制。用七个问题看清一个“不限量”AI 套餐对你的智能体来说是否真的不限量。 - [Whisper API 替代方案:什么时候该换,怎么测试](https://broadrouter.com/zh/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): 什么时候值得换掉 OpenAI Whisper API?从你自己音频上的词错误率、语言、每小时价格、限制和隐私几方面判断,并附测试代码。 - [零数据保留大模型 API:10 个问题的核对清单](https://broadrouter.com/zh/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): 大模型 API 的“零数据保留”应该涵盖什么:提示词、输出、日志、训练、子处理方和审计。发送数据之前要问的 10 个问题。 - [Guides in Chinese (Simplified)](https://broadrouter.com/zh/guides/) - [用 OpenAI Python SDK 调用 BroadNet Inference](https://broadrouter.com/zh/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): 只改一个 base_url,就能让官方 OpenAI Python SDK 调用 BroadNet Inference:安装、设置密钥、发出第一个请求,再把回答改成流式输出。 - [用 OpenAI Node.js SDK 调用 BroadNet Inference](https://broadrouter.com/zh/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): 设置 baseURL,让官方 OpenAI Node.js / TypeScript SDK 调用 BroadNet Inference:安装、发出第一个请求、流式输出 token,并记录请求 ID。 - [免注册,在浏览器里试用 BroadNet Inference](https://broadrouter.com/zh/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): 在 BroadNet 的浏览器 Playground 里免注册、免绑卡测试阿拉伯语语音转文字、文档识别和 Gemma 4 对话,再把同样的请求搬到 API 上。 - [用 curl 首次调用 BroadNet Inference API](https://broadrouter.com/zh/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): 用 curl 完成对 BroadNet Inference 的首次调用:一次对话补全、一次流式输出和一次音频转写,并学会读懂响应、usage 和错误。 - [对话补全的流式输出,以及连接中断的处理](https://broadrouter.com/zh/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): BroadNet Inference 上的流式输出如何工作:Server-Sent Events(SSE)、[DONE] 结束标志、最后一个数据块里的 token 用量,以及流提前结束时该怎么做。 - [推理模型为什么返回空回答](https://broadrouter.com/zh/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash、GLM 5.3 和 Qwen3.8 27B 会先思考再回答,思考内容计入 max_tokens。本文说明如何避免收到空回复。 - [BroadNet Inference 错误码,以及什么时候该重试](https://broadrouter.com/zh/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): BroadNet Inference API 每个 HTTP 状态码的含义:哪些要指数退避重试,哪些要先修正请求,以及联系技术支持时要带上请求 ID 等哪些信息。 - [用 IP 白名单把 API 密钥锁定到你的服务器](https://broadrouter.com/zh/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): 用 IP 白名单保护 BroadNet Inference 密钥,把推理密钥和控制台凭证分开,并在几秒内吊销泄露的密钥。附 curl 示例。 - [面向智能体的不限量 LLM API 套餐如何运作](https://broadrouter.com/zh/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): 不限量 DeepSeek V4.1 Flash 套餐没有 token 计量,也没有每周限额。本文说明会话、并行智能体、密钥和计费的实际运作方式。 - [用不限量的 DeepSeek V4.1 Flash 运行 Aider](https://broadrouter.com/zh/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): 通过 Aider 的 OpenAI 兼容设置把这款编程智能体接到 BroadNet Inference,在固定价格、不限量的 DeepSeek V4.1 Flash 套餐上运行。 - [在 LangChain 中调用 BroadNet Inference](https://broadrouter.com/zh/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): 用 base_url 和 api_key 把 LangChain 的 ChatOpenAI 接到 BroadNet Inference,并了解 LangChain 会保留和丢弃思考模型的哪些字段。 - [在 Cline 中使用 BroadNet Inference(OpenAI Compatible 提供商)](https://broadrouter.com/zh/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): 通过 OpenAI Compatible 提供商把 VS Code 插件 Cline 接到 BroadNet Inference:每个字段怎么填、模型 ID、上下文大小和常见问题的解决办法。 - [在 Roo Code 中使用 BroadNet Inference(OpenAI Compatible)](https://broadrouter.com/zh/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code 已于 2026 年 5 月停止维护。本文给出现有安装和分支版本的配置方法,以及接下来该迁移到哪里(Cline 或 Zoo Code 分支),并接入 BroadNet Inference。 - [在 Continue 中使用 BroadNet Inference(config.yaml 与 apiBase)](https://broadrouter.com/zh/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): 在 config.yaml 中用 provider openai 和 apiBase 把 BroadNet Inference 模型加入 Continue IDE 插件,包括 tool_use 能力和上下文长度。 - [把 BroadNet Inference 添加为 opencode 的自定义供应商](https://broadrouter.com/zh/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): 在 opencode.jsonc 中把 BroadNet Inference 添加为 opencode 编程 agent 的自定义 OpenAI 兼容供应商,并设置模型别名、能力和限制。 - [把 Open WebUI 接入 BroadNet Inference](https://broadrouter.com/zh/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): 在 Open WebUI 中把 BroadNet Inference 添加为 OpenAI 兼容连接,把模型 ID 加入白名单,并解决连接验证失败、空回复和速率限制问题。 - [把 BroadNet Inference 作为自定义端点添加到 LibreChat](https://broadrouter.com/zh/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): 在 librechat.yaml 中把 BroadNet Inference 添加为 LibreChat 的 OpenAI 兼容自定义端点:密钥放在 .env,配置模型列表、标题模型,以及常见问题的解决办法。 - [在 n8n 中使用 BroadNet Inference(OpenAI 凭据与 Base URL)](https://broadrouter.com/zh/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): 在 n8n 的 OpenAI 凭据里用 Base URL 字段指向 BroadNet Inference,再在 OpenAI Chat Model 节点中按 ID 选模型,并关闭 Responses API 开关。 - [用 Vercel AI SDK 调用 BroadNet Inference](https://broadrouter.com/zh/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): 用 @ai-sdk/openai-compatible 的 createOpenAICompatible 从 Vercel AI SDK 调用 BroadNet Inference:generateText、streamText、工具和思考强度。 - [阿拉伯语语音转文字 API 分步教程](https://broadrouter.com/zh/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): 通过 OpenAI 兼容端点,用 BroadNet Transcribe Arabic 转写阿拉伯语音频:支持的格式、language 字段、Python 和 curl 示例,以及价格。 - [用 LiteLLM 路由到 BroadNet Inference(SDK 与代理)](https://broadrouter.com/zh/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): 在 LiteLLM 的 Python SDK 或代理 config.yaml 中,用 openai/ 模型前缀和 api_base 调用 BroadNet Inference,并配好匹配的速率限制。 - [在 BroadNet Inference 上运行 OpenAI Agents SDK](https://broadrouter.com/zh/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): 通过 AsyncOpenAI 和 OpenAIChatCompletionsModel,让 Python 版 OpenAI Agents SDK 调用 BroadNet Inference,并关闭需要 OpenAI 密钥的 tracing。 - [用 API 把 WhatsApp 语音消息转成文字](https://broadrouter.com/zh/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): 用 ffmpeg 把 WhatsApp 的 .ogg 语音消息或手机的 .m4a 录音转成 WAV,再用一次 API 调用转写成阿拉伯语或其他 13 种语言的文字。 - [在 n8n 中转写 WhatsApp 语音消息](https://broadrouter.com/zh/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): 一个 n8n 工作流:接收 WhatsApp 语音消息,从 Meta 下载,用 ffmpeg 转换格式,再转写成阿拉伯语或其他 13 种语言的文字。 - [用视觉模型识别阿拉伯语发票和收据](https://broadrouter.com/zh/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): 把发票或收据照片发给 BroadNet Inference 上的视觉模型,用阿拉伯语或英语提问,以 JSON 形式拿回总金额、日期和商家。 - [在 Zed 中把 BroadNet Inference 配置为 OpenAI 兼容提供商](https://broadrouter.com/zh/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): 在 Zed 编辑器的 settings.json 中把 BroadNet Inference 添加为 OpenAI 兼容提供商,包括模型 ID、token 限制、能力开关和密钥环境变量。 - [在 Kilo Code 中使用 BroadNet Inference(OpenAI Compatible 供应商)](https://broadrouter.com/zh/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): 把 Kilo Code 接到 BroadNet Inference,作为 OpenAI Compatible 自定义供应商:可在 VS Code 对话框或 kilo.json 中配置模型 ID、上限和工具开关。 - [把 BroadNet Inference 作为自定义供应商接入 goose](https://broadrouter.com/zh/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): 把 goose AI agent 接到 BroadNet Inference,作为 OpenAI Compatible 自定义供应商:可以用 goose Desktop、goose configure 或 JSON 文件配置。 - [在 BroadNet Inference 上使用 Pydantic AI(OpenAICompatibleProvider)](https://broadrouter.com/zh/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): 用 OpenAIChatModel 和 OpenAICompatibleProvider 在 BroadNet Inference 上运行 Pydantic AI agent,包括工具调用、结构化输出和模型 profile 的调整。 - [在 CrewAI 中把 BroadNet Inference 用作自定义 OpenAI 端点](https://broadrouter.com/zh/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): 用 LLM(custom_openai=True, base_url=...) 把 CrewAI 智能体指向 BroadNet Inference,再按工具调用、长上下文和固定价格套餐选择模型。 - [通过 OpenAILike 在 LlamaIndex 中使用 BroadNet Inference](https://broadrouter.com/zh/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): 用 OpenAILike 类把 LlamaIndex 接入 BroadNet Inference,并正确设置 api_base、context_window、is_chat_model 和 is_function_calling_model。 - [用 OpenAI-API-compatible 插件把 BroadNet Inference 接入 Dify](https://broadrouter.com/zh/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): 通过 OpenAI-API-compatible 模型供应商插件把 Dify 接到 BroadNet Inference:每个字段的准确取值、上下文大小和函数调用设置。 - [在 Jan 中把 BroadNet Inference 添加为 OpenAI 兼容的自定义供应商](https://broadrouter.com/zh/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): 在 Jan 桌面应用中把 BroadNet Inference 添加为 OpenAI 兼容的自定义供应商,手动添加模型 ID,并为每个模型开启工具和视觉能力。 - [把 SillyTavern 接入 BroadNet Inference(自定义 OpenAI 兼容端点)](https://broadrouter.com/zh/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): 用 Custom (OpenAI-compatible) 这个 Chat Completion 来源把 SillyTavern(酒馆)接入 BroadNet Inference,包括正确的模型 ID、提示词后处理和 token 上限。 - [在 BroadNet Inference 上使用 Spring AI(OpenAI base-url)](https://broadrouter.com/zh/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): 用 spring.ai.openai.base-url、BroadNet 模型 ID 和 max-tokens 把 Spring AI 的 OpenAI chat starter 指向 BroadNet Inference,再通过 ChatClient 调用。 - [用 Go 和 openai-go SDK 调用 BroadNet Inference](https://broadrouter.com/zh/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): 通过 option.WithBaseURL,用官方 openai-go v3 SDK 调用 BroadNet Inference:Go 里的对话补全、带用量统计的流式输出和错误处理。 - [在 BroadNet Inference 上构建 LangGraph 智能体](https://broadrouter.com/zh/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): 用 ChatOpenAI、base_url 和 BroadNet 模型 ID 在 BroadNet Inference 上运行 LangGraph 的 StateGraph 或 create_agent,并配置工具、流式输出和记忆。 - [在 Haystack 的 pipeline 和 agent 中使用 BroadNet Inference](https://broadrouter.com/zh/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): 用 api_base_url、BroadNet 模型 ID 和 Secret 密钥,把 Haystack 的 OpenAIChatGenerator 指向 BroadNet Inference,支持 pipeline、工具、流式输出和 agent。 - [在 BroadNet Inference 上使用 Semantic Kernel](https://broadrouter.com/zh/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): 把 Microsoft Semantic Kernel 接入 BroadNet Inference:Python 用 AsyncOpenAI 客户端,C# 用自定义端点,支持对话、插件和智能体。 - [在 BroadNet Inference 上运行 Mastra 智能体](https://broadrouter.com/zh/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): 用一个设置了 id、url 和 apiKey 的 model 对象,在 BroadNet Inference 上运行 Mastra 智能体,并说明模型 ID 为什么要写两遍 broadnet/ 前缀。 - [在 BroadNet Inference 上用 Instructor 获取结构化输出](https://broadrouter.com/zh/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): 用 Instructor 从 BroadNet Inference 提取 Pydantic 对象:通过 from_openai 或 from_provider 设置自定义 base_url,并选对 Mode。 - [在 DSPy 中使用 BroadNet Inference(设置 model_type="chat")](https://broadrouter.com/zh/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): 为 BroadNet Inference 配置 dspy.LM:使用 openai/ 前缀、api_base 和 model_type="chat",因为 DSPy 3.5 会把 OpenAI 模型发到 Responses API。 - [用 promptfoo 评测 BroadNet 的模型](https://broadrouter.com/zh/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): 用 openai:chat 供应商、apiBaseUrl 和 apiKeyEnvar,在 BroadNet Inference 上运行 promptfoo 评测和模型并排对比测试,并配置评分模型。 - [用 OpenAIModel 在 BroadNet Inference 上运行 smolagents](https://broadrouter.com/zh/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): 用 OpenAIModel 和 api_base 把 Hugging Face smolagents 接入 BroadNet Inference:CodeAgent 可用任意模型,ToolCallingAgent 用 GLM 5.3 和 Qwen3.8 27B。 - [把 Flowise 接入 BroadNet Inference(OpenAI Custom Model 节点)](https://broadrouter.com/zh/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): 在 Flowise chatflow 中用 OpenAI Custom Model 节点调用 BroadNet Inference:设置 Base Path、OpenAI API 凭证,并手动填写 BroadNet 模型名。 - [在 Langflow 中接入 BroadNet Inference](https://broadrouter.com/zh/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): 在 Langflow 中通过 OpenAI 组件的 OpenAI API Base 字段和手动填写的模型名使用 BroadNet Inference,并说明为什么 OpenAI Compatible 服务商列不出它的模型。 - [Glossary entries in Chinese (Simplified)](https://broadrouter.com/zh/glossary/) - [AI 智能体(AI agent)](https://broadrouter.com/zh/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): AI 智能体是由语言模型自主决定步骤、朝目标推进的系统:它决定调用哪些工具、读取结果,然后重复,直到给出最终回答或触及上限。在固定位置调用模型的固定脚本,通常称为工作流(workflow)。 - [API 密钥(API key)](https://broadrouter.com/zh/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): API 密钥是一串用于识别和授权 API 调用方的秘密字符串。大模型 API 通常要求把它放在 Authorization: Bearer 请求头里,任何持有该密钥的人都能发起请求,费用记在拥有该密钥的账户上。 - [阿拉伯语变音符号(tashkeel)](https://broadrouter.com/zh/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): 阿拉伯语变音符号(tashkeel)是写在字母上方或下方的小符号,用来标示短元音、辅音重叠和无元音。它们是可选的:包括新闻在内,大多数阿拉伯语文本都省略它们,读者和软件需要根据上下文推断元音。 - [阿拉伯语方言识别(Arabic dialect identification)](https://broadrouter.com/zh/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): 阿拉伯语方言识别是判断一段文本或录音属于哪种阿拉伯语变体的任务,例如埃及、海湾、黎凡特、马格里布方言或现代标准阿拉伯语,粒度可以是地区、国家或城市。它用于消息路由、选择提示词和拆分评测集。 - [阿拉伯语文本规范化(Arabic text normalization)](https://broadrouter.com/zh/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): 阿拉伯语文本规范化是在匹配之前把同一个词的不同拼写统一成一种形式的步骤:去掉变音符号和 tatweel,统一 alef 的各种写法,通常还把 ى 映射为 ي、把 ة 映射为 ه。它用于搜索键和评测,不用于展示的文本。 - [Arabizi(拉丁字母阿拉伯语)](https://broadrouter.com/zh/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): Arabizi 是用拉丁字母非正式书写的阿拉伯语,没有拉丁对应的阿拉伯语音用数字表示,例如用 3 代表 ع。它在聊天和社交媒体中很常见,以方言为主,没有标准拼写,还经常夹杂英语或法语。 - [AWQ(激活感知权重量化)](https://broadrouter.com/zh/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ(激活感知权重量化)是一种训练后量化方法,把大模型的权重压缩到低精度,通常是 4 位整数。它用一小份校准集的激活值统计找出最重要的权重通道,并对它们做缩放,以减少精度损失。它不需要反向传播。 - [Base URL(base_url)](https://broadrouter.com/zh/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): base URL 是 API 固定的根地址,例如 https://api.example.com/v1,客户端会把它加在每个端点路径(例如 /chat/completions)前面。在 OpenAI 兼容 SDK 中,修改 base URL,再配合 API 密钥和模型 ID,就能让同一份代码改用另一家服务商。 - [批处理 API(Batch API)](https://broadrouter.com/zh/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): 批处理 API 是一种异步的大模型端点:它一次接收大量请求(通常以文件形式),在设定的时间窗口内处理而不是立即处理,最后统一返回结果。OpenAI 和 Anthropic 等服务商对批处理请求的定价低于其标准 API。 - [BM25](https://broadrouter.com/zh/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25(Okapi BM25)是关键词搜索的排序函数,它根据查询词在文档中出现的频率、这些词在整个文档集中有多罕见,以及文档有多长来给文档打分。它是 Lucene 和 Elasticsearch 默认的相似度算法。 - [字节对编码(BPE,byte pair encoding)](https://broadrouter.com/zh/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): 字节对编码(BPE)是一种子词分词算法:从单个字符或字节出发,反复把出现频率最高的相邻符号对合并成新的词表条目,直到达到目标词表大小。当前大多数大模型分词器使用 BPE 或它的字节级变体。 - [缓存输入 token(cached input tokens)](https://broadrouter.com/zh/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): 缓存输入 token 是大模型服务商近期已处理过、可以直接从缓存中复用而无需重新计算的提示词 token,通常是重复的前缀,例如系统提示词或文档。服务商对它们按低于新输入 token 的单价计费。 - [思维链(CoT,chain-of-thought)](https://broadrouter.com/zh/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): 思维链(CoT)是一种提示技巧:给语言模型示范或要求它在给出最终答案之前,先写出中间的推理步骤。它能提高算术、逻辑等多步问题的准确率,如今思考模型会自动生成思维链。 - [字错误率(CER,character error rate)](https://broadrouter.com/zh/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): 字错误率(CER)在字符层面衡量转写或 OCR 的准确度:把输出变成参考文本所需的替换、删除和插入次数,除以参考文本的字符数。越低越好,0 表示完全一致。 - [Chat Completions API(对话补全接口)](https://broadrouter.com/zh/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Chat Completions API 是 OpenAI 的 HTTP 端点 POST /v1/chat/completions,它接收模型名称和一组带角色标记的消息,返回模型的下一条消息。它的请求与响应结构已成为 OpenAI 兼容服务商和各类 SDK 通用的格式。 - [对话模板(chat template)](https://broadrouter.com/zh/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): 对话模板是一套格式规则,通常用 Jinja 编写并随模型的分词器一起发布,它把带角色的对话消息列表转换成模型期望的单一 token 序列,包括模型专用的控制 token。用错模板会让模型的回答质量大幅下降。 - [文本分块(chunking)](https://broadrouter.com/zh/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): 分块是检索流水线中把文档切成较小片段(称为块,chunk)的步骤,在嵌入和建索引之前进行。块大小决定每个片段容纳多少文本,块重叠则让相邻的块重复一部分文本,避免在边界处被切断的内容丢失。 - [语码转换(code-switching)](https://broadrouter.com/zh/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): 语码转换是在同一段对话或同一句话中使用两种或更多语言,或同一语言的不同变体,口语和书面语都有。例如讲阿拉伯语的人在方言里夹杂英语单词,或在现代标准阿拉伯语和方言之间来回切换。 - [AI 编程智能体(AI coding agent)](https://broadrouter.com/zh/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): AI 编程智能体是一种软件工具,其中的语言模型在循环中处理代码库:读取文件、编辑文件、运行测试等命令、读取输出并决定下一步,直到任务完成或达到某个上限。 - [并发(concurrency)](https://broadrouter.com/zh/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): 并发(concurrency)是客户端在同一时刻正在进行中的请求数。大模型 API 往往按密钥或套餐单独限制并发,与每分钟请求数分开计算,因为每个未完成的请求在响应结束前都会占用 GPU 资源。 - [上下文压缩(context compaction)](https://broadrouter.com/zh/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): 上下文压缩是指在对话接近上下文窗口上限时对其进行摘要,并从这份摘要继续会话的做法,这样智能体能保留关键决策和状态,同时丢掉旧的工具输出等原始历史。 - [上下文工程(context engineering)](https://broadrouter.com/zh/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): 上下文工程是选择并维护模型每次请求所见 token 集合的工作,包括系统提示词、工具定义、历史记录、检索文档和笔记,让模型以最少的高信号信息把任务做好。 - [上下文腐化(context rot)](https://broadrouter.com/zh/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): 上下文腐化是指语言模型的输入变长时,准确性和可靠性逐渐下降的现象,即使任务不变、输入远未超出上下文窗口也会发生。这个术语因 Chroma 2025 年一份测试了 18 个模型的技术报告而流行。 - [上下文窗口(context window)](https://broadrouter.com/zh/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): 上下文窗口是语言模型在单次请求中能处理的最大 token 数。它涵盖提示词、对话历史、工具定义、图片以及生成的输出(包括思考 token)。放不下的内容必须删减、摘要,或拆分到多个请求中。 - [数据驻留(data residency)](https://broadrouter.com/zh/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): 数据驻留是指数据须在特定国家或地区存储和处理的要求或承诺。对大模型 API 而言,它涵盖推理时提示词在哪里处理、存储的内容放在哪里,通常不包括账户和计费数据。 - [阿拉伯语方言(dialectal Arabic)](https://broadrouter.com/zh/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): 阿拉伯语方言是某个地区或城市日常使用的口语阿拉伯语,例如埃及、海湾、黎凡特或马格里布阿拉伯语。方言在语音、词汇和语法上都不同于现代标准阿拉伯语,没有标准拼写,而且越来越多地出现在网络文字中,这让语言模型更难处理。 - [说话人分离(speaker diarization)](https://broadrouter.com/zh/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): 说话人分离是把一段录音按说话人切分成片段的任务,回答“谁在何时说话”。它给每段语音打上 SPEAKER_00 这样的匿名说话人标签,通常与转写文本结合,显示每句话是谁说的。 - [知识蒸馏(knowledge distillation)](https://broadrouter.com/zh/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): 蒸馏(知识蒸馏)是一种训练方法:让较小的学生模型学习复现较大教师模型的输出,例如其概率分布或生成的回答,使学生模型以更低的运行成本保留教师模型的大部分质量。 - [数据处理协议(DPA)](https://broadrouter.com/zh/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): 数据处理协议(DPA)是控制者与处理者之间的合同,规定处理者如何处理个人数据。GDPR 第 28(3) 条要求签订此类合同,并列出其内容:书面指示、保密、安全措施、次级处理者规则、数据删除或返还,以及审计。 - [嵌入向量(embeddings)](https://broadrouter.com/zh/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): 嵌入向量(embeddings)是模型为一段文本、一张图片或其他输入生成的浮点数列表(向量)。意思相近的输入会得到彼此接近的向量,因此可以用向量之间的距离来给搜索结果排序、对文档聚类或查找相关条目。 - [评测(eval,LLM evaluation)](https://broadrouter.com/zh/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): 评测(eval,evaluation 的简称)是对语言模型或提示词的可重复测试:一组固定的输入、对好输出的描述,以及给每个输出打分的评分器(代码、人工或另一个模型),以便比较不同版本。 - [指数退避(exponential backoff)](https://broadrouter.com/zh/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): 指数退避(exponential backoff)是一种重试策略:客户端每失败一次,就等待更长时间再重试,通常把延迟翻倍,直到达到上限。在每次等待中加入随机抖动(jitter),可以把众多客户端的重试错开,避免它们在同一时刻打到服务器上。 - [少样本提示(few-shot prompting)](https://broadrouter.com/zh/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): 少样本提示是一种在真实输入之前,先在提示词中放入少量“输入—输出”示例的技巧,让语言模型从中推断任务和格式。模型权重不会改变;零样本不用示例,单样本只用一个。 - [微调(fine-tuning)](https://broadrouter.com/zh/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): 微调是在一个较小的特定任务数据集(例如示例对话)上继续训练预训练语言模型的过程,使其权重发生变化,从而更稳定地完成该任务或风格,而无需在每个提示词中写长指令或示例。 - [finish_reason(结束原因)](https://broadrouter.com/zh/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason 是对话补全响应中每个 choice 上的字段,说明模型为什么停止生成。在 OpenAI 格式中,取值有 stop、length、tool_calls、content_filter 以及已弃用的 function_call。Anthropic 的 Messages API 用 stop_reason 表达同样的含义。 - [大模型包月套餐(flat-rate LLM plan)](https://broadrouter.com/zh/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): 大模型包月套餐(固定价格套餐)是按一个周期(如一个月)收取固定费用、提供大模型 API 使用权的计价方式,不按 token 计费。这类套餐通常会限制别的东西,如并发、速度、模型或使用时间窗口,所以限制条件和价格同样重要。 - [FP8(8 位浮点数)](https://broadrouter.com/zh/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 是深度学习中使用的 8 位浮点数格式,有两种编码:E4M3(4 位指数、3 位尾数)和 E5M2(5 位指数、2 位尾数)。与 16 位格式相比,它的显存占用减半,并在较新的 GPU 上原生运行。 - [频率惩罚与存在惩罚(frequency / presence penalty)](https://broadrouter.com/zh/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): 频率惩罚与存在惩罚是一类采样设置,会降低输出中已出现 token 的分数。存在惩罚在 token 出现过一次后就生效;频率惩罚随每次重复而加大。在 OpenAI 的 API 中,两者的取值范围都是 -2.0 到 2.0,默认值为 0。 - [GDPR(通用数据保护条例)](https://broadrouter.com/zh/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): GDPR(《通用数据保护条例》,欧盟第 2016/679 号条例)是欧盟关于个人数据处理的法律,自 2018 年 5 月 25 日起适用,覆盖在欧盟设立的组织以及部分欧盟以外的组织。对大模型 API 而言,它管辖处理者合同、跨境传输和数据最小化。 - [GGUF](https://broadrouter.com/zh/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF 是一种二进制文件格式,用于存储语言模型,供 llama.cpp 等基于 GGML 的运行时推理。一个 GGUF 文件包含权重(通常已量化)以及加载模型所需的元数据,如分词器和架构信息,并且可以通过内存映射快速加载。 - [GPU 小时(GPU-hour)](https://broadrouter.com/zh/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): GPU 小时是指一块图形处理器使用一小时。云服务商按小时或按秒对租用的 GPU 实例计费,研究人员也用 GPU 小时来说明一项任务消耗了多少算力。你为机器的时间付费,无论它是否在忙。 - [事实依据(grounding)](https://broadrouter.com/zh/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): Grounding(事实依据)是把语言模型的输出与具体、可核查的信息来源关联起来的做法,例如提供的文档、搜索结果或数据库记录,并要求模型依据这些来源作答。有依据的回答可以追溯到来源,也更不容易编造事实。 - [护栏(guardrails)](https://broadrouter.com/zh/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): 护栏是应用在调用语言模型前后运行的检查,让调用保持在设定的范围内。输入护栏筛查进入模型的内容,输出护栏检查模型产出的内容,工具护栏在操作执行前进行校验。检查不通过时,会拦截、改写或标记该请求。 - [AI 幻觉(hallucination)](https://broadrouter.com/zh/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): 幻觉是语言模型输出的、听起来合理但实际错误、凭空编造或没有输入依据的内容,例如虚构的引用、错误的数字或不存在的 API。它源于模型预测"可能的文本",而不是核查事实。 - [海姆宰(hamza)](https://broadrouter.com/zh/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): Hamza(ء,海姆宰)是阿拉伯语表示声门塞音的字母。它可单独书写,也可写在 alef(أ إ)、waw(ؤ)或 yeh(ئ)上,آ 则是 alef 加 madda。由于输入文本对它的写法不一致,搜索引擎和语音评分常把 alef 形式统一为 ا。 - [混合检索(hybrid search)](https://broadrouter.com/zh/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): 混合检索是一种检索方法:对同一个查询同时运行关键词检索(如 BM25)和基于嵌入向量的向量检索,再把两份排序列表合并成一份,常用倒数排名融合(RRF)或归一化分数的加权求和。 - [幂等键(idempotency key)](https://broadrouter.com/zh/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): 幂等键是客户端附在会改变状态的请求(如付款或购买)上的唯一值,通常是 UUID 之类的随机 ID。如果用同一个键再次发送同一请求,服务器会返回原来的结果,而不会把操作执行两次。 - [图像 token(image tokens)](https://broadrouter.com/zh/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): 图像 token 是视觉语言模型用来在提示词中表示一张图片的输入 token。服务商根据图片尺寸计算其数量,通常是先缩放、再切成固定大小的图块,并像文本输入 token 一样计费,所以图片越大越贵。 - [推理服务商(inference provider)](https://broadrouter.com/zh/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): 推理服务商是在自有、租赁或租用的 GPU 上运行机器学习模型(通常是开放权重的大模型),并通过 API 出售使用权的公司,通常按 token、按请求或按时间计价。它负责模型部署、扩缩容和硬件。 - [输入 token 与输出 token](https://broadrouter.com/zh/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): 输入 token 是语言模型在一次请求中读取的 token,包括系统提示词、对话历史、工具定义和图片。输出 token 是模型生成回复的 token,包括思考内容。大模型 API 通常分别计价,输出一般更贵。 - [INT4(4-bit 量化)](https://broadrouter.com/zh/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 是一种能表示 16 个等级的 4-bit 整数格式。在大模型推理中,它通常指 4-bit 仅权重量化(W4A16):权重以 INT4 加缩放因子存储,计算时再展开为 16-bit,显存占用约为 16-bit 权重的四分之一。 - [IP 白名单(IP allow list)](https://broadrouter.com/zh/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): IP 白名单是一种安全控制:只接受来自列出的 IP 地址或 CIDR 网段的请求,拒绝其他所有请求。用在 API 上,它能限制密钥泄露造成的损失,因为密钥在不在列表中的网络上无法使用。 - [越狱(LLM jailbreak)](https://broadrouter.com/zh/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): 越狱是专门构造的输入,目的是让语言模型无视其安全训练或部署应用设定的规则,从而生成它通常会拒绝的内容或执行这类操作。OWASP 把越狱视为提示词注入的一种形式。 - [JSON 模式(JSON mode)](https://broadrouter.com/zh/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): JSON 模式(JSON mode)是大模型 API 的一项设置,在 OpenAI 风格的 API 中写作 response_format {"type": "json_object"},它让模型的回复成为合法、可解析的 JSON。与使用 JSON Schema 的结构化输出不同,它不保证对象里有哪些键、类型或取值。 - [JSON Schema](https://broadrouter.com/zh/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema 是一套用 JSON 编写的词汇,用于描述和校验 JSON 数据的结构:类型、必填字段、允许的取值和嵌套。当前版本是 2020-12。大模型 API 用它来定义工具参数,以及结构化输出必须遵循的确切结构。 - [KV 缓存(KV cache)](https://broadrouter.com/zh/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): KV 缓存是 Transformer 语言模型存放已处理的每个 token 的注意力键(key)和值(value)的显存,让每个新 token 直接复用它们,而不必重新计算整个序列。它随上下文长度增长,常常决定一张 GPU 能同时服务多少请求。 - [延迟(latency)](https://broadrouter.com/zh/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): 对于语言模型 API,延迟是一个请求从发出到收到最后一个输出 token 所用的时间。它等于首 token 延迟加上生成时间,通常以 p50、p95 等百分位而不是平均值来报告。 - [LLM-as-a-judge(大模型评审)](https://broadrouter.com/zh/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge 是一种评测方法:由一个语言模型给其他模型的输出打分,要么按评分标准给单个回答打分,要么从两个回答中挑出更好的一个。它以低得多的成本近似人工判断,但有已知偏差,比如偏爱排在前面或更长的回答。 - [大模型基准测试(LLM benchmark)](https://broadrouter.com/zh/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): 大模型基准测试是一组固定、通常公开的任务和评分方法,用来在某一项能力上衡量和比较语言模型,比如知识问答、编程或语音识别。基准分数反映的是模型在该测试上的相对能力,而不是它在你自己的业务负载上的表现。 - [大模型路由(LLM router)](https://broadrouter.com/zh/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): LLM 路由是位于应用与一个或多个语言模型 API 之间的软件或服务,它根据成本、延迟、可用性或任务类型等规则,为每个请求决定由哪个模型或服务商来处理。 - [Logprobs(对数概率)](https://broadrouter.com/zh/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Logprobs 是语言模型为其生成的各个 token 给出的对数概率。返回 logprobs 的 API 会显示每个输出 token 被模型判断的可能性有多大,还可以选择列出该位置上最可能的候选 token。 - [LoRA(低秩适配)](https://broadrouter.com/zh/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA(低秩适配)是一种参数高效的微调方法:冻结预训练模型的权重,只为选定的层训练一对小的低秩矩阵。学到的更新可以存成一个小的适配器,也可以合并进权重,不会增加推理延迟。 - [中间迷失(lost in the middle)](https://broadrouter.com/zh/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): 中间迷失是指语言模型的一种已观察到的倾向:相关信息出现在长输入的开头或结尾时,模型利用得最好;出现在中间时,利用得最差。这一名称来自 Liu 等人 2023 年的研究。 - [max_tokens(最大输出 token 数)](https://broadrouter.com/zh/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens 是大模型 API 的一个参数,用于设定模型在一次回复中最多能生成多少 token。生成达到上限时就会停止,响应会被标记为截断,例如 finish_reason 为 "length"。 - [MCP(模型上下文协议,Model Context Protocol)](https://broadrouter.com/zh/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP(模型上下文协议)是一个基于 JSON-RPC 2.0 的开放协议,规范了大模型应用连接外部工具和数据的方式。MCP 服务器提供工具、资源和提示词;宿主应用中的 MCP 客户端发现它们,并把模型发出的工具调用转交给服务器。 - [混合专家模型(MoE)](https://broadrouter.com/zh/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): 混合专家模型(MoE)是一种神经网络设计:一层中包含许多专家子网络,由一个经过训练的路由器把每个 token 只发送给其中少数几个。MoE 语言模型的总参数量很大,但每个 token 只使用一小部分激活参数。 - [模型弃用(model deprecation)](https://broadrouter.com/zh/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): 模型弃用是指 API 服务商宣布某个模型将停止提供,通常会给出下线(retirement 或 shutdown)日期和推荐的替代模型。过了这个日期,对该模型的请求会失败,所以应用必须提前迁移到另一个模型 ID。 - [模型回退(model fallback)](https://broadrouter.com/zh/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): 模型回退是一种可靠性模式:失败的 LLM 请求会按一个有序列表在另一个模型或服务商上重试。触发条件可以是错误、速率限制、上下文长度错误或内容策略拒绝。它运行在你的客户端里,或运行在 API 前面的网关中。 - [模型许可证(model license)](https://broadrouter.com/zh/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): 模型许可证是规定模型权重、代码和输出可以如何使用、修改和分发的法律条款。开放权重模型采用 Apache 2.0、MIT 等标准软件许可证,或采用自定义模型许可证,后者可能附加条件,例如限制大型商业服务。 - [现代标准阿拉伯语(MSA)](https://broadrouter.com/zh/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): 现代标准阿拉伯语(MSA)是阿拉伯世界在书面、新闻、教育和正式讲话中使用的标准化阿拉伯语。它不是任何人日常使用的母语:人们平时说埃及阿拉伯语、海湾阿拉伯语等地区方言,在正式场合才切换到 MSA。 - [多智能体系统(multi-agent system)](https://broadrouter.com/zh/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): 多智能体系统是一种大模型应用:多个智能体各自拥有提示词、工具和上下文,分别处理同一任务的不同部分。常见设计是编排者-工作者模式:主智能体负责规划,并行启动多个工作智能体,再汇总它们的结果。 - [多模态 AI(multimodal AI)](https://broadrouter.com/zh/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): 多模态 AI 是能处理多种数据类型(如文本、图像、视频和音频)的模型或系统。多模态模型可以同时读取一张照片和一个问题,或者输入语音、输出文本。视觉语言模型是最常见的一类。 - [命名实体识别(NER)](https://broadrouter.com/zh/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): 命名实体识别(NER)是在文本中找出名称、并为每个名称标注类型(如人物、机构或地点)的任务。传统系统给每个 token 打标签;大模型则通常被要求以 JSON 返回实体列表,再与原文核对。 - [光学字符识别(OCR)](https://broadrouter.com/zh/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): 光学字符识别(OCR)是一种技术,能检测图片、扫描件或 PDF 中的印刷或手写文字,并转换成机器可读的字符。传统 OCR 输出文字及其在页面上的位置;视觉语言模型则能直接读取文字并回答相关问题。 - [Ogg Opus](https://broadrouter.com/zh/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus 是用 Opus 编解码器(RFC 6716)编码、存放在 Ogg 容器(RFC 7845)中的音频。它专为低码率下的语音和实时交互音频设计,也是 WhatsApp 语音消息使用的格式,文件通常为 .ogg 或 .opus。 - [开放权重模型(open-weight model)](https://broadrouter.com/zh/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): 开放权重模型是公开发布训练好的权重、任何人都可以下载和运行的机器学习模型,通常附带有条件的许可证。它不同于完全开源的 AI,后者还会公开训练代码和训练数据的详细信息,让其他人能够研究并复现模型。 - [OpenAI 兼容 API(OpenAI-compatible API)](https://broadrouter.com/zh/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): OpenAI 兼容 API 是一种 HTTP API,它接受与 OpenAI API 相同的请求和响应格式,例如 POST /v1/chat/completions,因此 OpenAI 的 SDK 和工具只需修改 base URL、API 密钥和模型名称就能使用它。 - [并行工具调用(parallel tool calls)](https://broadrouter.com/zh/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): 并行工具调用是指模型在一次响应中返回多个工具调用,例如同时读取三个文件。在 OpenAI 格式中,它们表现为 tool_calls 里的多个条目;把 parallel_tool_calls 设为 false,可把模型限制为每轮调用零个或一个工具。 - [按量付费(pay-as-you-go)](https://broadrouter.com/zh/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): 按量付费是一种计价模式:你只为实际用量付费,按 token、请求次数或音频分钟等单位计量,没有订阅费,也没有最低消费承诺。大多数大模型 API 都这样计费,从预付余额中扣减或定期开具账单。 - [PDPL(沙特与阿联酋个人数据保护法)](https://broadrouter.com/zh/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL(个人数据保护法)通常既指沙特阿拉伯依皇家法令 M/19 颁布、由 SDAIA 执行的个人数据法,也指阿联酋 2021 年第 45 号联邦法令。两者都规范个人数据处理,并对向境外传输(包括发往外国大模型 API 的提示词)设定条件。 - [个人身份信息(PII)](https://broadrouter.com/zh/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): 个人身份信息(PII)是可用于区分或追溯个人身份的信息,例如姓名、身份证号或生物特征记录,以及与该人关联或可关联的其他信息。GDPR 用的是更宽泛的“个人数据”。在大模型应用中,PII 脱敏会在提示词离开你的系统之前将其移除。 - [Prefill 与 Decode(预填充与解码)](https://broadrouter.com/zh/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill(预填充)和 Decode(解码)是大模型推理的两个阶段。Prefill 阶段模型并行处理整个提示词并填充 KV 缓存;Decode 阶段模型复用该缓存,逐个 token 生成回复。Prefill 主要决定首 token 延迟,Decode 决定输出速度。 - [预付额度(prepaid credit)](https://broadrouter.com/zh/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): 预付额度是在使用前存入 API 账户的资金。每个请求的费用从余额中扣除,余额用完后,计量计费的请求会停止,直到你充值。它把支出上限锁定在你已支付的金额之内。 - [提示词工程(prompt engineering)](https://broadrouter.com/zh/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): 提示词工程是编写、组织和测试发给语言模型的指令的过程,目的是让模型稳定地产出符合要求的结果。它包括角色消息、清晰的指令、示例、格式,以及用测试用例衡量结果。 - [提示词注入(prompt injection)](https://broadrouter.com/zh/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): 提示词注入是一种攻击:用户输入的文本,或藏在网页、邮件、文件、工具结果中的文本,被提供给语言模型后,违背开发者的意图改变了模型的行为。OWASP 在其大模型应用 Top 10 中把它排在第一位(LLM01:2025)。 - [量化(quantization)](https://broadrouter.com/zh/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): 量化是把神经网络中的数值(主要是权重)从 16 位改用 8 位或 4 位等更低精度格式存储的过程,这样模型占用的显存更少,通常也运行得更快。好的量化方法能把精度损失控制得很小。 - [RAG(检索增强生成)](https://broadrouter.com/zh/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG(检索增强生成)是一种模式:应用先从外部存储(例如搜索引擎或向量索引)中检索与问题相关的段落,再把它们加入提示词,让语言模型依据这些文本回答,而不只是凭记忆作答。 - [速率限制(rate limit)](https://broadrouter.com/zh/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): 速率限制是 API 对客户端在一个时间窗口内的使用量设定的上限,例如每分钟请求数或每分钟 token 数。超出限制的请求会被拒绝,通常返回 HTTP 429 Too Many Requests,直到用量回落到限制以下。 - [ReAct(推理与行动)](https://broadrouter.com/zh/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct(推理与行动)是 Yao 等人于 2022 年提出的一种提示模式:语言模型把推理过程与搜索、工具调用等行动交替进行。它读取每次的观察结果,更新计划,如此循环,直到能给出最终答案。 - [思考 token(reasoning tokens)](https://broadrouter.com/zh/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): 思考 token(reasoning tokens)是推理模型在写出可见回答之前、为解决问题而生成的 token。它们通常按输出 token 计费,计入输出上限和上下文窗口,API 可能将其隐藏或只返回摘要。 - [请求超时(request timeout)](https://broadrouter.com/zh/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): 请求超时(request timeout)是 HTTP 客户端在放弃请求之前等待服务器的最长时间。客户端通常把它拆成建立连接的连接超时和等待数据的读取超时。大模型调用需要宽松的读取超时,因为长回答本来就要花时间。 - [重排序(reranking)](https://broadrouter.com/zh/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): 重排序是检索的第二阶段:用一个更精确但更慢的模型(通常是交叉编码器),对初次检索得到的每个“查询-候选文档”对打分并重新排序,只把最相关的少数几条交给语言模型。 - [从右到左(RTL)文本](https://broadrouter.com/zh/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): 从右到左(RTL)文本是指阿拉伯语、希伯来语等从右向左显示的文字。计算机按逻辑阅读顺序存储它;当它与从左到右的文字、数字、模型名或代码混排时,由 Unicode 双向算法决定显示顺序。 - [采样率(sample rate)](https://broadrouter.com/zh/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): 采样率是每秒采集的音频样本数,单位为赫兹(Hz)。语音识别模型通常要求 16,000 Hz(16 kHz)单声道音频。采样率与位深、声道数一起,决定了未压缩音频每分钟的文件大小。 - [seed 参数(随机种子)](https://broadrouter.com/zh/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): seed 参数是随请求发送的一个整数,让 API 在重复同一请求时尽量采样出相同的 token。OpenAI 称其为尽力而为:不保证确定性,system_fingerprint 字段会显示后端配置何时发生了变化。 - [Serverless 推理(serverless inference)](https://broadrouter.com/zh/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): Serverless 推理是一种运行机器学习模型的方式:平台自动启动、扩缩和停止端点背后的算力,空闲时常常缩到零,并按使用量而不是按预留机器计费。代价是可能出现冷启动延迟。 - [SLA(服务等级协议)](https://broadrouter.com/zh/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): SLA(服务等级协议)是服务商在合同中对可衡量服务水平的承诺,通常是每月可用性百分比,并规定未达标时的补救方式,通常是服务抵扣。SLA 会定义什么算停机、哪些情况除外,以及如何申请赔付。 - [投机解码(speculative decoding)](https://broadrouter.com/zh/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): 投机解码是一种推理技术:由一个小而快的草稿模型提出接下来的若干 token,大的目标模型在一次前向计算中验证它们,并保留被接受的部分。它能加快生成,而不改变输出的分布。 - [语音转文字(speech-to-text)](https://broadrouter.com/zh/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): 语音转文字,也叫自动语音识别(ASR),是把口语音频转换成书面文字的技术。语音转文字 API 接收音频文件或音频流并返回转写文本,通常按音频分钟计价。准确率用词错误率衡量,会随口音、方言和录音质量变化。 - [消费上限(spend limit)](https://broadrouter.com/zh/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): 消费上限是对 API 账户、项目或密钥在一段时间(通常是一个月)内可花费金额的限制。用量达到上限后,服务商会拒绝后续请求,常返回 HTTP 429 或 400 错误,直到周期重置或上限被调高。 - [SRT 字幕(SRT subtitles)](https://broadrouter.com/zh/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT(SubRip)是一种纯文本字幕格式。每条字幕由序号、写成 hh:mm:ss,mmm --> hh:mm:ss,mmm 的开始和结束时间、一行或多行文字,以及一个空行组成。WebVTT 是 HTML 视频使用的类似网页格式。 - [SSE(服务器发送事件,Server-sent events)](https://broadrouter.com/zh/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): SSE(服务器发送事件,Server-sent events)是一项 Web 标准,让服务器通过一个内容类型为 text/event-stream 的长连接 HTTP 响应,向客户端推送一连串文本事件。每个事件由一行或多行字段(如 data:)组成,以一个空行结束。 - [停止序列(stop sequences)](https://broadrouter.com/zh/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): 停止序列(stop sequences)是随请求发送的字符串,模型一旦生成其中任意一个就停止输出。OpenAI 的字段叫 stop,最多接受 4 个序列;Anthropic 叫 stop_sequences。匹配到的文本通常不会出现在返回结果里。 - [流式输出(streaming)](https://broadrouter.com/zh/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): 流式输出是一种在语言模型仍在生成时,就把回复分成小块陆续返回的方式,而不是在最后一次性返回。当请求把 stream 设为 true 时,OpenAI 兼容 API 会通过服务器发送事件(SSE)进行流式输出。 - [结构化输出(structured output)](https://broadrouter.com/zh/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): 结构化输出是大模型 API 的一项功能,它把模型的回复约束为开发者指定的格式(通常是 JSON Schema),让代码能可靠地解析。它比 JSON mode 更严格:JSON mode 只保证输出是合法 JSON,不保证特定结构。 - [子智能体(subagent)](https://broadrouter.com/zh/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): 子智能体是主智能体为处理一项明确任务而启动的独立智能体。它有自己的上下文窗口、系统提示词和工具权限,完成工作后只把结果或摘要返回给主智能体,从而让主对话保持精简。 - [次级处理者(subprocessor)](https://broadrouter.com/zh/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): 次级处理者是处理者为代表控制者处理个人数据而委托的第三方。对大模型 API 而言,次级处理者可能包括 GPU 托管方、合作推理服务商和备用算力。根据 GDPR 第 28 条,它们需要获得控制者授权,并承担同样的数据保护义务。 - [系统提示词(system prompt)](https://broadrouter.com/zh/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): 系统提示词是应用在对话开始前发给语言模型的指令文本,通常是一条 system 或 developer 角色的消息。它设定模型的角色、规则、语气和输出格式,当用户消息与它冲突时,模型应当优先遵循它。 - [塔马尔布塔(taa marbuta,ة)](https://broadrouter.com/zh/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): Taa marbuta(ة,U+0629)是只出现在词尾的阿拉伯字母,常用来标记阴性名词和形容词。它看起来像带两个点的 heh(ه,U+0647),所以输入的文本常把两者混用,许多搜索规范化工具会把 ة 映射为 ه。 - [温度(temperature)](https://broadrouter.com/zh/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): 温度(temperature)是控制语言模型输出随机程度的采样参数。模型先把每个 token 的分数除以温度,再转换成概率:温度低时,最可能的 token 占据主导;温度高时,分布变得平坦,可能性较低的 token 被选中的次数更多。 - [文本转语音(TTS,text-to-speech)](https://broadrouter.com/zh/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): 文本转语音(TTS),也叫语音合成,是把书面文字转换成语音的技术。TTS API 接收文本、音色和输出格式(如 MP3 或 WAV),返回生成的语音。它与把音频转成文字的语音转文字方向相反。 - [吞吐量(throughput)](https://broadrouter.com/zh/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): 在大模型推理服务中,吞吐量(throughput)是系统在所有并发请求上单位时间内完成的总工作量,通常以每秒输出 token 数或每秒请求数衡量。它不同于单路流速度,后者是单个请求接收 token 的快慢。 - [首 token 延迟(TTFT,time to first token)](https://broadrouter.com/zh/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): 首 token 延迟(TTFT)是从向语言模型发出请求到第一个输出 token 到达的时间。它包括网络时间、服务器上的排队时间和预填充(模型读取整个提示词的步骤),所以会随提示词长度和负载增加。 - [时间戳(语音转文字,timestamps)](https://broadrouter.com/zh/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): 语音转文字中的时间戳,是附在转写片段上的开始和结束时间,单位为秒。片段级时间戳覆盖一个短语或句子,词级时间戳覆盖每个词。制作字幕、在录音中搜索以及把文字与说话人对齐都需要它。 - [Token(词元)](https://broadrouter.com/zh/glossary/token/) (translation of https://broadrouter.com/glossary/token/): token(词元)是大语言模型读写文本的基本单位:由模型的分词器切分出的一个词、词的一部分、一个标点或一段字节序列。大模型 API 按 token 计数、设定限制,通常也按 token 计费。 - [分词器(tokenizer)](https://broadrouter.com/zh/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): 分词器是把文本转换成语言模型读取的整数 token ID 序列、并把生成的 ID 还原成文本的组件。每个模型都自带分词器,所以同一段文本在不同模型上可能产生不同数量的 token。 - [每秒 token 数(TPS,tokens per second)](https://broadrouter.com/zh/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): 每秒 token 数(TPS)是语言模型生成输出 token 的速率,通常针对单个请求、在首个 token 到达之后测量。它决定了回复开始后流得有多快。把所有并发请求加起来就是系统吞吐量;首个 token 之前的等待时间则单独测量。 - [工具调用(tool calling / function calling)](https://broadrouter.com/zh/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): 工具调用(也叫函数调用,function calling)是大模型 API 的一项功能:模型返回一个结构化请求,要求执行开发者定义的某个函数,并附上 JSON 参数。应用执行该函数并把结果发回;模型自己从不执行代码。 - [tool_choice(工具选择参数)](https://broadrouter.com/zh/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice 是控制工具使用的请求参数。在 OpenAI 的 Chat Completions 中,它可以是 auto(由模型决定)、required(必须调用一个或多个工具)、none(不用工具),或指定某个函数强制调用。Anthropic 对应的取值是 auto、any、tool 和 none。 - [Top-p(核采样,nucleus sampling)](https://broadrouter.com/zh/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Top-p 又叫核采样(nucleus sampling),是一种解码方法:语言模型只在“累计概率达到阈值 p(如 0.9)的最小高概率 token 集合”中采样下一个 token,丢弃不太可能的长尾。 - [Unicode 规范化(Unicode normalization)](https://broadrouter.com/zh/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Unicode 规范化把文本转换成标准形式,让看起来相同的字符串使用相同的码位。共有 NFD、NFC、NFKD 和 NFKC 四种形式;带 K 的形式还会把兼容字符(如阿拉伯语表现形式和连字)折叠成普通字母。 - [语音活动检测(VAD)](https://broadrouter.com/zh/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): 语音活动检测(VAD)是一种信号处理步骤,把音频切成短帧并判断每帧是否为语音。语音系统用它跳过静音、在停顿处切分长录音、判断说话人是否说完,并避免转写背景噪声。 - [向量数据库(vector database)](https://broadrouter.com/zh/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): 向量数据库是存放嵌入向量的数据存储,通常连同原文和元数据一起保存,并回答最近邻查询:给定查询向量,返回与之最相似的已存向量,一般按余弦距离、欧氏距离或内积计算。 - [视觉语言模型(VLM,vision-language model)](https://broadrouter.com/zh/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): 视觉语言模型(VLM)是一种同时接收图像和文本作为输入、输出文本的模型,可以描述图片、回答关于图片的问题,并读取文档、图表或截图。VLM 也被称为图文生文(image-text-to-text)模型或多模态模型。 - [vLLM](https://broadrouter.com/zh/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM 是一个用于 LLM 推理和服务部署的开源库,最初由加州大学伯克利分校 Sky Computing Lab 开发。它用 PagedAttention 管理 KV 缓存内存,对请求做连续批处理,支持多种量化格式,并提供 OpenAI 兼容的 HTTP 服务器。 - [词表大小(vocabulary size)](https://broadrouter.com/zh/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): 词表大小是语言模型分词器中不同 token 的数量,每个 token 都有自己的 ID 和嵌入向量。更大的词表能用更少的 token 编码常见文本,但会让嵌入层和输出层变大,而且只对词表训练时见过的文本有效。 - [WAV](https://broadrouter.com/zh/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV(Waveform Audio File Format,又称 WAVE)是建立在微软与 IBM 的 RIFF 格式之上的音频容器。它通常存放未压缩的线性 PCM 样本,因此大小可以预测:每秒字节数 = 采样率 × 每样本字节数 × 声道数。 - [词错误率(WER,word error rate)](https://broadrouter.com/zh/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): 词错误率(WER)是语音识别的标准准确率指标:把转写结果改成正确参考文本所需的替换、删除和插入词数,除以参考文本的词数。数值越低越好。 - [零数据保留(ZDR)](https://broadrouter.com/zh/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): 零数据保留(ZDR)是一种数据处理政策:AI 服务商只为完成请求而处理提示词和输出,之后不在日志、缓存或训练数据中存储这些内容。不含内容的记录,如 token 数和耗时,通常仍会保留。 ### Turkish (tr, Türkçe): 294 translated pages - [Articles in Turkish](https://broadrouter.com/tr/blog/) - [Ekim 2026 LLM API fiyat değişiklikleri: 3 zam, hiç indirim yok](https://broadrouter.com/tr/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): Ekim 2026'da 7 LLM API sağlayıcısının 5'inde fiyat sayfalarında 23 değişiklik: 3 zam, hiç indirim yok, 9 yeni fiyatlı model, 11 kaldırılan. Hepsi arşiv linkiyle. - [Bir yapay zeka kodlama ajanı çalıştırması ne kadar tutar?](https://broadrouter.com/tr/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): Ajan çalıştırması uzadıkça adım başına maliyet neden artar, önbellek toplamı nasıl değiştirir, sabit oturum ne zaman token başına ödemeyi geçer? Hesaplı örnek. - [LLM ile Arapça metne hareke ekleme: istem, kontrol, DER](https://broadrouter.com/tr/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): Arapça harekeleri (teşkil) bir LLM API ile geri getirin: her harfi koruyan istem, düzenlemeyi reddeden kontrol, DER ve WER puanlaması, CATT ne zaman seçilmeli. - [LLM ile Arapça dil bilgisi ve yazım düzeltme: adım adım](https://broadrouter.com/tr/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Arapça yazım ve dil bilgisi hatalarını bir LLM API ile düzeltin: en az düzenleme istemi, yeniden yazıma karşı koruma, hata sınıfına göre düzenlemeler ve M2 puanı. - [Kayıtlardan Arapça toplantı tutanağı: bir API hattı](https://broadrouter.com/tr/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Arapça toplantı kayıtlarını kararlar ve aksiyon maddeleriyle tutanağa çevirin: metne dökün, temizleyin, LLM ile JSON özetleyin, adları ve tarihleri doğrulayın. - [LLM ile Arapça varlık çıkarımı: JSON, tarihler ve denetimler](https://broadrouter.com/tr/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Arapça kişi, kurum, yer adlarını ve Hicri tarihleri LLM ile JSON olarak çıkarın: istem, harekeye duyarsız aralık denetimi, tarih dönüştürme ve puanlama betiği. - [Arapça OCR API karşılaştırması: Google, Azure, Textract, Mistral, LLM](https://broadrouter.com/tr/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): Hangi OCR API'leri basılı ve el yazısı Arapçayı resmî olarak destekliyor, her biri ne döndürüyor, 1.000 sayfa ne tutuyor; görsel LLM API'leriyle karşılaştırma. - [Konuşmadan metne API ile Arapça altyazı (SRT): neye ihtiyacınız var](https://broadrouter.com/tr/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): SRT dosyası neye ihtiyaç duyar, hangi API'ler Arapça için zaman damgası veya SRT döndürür ve yalnızca metin veren bir API'den altyazı üreten test edilmiş bir hat. - [LLM ile Arapça metin sınıflandırma: duygu, niyet, önceliklendirme](https://broadrouter.com/tr/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): Arapça yorumları, mesajları ve destek taleplerini bir LLM API ile sınıflandırın: etiket tasarımı, lehçeye duyarlı istem, JSON çıktı, makro F1 testi, maliyet. - [Arama ve RAG için Arapça metin normalleştirme: neyi değiştirmeli?](https://broadrouter.com/tr/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Teşkil, elif biçimleri, tâ merbûta, hemze oturakları, Farsça harfler, rakamlar ve Unicode: hangi Arapça normalleştirme kuralı aramaya yarar, hangisi zarar verir. - [LLM ile Arapça metin özetleme: istem, uzunluk ve puanlama](https://broadrouter.com/tr/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): Arapça makale, rapor ve lehçe mesajlarını LLM API ile özetleyin: istem kalıpları, kelime sayısı kontrolü, sayı denetimi ve varsayılan ROUGE neden 0 verir. - [Artificial Analysis alternatifleri: 2026'da LLM karşılaştırma](https://broadrouter.com/tr/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Artificial Analysis alternatifleri: Arena, Epoch AI, Vals AI, OpenRouter sıralamaları, Hugging Face ve fiyat verileri; her biri neyi ölçüyor, ne zaman kullanılmalı? - [BroadNet'in GPU donanımı ve sunum yaklaşımı, BroadNet'in anlatımıyla](https://broadrouter.com/tr/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): BroadNet, BroadNet Inference'ın donanımı hakkında ne diyor: B300'den müşteri cihazlarına üç GPU katmanı, GPU başına işlem hacmi ve testlerden çıkan dersler. - [LLM API yanıtlarını önbelleğe alıp maliyeti düşürün: Python örneği](https://broadrouter.com/tr/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Python'da LLM API yanıtları için test edilmiş birebir eşleşmeli önbellek: anahtara ne girer, neler asla önbelleğe alınmaz, TTL, gizlilik ve isabet oranını ölçme. - [Claude Code ile açık modeller: proxy, yerel uç noktalar ve sınırlar](https://broadrouter.com/tr/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code, Anthropic Messages API'yi konuşur. Açık modellerde LiteLLM proxy'si veya yerel bir uç noktayla nasıl çalıştırılır ve yolda neler bozulur? - [Belge yapay zekâsı iş akışınıza insan incelemesini ekleyin](https://broadrouter.com/tr/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): Belge yapay zekâsı hattında insan incelemesini tasarlayın: doğrulama kuralları, kişiye ne gönderilir, rastgele denetim örneği, denetim izi ve inceleme oranı. - [Mısır veri koruma yasası (151 sayılı Kanun, 2020) ve LLM API'leri](https://broadrouter.com/tr/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Mısırlıların kişisel verilerini yurt dışındaki LLM veya konuşmadan metne API'ye göndermek: 151 sayılı Kanun ve 816/2025 Kararı aktarım, lisans, ihlal için ne der? - [İnce ayardan önce: çözülecek sorunu tanımlayın](https://broadrouter.com/tr/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): İstem, RAG veya değerlendirme sorunu ne zaman çözer ve ince ayar gereksizdir; hangi eğitim verisi gerekir ve bir ince ayarı istemeden önce nasıl test edersiniz. - [LLM'den gelen geçersiz JSON: nedenler, onarım ve yeniden deneme](https://broadrouter.com/tr/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): LLM'ler neden bozuk JSON döndürür (kod çiti, ek metin, max_tokens kesintisi, sondaki virgül) ve güvenli olanı onarıp gerisini reddeden test edilmiş Python kodu. - [LLM API fiyatları ne sıklıkla değişir? 3 yıllık arşiv sayfaları](https://broadrouter.com/tr/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): 11 sağlayıcının 2023–2026 arası 308 arşiv fiyat sayfasını okuduk. Modellerin %80'inin fiyatı hiç değişmedi; liste fiyatı değişikliklerinin %75'i DeepInfra'dan. - [LLM API maliyetleri nasıl düşürülür: sıralı 10 adımlık kontrol listesi](https://broadrouter.com/tr/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): LLM API faturasını doğru sırayla küçültün: ölçün, istem önekini önbelleğe alın, istemi kısaltın, çıktıyı ve akıl yürütmeyi sınırlayın, modeli doğru seçin. - [LLM API failover: kendi kodunuzda ikinci bir sağlayıcıya geçiş](https://broadrouter.com/tr/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): OpenAI uyumlu LLM sağlayıcıları arasında failover: hangi hatalarda geçilir, devre kesici, model kimliği eşleme, akış, maliyet ve gizlilik koşullarının denkliği. - [2023–2026 LLM API fiyat indirimleri: en büyük düşüşler ve kaynakları](https://broadrouter.com/tr/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): 2023–2026 arasındaki en büyük LLM API fiyat indirimleri, her biri önce ve sonra Internet Archive bağlantılarıyla: o3, Codestral, Gemini 1.5 ve gözden kaçan zamlar. - [LLM API fiyatları hep düşer mi? 52 zam ve bütçeye nasıl katılır](https://broadrouter.com/tr/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): LLM API fiyatları da artıyor. Arşivlenmiş fiyat sayfalarında 2024–2026 arası aynı modelde 52 zam bulduk, medyan %40. Tüm kaynaklar ve bir bütçe kuralı. - [Sağlayıcılara göre LLM API hız sınırları: RPM, TPM ve eşzamanlılık](https://broadrouter.com/tr/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): 13 sağlayıcının giriş katmanı LLM API hız sınırları, kendi belgelerinden: OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks ve diğerleri; RPM ve TPM ne demek? - [LLM API SLA'ları: Azure, Bedrock, Google ve OpenAI ne vaat ediyor?](https://broadrouter.com/tr/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): Her LLM API sağlayıcısının yazılı taahhüdü: çalışma süresi yüzdesi, neyin kesinti sayıldığı, hizmet kredileri ve gecikme hedefleri, SLA sayfalarından okundu. - [Batch API olmadan LLM toplu işleme: kaldığı yerden devam eden betik](https://broadrouter.com/tr/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Batch API olmadan 10.000 ila 1.000.000 istemi chat completions API'sinden geçirin: eş zamanlılık sınırı, hız ayarı, yeniden deneme, kontrol noktası ve süre hesabı. - [Safety Map: 39 LLM hassas verileri nasıl ele alıyor?](https://broadrouter.com/tr/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet 14 laboratuvardan 39 LLM'i kimlik, kişisel, sağlık ve finans verisiyle test etti: ne aktarıldı, ne aktarılmadı; araç çağrısı ve akıl yürütme sızıntıları. - [LLM hassas veri güvenliği nasıl test edilir: BroadNet'in yöntemi](https://broadrouter.com/tr/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): BroadNet'in, bir LLM'in kimlik, kişisel, sağlık veya finans verisi sızdırıp sızdırmadığını yayından önce test yöntemi: kategoriler, sızıntı/kaçırma, 3+ çalıştırma. - [OpenAI API'den geçiş: Python, Node ve cURL ile adım adım kod](https://broadrouter.com/tr/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): OpenAI'dan OpenAI uyumlu bir API'ye geçmek için kod değişiklikleri: istemci kurulumu, Responses'tan Chat Completions'a, araçlar, akış, embedding ve duman testi. - [BroadNet'in test edip elediği modeller ve nedenleri](https://broadrouter.com/tr/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet reddettiği modelleri gerekçeleriyle yayımlıyor: kimlik bilgisi güvenliği testini geçemeyen sohbet modelleri ve kaybeden konuşma, OCR, embedding modelleri. - [Çok dilli yapay zekâ ürünü tasarımı: Arapça, İngilizce, İspanyolca](https://broadrouter.com/tr/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): Arapça, İngilizce ve İspanyolca kullanıcılar için yapay zekâ ürünü: arayüz dili ile model dili, RTL ve karışık metin, sayılar, tarihler, dil tespiti ve maliyet. - [Açık ağırlıklı ve kapalı model API maliyeti: fiyat verileri ne diyor?](https://broadrouter.com/tr/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): Açık ağırlıklı modeller API üzerinden GPT, Claude veya Gemini'den ucuz mu? Sağlayıcı sayfalarından token fiyatları, sağlayıcı farkları ve kullanıcı başına maliyet. - [OpenAI, Anthropic ve Google API fiyat geçmişi, 2023–2026](https://broadrouter.com/tr/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): Arşivlenmiş fiyat sayfalarından 2023–2026 GPT, Claude ve Gemini API fiyatları yan yana: hangi seriler ucuzladı, hangileri pahalandı, her biri arşiv bağlantılı. - [OpenAI uyumlu API base URL listesi: 14 sağlayıcı kontrol edildi](https://broadrouter.com/tr/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): OpenAI uyumlu 14 LLM API'sinin base_url adresi, her biri sağlayıcının kendi belgelerinden; listelenen uç noktalar ve sorunsuz geçişi bozan farklar. - [OpenClaw ve Hermes Agent ucuz modelle: kurulum ve token hesabı](https://broadrouter.com/tr/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): OpenClaw veya Hermes Agent'ı OpenAI uyumlu bir uç noktaya bağlayın, zamanlanmış kontrollerin token yakmasını anlayın, faturayı hafif bağlam ve ucuz modelle düşürün. - [Özel uç nokta, net tanımlanmış bir iş yüküyle başlar](https://broadrouter.com/tr/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): Özel bir LLM uç noktası istemeden önce tanımlanacaklar: iş yükü, saniyede token, eşzamanlılık, bağlam, bölge, lisans, bütçe; ardından uç noktanın testi. - [Kodlama ajanlarında istem enjeksiyonu: gerçekten işe yarayan önlemler](https://broadrouter.com/tr/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): İstem enjeksiyonu kodlama ajanına nereden ulaşır; sağlayıcı ve araştırmacıların belgelediği katmanlı savunmalar: sanal alan, çıkış sınırı, izin listesi, onay. - [RAG parça boyutu ve örtüşme: getirmeye ve maliyete etkisi](https://broadrouter.com/tr/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): Parça boyutu ve örtüşme; RAG isabetini, indeks boyutunu, embedding maliyetini ve her sorgunun LLM'e gönderdiği token'ları nasıl değiştirir? Testler ve hesaplar. - [Akıl yürütme modellerinde max_tokens ayarı: bütçeleme yöntemi](https://broadrouter.com/tr/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Akıl yürütme modeli düşük max_tokens ile neden boş yanıt döner; bütçeyi ölçülen yüzdeliklerle, bağlam penceresiyle ve büyüyen yeniden denemeyle nasıl belirlersiniz? - [Tarayıcıda ses kaydedip ses metne çevirme API'sine gönderme](https://broadrouter.com/tr/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): Mikrofonu MediaRecorder ile kaydedin, WebM veya Safari MP4'ü 16 kHz WAV'a çevirin, anahtarı tutan küçük bir sunucuyla transkripsiyon API'sine gönderin. - [İstem token sayısını azaltma: istem, araç ve veride ölçülen kazanç](https://broadrouter.com/tr/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): Sistem istemi, few-shot örnekler, araç şemaları, veri formatı, sohbet geçmişi ve Arapça harekeleri kırpmanın gerçek tokenizer'larla ölçülmüş token tasarrufu. - [Aynı model, farklı sağlayıcı: LLM API fiyatları ne kadar değişiyor?](https://broadrouter.com/tr/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): Birden fazla sağlayıcının sattığı 119 modelde en pahalı liste fiyatı medyan olarak en ucuzun 1,76 katı. 41 sağlayıcının kendi fiyat sayfalarından veri, Kasım 2026. - [Suudi ve BAE PDPL ile LLM API'leri: yurt dışı aktarım listesi](https://broadrouter.com/tr/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Suudi Arabistan veya BAE'den bir LLM API'ye kişisel veri göndermek: PDPL aktarım kuralları ne diyor, sağlayıcıya ne sorulur, yerel çıkarım ne zaman şart? - [Metin, ses mi görsel mi? İşe görevden başlayın](https://broadrouter.com/tr/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): Metin sohbeti, görsel girdi veya ses transkripsiyonunu göreve göre seçin: BroadNet'in belgelediği modeller ve limitler, ayrıca sunmadıkları (ör. metinden sese). - [25 MB üstü uzun sesi metne çevirme: bölün, gönderin, birleştirin](https://broadrouter.com/tr/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): Konuşmadan metne API dosya sınırını aşan kayıtları metne çevirme: dosyayı küçültün, ffmpeg ile sessizliklerden bölün, parçaları paralel gönderip metni birleştirin. - [Uygulama metinlerini LLM ile Arapçaya çevirme: yer tutucular, çoğullar](https://broadrouter.com/tr/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): i18n JSON dosyalarını LLM API ile Arapçaya çevirin: yer tutucuları dondurun, altı Arapça çoğul biçimini üretin, ICU sözdizimini CI'da doğrulayın, adları yalıtın. - [Arapça sesi API ile İngilizce metne çevirme: test edilmiş bir hat](https://broadrouter.com/tr/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): Arapça kayıtları iki API çağrısıyla İngilizce metne çevirin: language=ar ile yazıya dökün, LLM ile çevirin; sayı denetimi, bölme, maliyet ve puanlama dahil. - [Canlı, talep üzerine, araştırma: model kataloğunu doğru okuyun](https://broadrouter.com/tr/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): Katalogdaki hangi modeli bugün çağırabileceğinizi, hangisinin talep gerektirdiğini, hangisinin yalnızca test edildiğini BroadNet kataloğu ve /v1/models ile görün. - [vLLM'de Uno: BroadNet'in sunum çalışması neyi ölçtü](https://broadrouter.com/tr/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet, Uno öz-taslak adaptörünü vLLM'e yerleştirip bir H100'de ölçtü: tek akışta düz işlem hacminin 2,58 katı, 32 akışta 1,74 katı. Yöntem, DOI ve sınırlar. - [Gemma 4 26B üzerinde Uno: BroadNet'in uzmanlar karışımı çalışması](https://broadrouter.com/tr/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet ilk Uno öz-taslak adaptörünü bir uzmanlar karışımı modeli olan Gemma 4 26B A4B için eğitti. Neyi ölçtü, öz-taslak ne zaman kazandırır, DOI. - [Ajan token bütçesi: bir ajan döngüsünün token'ları nereye gidiyor?](https://broadrouter.com/tr/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): Sistem istemi, araç şemaları, dosya okumaları, geçmiş ve akıl yürütme: ajan döngüsü token'ları nasıl harcar, her parça nasıl ölçülür, önce ne kırpılır? - [Arapça ASR için kelime hata oranı (WER) nedir?](https://broadrouter.com/tr/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): Kelime hata oranı ne demek, nasıl hesaplanır ve Arapça WER neden harekelere, yazım normalleştirmesine ve lehçelere bu kadar duyarlı? Python ve gerçek sayılarla. - [Arapça lehçelerde konuşma tanıma: Körfez, Mısır, Levant, Mağrip](https://broadrouter.com/tr/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Konuşmadan metne modelleri Körfez, Mısır, Levant ve Mağrip Arapçasında neden zorlanır, kendi sesinizle bir lehçe test seti nasıl kurulur? - [LLM ile Arapça–İngilizce çeviri: istemler, terim sözlükleri, test](https://broadrouter.com/tr/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): Bir LLM API ile Arapçadan İngilizceye ve tersine çevirin: tonu koruyan istemler, sabit terimler için sözlük, sayılar ve adlar, kaliteyi değerlendirme yolları. - [Körfez, Mısır ve Levant Arapçası: LLM'ler lehçelerle nasıl başa çıkar?](https://broadrouter.com/tr/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): LLM'ler Körfez, Mısır ve Levant Arapçasında nerede hata yapıyor, lehçe için nasıl istem yazılır ve test edilir, BroadNet Qwen3.8 27B için ne bildiriyor. - [Arapça OCR mı, görü LLM'i mi: Arapça belge ve fatura okuma](https://broadrouter.com/tr/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): Arapça fatura, fiş ve formlarda klasik OCR'ı ne zaman, görü LLM'ini ne zaman kullanmalı: doğruluk tuzakları, JSON çıktı, sayfa başı maliyet ve doğrulama hattı. - [Arapça istem mühendisliği: dil düzeyi, RTL, rakamlar, karışık metin](https://broadrouter.com/tr/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): LLM'den iyi Arapça alan istemler: dil düzeyini ve lehçeyi seçin, rakamları ve tarihleri kontrol edin, RTL biçimini temiz tutun, Arapça–İngilizce karışımı yönetin. - [Arapça RAG: işe yarayan normalleştirme, parçalama ve erişim](https://broadrouter.com/tr/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): Arapça belgelerde erişimle zenginleştirilmiş üretim: elif, yâ ve ta marbuta normalleştirme, akıllı parçalama, çok dilli embedding ve anahtar kelime araması. - [Arapça ses metne çevirme API: 2026'da nasıl seçilir?](https://broadrouter.com/tr/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): Arapça transkripsiyon API'si nasıl seçilir: kelime hata oranı, lehçeler, dosya sınırları, ses saati başına fiyat, gizlilik ve kendi kayıtlarınız için test planı. - [Arapça WhatsApp sohbet botu mimarisi: sesli mesajdan yanıta](https://broadrouter.com/tr/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): Arapça WhatsApp asistanı için çalışan bir mimari: webhook, kuyruk, sesli mesaj transkripsiyonu, LLM yanıtı, politika kuralları ve maliyet; uyarlanabilir Python ile. - [Arabizi ve LLM'ler: 3 ve 7 gibi rakamlarla Franko-Arapça](https://broadrouter.com/tr/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): Arabizi, Arapçayı Latin harfleri ve ع için 3, ح için 7 gibi rakamlarla yazar. Nasıl tespit edilir, LLM ile Arap yazısına nasıl çevrilir ve doğru nasıl yanıt verilir? - [Aider için en iyi LLM: model nasıl seçilir ve bağlanır?](https://broadrouter.com/tr/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Aider için en iyi LLM nasıl seçilir: düzenleme biçimi disiplini, bağlam boyutu, saatlik maliyet, kendi deponuzda bir saatlik test ve tam OPENAI_API_BASE kurulumu. - [Kodlama ajanları için en iyi açık model: araçlar, bağlam ve maliyet](https://broadrouter.com/tr/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): Kodlama ajanınız için GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B mi Gemma 4 mü? Araç çağırma, bağlam boyutu, akıl yürütme ve görev başına maliyete göre, hesabıyla. - [Önbellekli girdi tokenları: LLM API faturası nasıl düşürülür](https://broadrouter.com/tr/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): Önbellekli girdi, normal girdinin küçük bir kesrine mal olur. İstem önbellekleme nasıl faturalanır, usage içinde cached_tokens nasıl okunur, istem nasıl kurulur? - [En ucuz görsel LLM API: görsel ve sayfa başına maliyet hesabı](https://broadrouter.com/tr/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): Görsel LLM çağrısının görsel ve sayfa başına maliyeti: Gemma 4, Qwen3.8 27B ve DeepSeek V4.1 Flash için token hesabı ve DeepInfra'nın nerede daha ucuz olduğu. - [LLM API sağlayıcısı nasıl seçilir: 20 maddelik kontrol listesi](https://broadrouter.com/tr/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): LLM API seçmeden önce sorulacak 20 soru: fiyat, limitler, OpenAI uyumluluğu, gizlilik, modeller, destek ve erişilebilirlik; ayrıca puanlama tablosu ve test planı. - [Arapça için LLM API seçimi: neler test edilmeli?](https://broadrouter.com/tr/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): Arapça uygulamalar için LLM seçimi: lehçe anlama, yazım kalitesi, çeviri, Arapça belgeler ve görseller, maliyet ve bir günde kurabileceğiniz bir test seti. - [Cline vs Roo Code vs Aider: açık modelli ekipler için (2026)](https://broadrouter.com/tr/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (artık arşivlendi, Zoo Code olarak sürüyor) ve Aider; açık modellerde önemli olan konularda karşılaştırma: araç çağrıları, bağlam, MCP, git, onaylar. - [LLM isteği göndermeden önce token sayısı nasıl hesaplanır?](https://broadrouter.com/tr/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): LLM API'sini çağırmadan önce istem token'larını dile ve modele göre tahmin edin: tokenizer farkları, beş dilde ölçümler ve güvenli max_tokens payları. - [Kodlama ajanı LLM maliyetini düşürmenin 10 yolu ve tahmini tasarruf](https://broadrouter.com/tr/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Cline, Aider, opencode veya kendi kodlama ajanınızın token harcamasını düşürmenin on somut yolu; her birinin tasarruf hesabı ve nasıl ölçüleceği. - [DeepSeek API alternatifleri: DeepSeek modellerini çağırmanın 5 yolu](https://broadrouter.com/tr/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): DeepSeek modellerini resmi API, bir aracı platform, bulut kataloğu, kendi GPU'larınız veya BroadNet Inference ile çağırmak: fiyat, gizlilik ve sınır farkları. - [DeepSeek V4.1 Flash API fiyatı: token başına, önbellekli ve planlar](https://broadrouter.com/tr/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): DeepSeek V4.1 Flash API fiyatı: 1M girdi, önbellekli girdi ve çıktı token ücretleri, aylık, haftalık, saatlik sabit planlar ve DeepSeek, DeepInfra, Fireworks kıyası. - [Bir özelliği yayına almadan önce LLM API maliyeti nasıl hesaplanır](https://broadrouter.com/tr/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): Yeni bir yapay zeka özelliği için adım adım maliyet modeli: kullanıcı × eylem × çağrı × token; önbellek, akıl yürütme, yeniden deneme, hız sınırı ve örnek tablo. - [Sağlayıcı değiştirmeden önce bir LLM nasıl değerlendirilir?](https://broadrouter.com/tr/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): Gerçek trafikten küçük bir eval seti kurun, kod kontrolleri ve yanlılığı giderilmiş LLM hakemle puanlayın; geçişten önce kalite, maliyet ve gecikmeyi karşılaştırın. - [Fatura, fiş ve formlardan görsel LLM ile JSON veri çıkarma](https://broadrouter.com/tr/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): Fatura, fiş ve formlar için üretim hattı: görüntü hazırlığı, JSON şemalı istem, Pydantic doğrulaması, aritmetik kontroller, yeniden deneme ve insan inceleme kuyruğu. - [Sabit fiyatlı mı, token başına mı LLM API: başa baş hesabı](https://broadrouter.com/tr/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): Sabit fiyatlı bir LLM planı token başına ödemeyi ne zaman geçer? DeepSeek V4.1 Flash için dört örnek aylık fatura, başa baş hacmi ve kendi kullanımınızı ölçme yolu. - [Ücretsiz LLM API kredisi: başlangıç kredisi gerçekte ne kadar yeter?](https://broadrouter.com/tr/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): Ücretsiz LLM API kredisiyle gerçekte ne alırsınız: BroadNet başlangıç kredisinde token, ses dakikası ve saatlik anahtar; Gemini ve OpenRouter ile kıyas. - [GDPR ve LLM API'leri: sağlayıcınıza sormanız gereken 15 soru](https://broadrouter.com/tr/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): LLM API'leri için GDPR kontrol listesi: DPA, alt işleyenler, saklama, model eğitimi, yurt dışı aktarım ve ihlal bildirimi; her sorunun dayandığı GDPR maddesiyle. - [GLM 5.3 mü DeepSeek V4.1 Flash mı: hangisi, ne zaman kullanılmalı?](https://broadrouter.com/tr/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 ile DeepSeek V4.1 Flash; fiyat, bağlam, araç çağırma, yapılandırılmış çıktı, görsel ve akıl yürütme açısından kıyaslandı. Örnek maliyetler ve seçim kuralı. - [LLM API 429 hatası nasıl çözülür: geri çekilme, jitter, sınırlayıcılar](https://broadrouter.com/tr/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): LLM API'lerinde 429 ve 5xx farkı, tam jitter'lı üstel geri çekilme, eşzamanlılık sınırlayıcıları ve kuyruklar; üretime kopyalayabileceğiniz Python ve Node koduyla. - [Saatlik LLM API anahtarı: saat başına sabit fiyat, açıklamalı](https://broadrouter.com/tr/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): BroadNet'in saatlik anahtarı, sabit bir fiyata tek akışta bir saatlik Gemma 4 VL 26B sunar. Neleri kapsar, limitleri ne, kime uygun ve ne zaman kârlı olur? - [LLM API veri saklama: OpenAI, Anthropic, Gemini, Mistral kıyaslaması](https://broadrouter.com/tr/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): OpenAI, Anthropic, Gemini API, Mistral ve BroadNet, API istemlerini saklama ve onlarla eğitim konusunda resmi olarak ne diyor? Kasım 2026 sayfalarından alıntılarla. - [LLM API anahtarı güvenliği: saklama, rotasyon ve IP izin listeleri](https://broadrouter.com/tr/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): LLM API anahtarlarını koddan ve loglardan uzak tutun: iş yükü başına anahtar, düzenli rotasyon, sunucu IP'sine kilit ve sızıntıyı saniyeler içinde iptal. - [LLM API fiyatları: girdi, önbellekli girdi, çıktı ve akıl yürütme](https://broadrouter.com/tr/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): LLM API'lerde token başına fiyatlandırma nasıl işler: girdi, önbellekli girdi, çıktı, akıl yürütme tokenları, görseller ve dakika başına ses, örnek hesaplarla. - [LLM API'ye hassas veri göndermek: maskeleyin, azaltın, az kaydedin](https://broadrouter.com/tr/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): Müşteri, sağlık veya finans verisini LLM API'ye güvenle göndermek: alanları sınıflandırın, geri alınabilir yer tutucularla maskeleyin, sıfır veri saklama seçin. - [LLM API harcama kontrolü: bütçe, anahtar sınırı, döngü koruması](https://broadrouter.com/tr/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): LLM API faturasının kontrolden çıkmasını önleyin: ön ödemeli bakiye, uygulama başına bir anahtar, IP izin listesi, kullanım kaydı, kod içi bütçe, döngü sınırları. - [LLM API kullanım günlüğü: kullanıcı ve özellik başına maliyet, PII yok](https://broadrouter.com/tr/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): Her LLM API çağrısını token, maliyet, gecikme ve istek kimliğiyle kaydedin, harcamayı kullanıcıya ve özelliğe atayın, PII'yi yazmadan önce maskeleyin. Python kodu. - [Kullanıcı başına LLM maliyeti: SaaS yapay zeka özelliği ekonomisi](https://broadrouter.com/tr/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): Yapay zeka özelliğinin aktif kullanıcı başına aylık maliyeti, yoğun kullanıcıların ortalamadan neden önemli olduğu ve faturayı karşılayan SaaS paket fiyatı. - [Açık model API fiyatları: 6 çıkarım sağlayıcısı nasıl fiyatlandırıyor?](https://broadrouter.com/tr/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): OpenRouter, Together AI, Fireworks, Groq, DeepInfra ve BroadNet açık modelleri nasıl fiyatlandırıyor: token başına, önbellekli, toplu iş, GPU saati ve sabit plan. - [LLM gecikmesi: TTFT, saniye başına token ve uçtan uca süre](https://broadrouter.com/tr/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): İlk token süresi, saniye başına token ve uçtan uca gecikme neyi ölçer, her biri istemciden kodla nasıl ölçülür ve hangi iyileştirme hangi sayıyı değiştirir. - [LLM akışı (streaming): SSE parçaları, delta, usage ve iptal](https://broadrouter.com/tr/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Chat completion akışı nasıl iletilir, yanıt delta'lardan nasıl kurulur, usage son parçadan nasıl okunur ve Python, Node, curl'de akış nasıl temizce iptal edilir? - [Uzun bağlam mı RAG mı: 1M token vektör veritabanının yerini alır mı?](https://broadrouter.com/tr/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): 1M token'lık bir bağlam penceresinin vektör veritabanının yerini ne zaman alabileceği, RAG'in ne zaman hâlâ kazandığı: sorgu başına maliyet, gecikme, isabet, bakım. - [Uzun süren ajanlarda 1M token pencereyle bağlam yönetimi](https://broadrouter.com/tr/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): 1M token pencereyle uzun ajan oturumlarını isabetli ve uygun maliyetli tutun: yalnızca eklenen geçmiş, önbellek dostu istem, sıkıştırma zamanı ve ilk ne atılır. - [Uzun süren ajanlar: zaman aşımı, yeniden deneme ve kontrol noktası](https://broadrouter.com/tr/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): Bir LLM ajanını saatlerce ayakta tutun: çağrı başına zaman aşımı, takılmaları gösteren akış, güvenli yeniden deneme, idempotent araçlar ve adım adım kayıt. - [MCP sunucularını OpenAI uyumlu modellerle kullanma (her sağlayıcıda)](https://broadrouter.com/tr/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP her LLM ile çalışır: istemci MCP araçlarını fonksiyon çağrılarına çevirir. Açık modellerle MCP'yi Cline, Continue, opencode, n8n veya Python'da kullanın. - [1M token'lık bağlam penceresi ne zaman değer (ve ne zaman değmez)?](https://broadrouter.com/tr/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): 1.048.576 token'a neler sığar, DeepSeek V4.1 Flash'ta maliyeti nedir, akıl yürütme ve önbellekleme hesabı nasıl değiştirir ve RAG ne zaman hâlâ daha iyi seçimdir? - [Çoklu ajan maliyeti: token başına mı, sabit oturum mu?](https://broadrouter.com/tr/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): Çoklu ajan sistemleri tek bir ajanın birkaç katı token yakar. Sabit plan ve token başına ödeme için oturum hesabı, gizli çarpanlar ve hangisinin ne zaman kazandığı. - [Çok dilli ses metne çevirme API: 14 dil, tek uç nokta](https://broadrouter.com/tr/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): İngilizce, İspanyolca, Fransızca, Çince, Arapça ve 9 dili dakika başına fiyatlı tek bir OpenAI uyumlu uç noktayla metne çevirin; Arapça modeli ne zaman seçmeli? - [Ucuz LLM ile n8n AI Agent: kurulum, model seçimi ve maliyet](https://broadrouter.com/tr/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): Düşük maliyetli OpenAI uyumlu bir modelle n8n AI Agent iş akışları kurun: araçlar için hangi model, token azaltan ayarlar ve 10.000 çalıştırma başına maliyet hesabı. - [İşletmeler için açık ağırlıklı ve kapalı yapay zekâ modelleri](https://broadrouter.com/tr/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): Açık ağırlıklı mı kapalı LLM mi? Maliyet, kontrol, gizlilik, kalite farkı ve bağımlılığı karşılaştırın; açık modelleri çalıştırmanın üç yolu ve hangisi ne zaman. - [OpenAI uyumlu API'ye geçiş: adım adım kontrol listesi](https://broadrouter.com/tr/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): base_url'i değiştirmek işin kolay kısmı. Bozulanlar model kimlikleri, akıl yürütme token'ları, akışta kullanım, hatalar, hız sınırları ve anahtarlar. Kodlu liste. - [Yapay zeka ajanları için OpenRouter mı, sabit fiyatlı plan mı?](https://broadrouter.com/tr/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter token başına ücret ve kredi alımında komisyon alır; sabit plan oturum başına. Sürekli çalışan ajanlarda maliyet, ücretler, sınırlar ve doğru seçim. - [Kodlama ajanlarını gece boyunca gözetimsiz ve güvenle çalıştırma](https://broadrouter.com/tr/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): Gözetimsiz kodlama ajanı çalışmaları için pratik bir kurulum: sanal alan, git kontrol noktaları, görev dosyası, adım sınırı, sabah incelemesi, gece için sabit plan. - [Önek önbelleğe alınsın diye istem nasıl yapılandırılır?](https://broadrouter.com/tr/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): Önbelleklenen önek aynı kalsın diye LLM istemlerinin düzeni: sistem istemi, araç şemaları, belgeler, yalnızca sona eklenen geçmiş ve bunu kanıtlayan bir test. - [Akıl yürütme token'ları: neden gördüğünüzden fazla çıktı ödersiniz](https://broadrouter.com/tr/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): Akıl yürütme modelleri gizli düşünmeyi çıktı olarak faturalandırır. Nasıl ölçülür, modele göre maliyeti ve reasoning_effort, max_tokens ve yönlendirmeyle düşürme. - [Sabit fiyatlı LLM API ile paralel kodlama ajanları nasıl çalıştırılır?](https://broadrouter.com/tr/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Sınırsız DeepSeek V4.1 Flash'ta alt ajanlar, orkestratörler ve toplu işçiler: kaç plan oturumu gerekir, anahtarlara nasıl bölünür, çakışma nasıl sınırlanır. - [LLM barındırmak mı, API mi: gerçek maliyet ve başa baş noktası](https://broadrouter.com/tr/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): Resmî sayfalardan GPU kiralama fiyatları, kendi GPU'nuzda 1M token maliyeti formülü, bir API'yi geçmek için gereken kullanım oranı ve unutulan operasyon maliyetleri. - [Konuşmadan metne API fiyatları: 2026 dakika başına karşılaştırma](https://broadrouter.com/tr/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq ve BroadNet Transcribe için resmi sayfalardan dakika ve 1.000 saat başına fiyatlar, ek ücretlerle. - [JSON modu ve json_schema: sağlam yapılandırılmış çıktı](https://broadrouter.com/tr/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON modu geçerli JSON verir, json_schema ise sizin yapınızı. OpenAI uyumlu modellerde ikisi nasıl istenir, Pydantic ile doğrulanır ve hatalı çıktı döngüde onarılır? - [Arapça ve İngilizcede kelime başına token: Arapça neden pahalı?](https://broadrouter.com/tr/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): Arapça aynı anlam için neden çoğu zaman İngilizceden fazla LLM token'ı kullanır, kelime başına token neden yanıltır; usage alanından oranınızı ölçen bir betik. - [Açık modellerle araç çağırma: şema, paralel çağrı, yeniden deneme](https://broadrouter.com/tr/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): OpenAI uyumlu açık modellerde function calling: şema ipuçları, paralel araç çağrıları, beş hata biçimi ve doğrulayıp yeniden deneyen bir Python döngüsü. - [Çağrı merkezi kayıtlarını metne çevirme: Arapçaya hazır bir hat](https://broadrouter.com/tr/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): Stereo çağrıları kanala ayırın, konuşma turlarını kesin, Arapça ve İngilizceyi metne çevirin, kişisel verileri maskeleyip LLM ile özetleyin. Kod ve maliyet. - [Kodlama ajanlarında açık modeller: Aider, Cline, opencode, Codex](https://broadrouter.com/tr/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Aider, Cline, Roo Code, opencode, Continue ve Codex CLI'ı OpenAI uyumlu herhangi bir uç noktaya bağlayın. Her biri için alan adları ve sorun çıkaran ayarlar. - [Sesli mesajı metne çevirme: WhatsApp ve Telegram'dan destek talebine](https://broadrouter.com/tr/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): WhatsApp Business ve Telegram botlarındaki müşteri sesli mesajlarını destek talebine çevirin: indirin, metne dökün, LLM ile alanları çıkarın, yönlendirin. - [Kısa projeler için haftalık LLM planı: Orbit, saatlik ve aylık](https://broadrouter.com/tr/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): Bir haftalık sınırsız DeepSeek V4.1 Flash planı ne zaman saatlik oturumdan, aylık plandan veya token başına ödemeden iyidir? Hackathon, sprint ve geçiş hesabı. - [Sınırsız LLM API planında "sınırsız" gerçekte ne demek?](https://broadrouter.com/tr/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Token sayacı, 5 saatlik pencere, haftalık sınır, hız sınırı, eşzamanlılık: "sınırsız" bir yapay zeka planı ajanınız için gerçekten sınırsız mı? 7 soru. - [Whisper API alternatifi: ne zaman geçilir, nasıl test edilir](https://broadrouter.com/tr/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): OpenAI Whisper API yerine barındırılan alternatif ne zaman mantıklı: kendi sesinizde kelime hata oranı, diller, saatlik fiyat, sınırlar ve gizlilik, test koduyla. - [Sıfır veri saklama (ZDR) LLM API: 10 soruluk kontrol listesi](https://broadrouter.com/tr/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): Bir LLM API için "sıfır veri saklama" ne anlama gelmeli: istemler, çıktılar, loglar, eğitim, alt işleyenler ve denetimler. Veri göndermeden önce sorulacak on soru. - [Guides in Turkish](https://broadrouter.com/tr/guides/) - [OpenAI Python SDK'yı BroadNet Inference ile kullanın](https://broadrouter.com/tr/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): Resmi OpenAI Python SDK'yı tek bir base_url değişikliğiyle BroadNet Inference'a yönlendirin. Kurun, anahtarı ayarlayın, ilk isteği gönderin ve yanıtı akışla alın. - [OpenAI Node.js SDK'yı BroadNet Inference ile kullanın](https://broadrouter.com/tr/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): Resmi OpenAI Node.js ve TypeScript SDK'yı baseURL ayarıyla BroadNet Inference'a yönlendirin. Kurun, ilk isteği gönderin, akış alın, istek kimliğini kaydedin. - [BroadNet Inference'ı hesap açmadan tarayıcıda deneyin](https://broadrouter.com/tr/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): Arapça konuşmadan metne, belge okuma ve Gemma 4 sohbetini BroadNet'in tarayıcı playground'unda kayıt ve kart olmadan deneyin, sonra aynı isteği API'ye taşıyın. - [curl ile BroadNet Inference'a ilk istek](https://broadrouter.com/tr/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): curl ile ilk BroadNet Inference isteklerinizi gönderin: sohbet yanıtı, akışla gelen yanıt ve ses transkripsiyonu. Yanıtı, kullanımı ve hataları okuyun. - [Chat completions akışı ve kopan bağlantıları ele alma](https://broadrouter.com/tr/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): BroadNet Inference'ta akış (streaming) nasıl çalışır: Server-Sent Events, [DONE] işareti, son parçadaki token kullanımı ve akış erken biterse ne yapılmalı. - [Akıl yürüten model neden boş yanıt döndürür?](https://broadrouter.com/tr/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3 ve Qwen3.8 27B yanıt vermeden önce düşünür ve bu akıl yürütme max_tokens sınırına dahildir. Boş yanıtlardan nasıl kurtulursunuz? - [BroadNet Inference hata kodları ve ne zaman yeniden denenmeli](https://broadrouter.com/tr/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): BroadNet Inference API durum kodlarının anlamı; hangilerinin geri çekilmeyle yeniden deneneceği, hangilerinin önce düzeltileceği, desteğe ne gönderileceği. - [API anahtarlarını IP izin listeleriyle sunucularınıza kilitleyin](https://broadrouter.com/tr/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): BroadNet Inference anahtarlarını IP izin listeleriyle koruyun, portal kimlik bilgilerinden ayırın ve sızan anahtarı saniyeler içinde iptal edin. curl örnekleriyle. - [Ajanlar için sınırsız LLM API planları nasıl çalışır?](https://broadrouter.com/tr/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): Sınırsız DeepSeek V4.1 Flash planlarında token sayacı ve haftalık sınır yok. Oturumlar, paralel ajanlar, anahtarlar ve faturalandırma aslında böyle çalışır. - [Aider'ı sınırsız DeepSeek V4.1 Flash ile çalıştırın](https://broadrouter.com/tr/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Aider kodlama ajanını OpenAI uyumlu ayarıyla BroadNet Inference'a bağlayın ve sabit fiyatlı, sınırsız DeepSeek V4.1 Flash planında çalıştırın. - [LangChain'i BroadNet Inference ile kullanın](https://broadrouter.com/tr/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): LangChain'in ChatOpenAI sınıfını base_url ve api_key ile BroadNet Inference'a bağlayın; LangChain'in akıl yürütme modellerinden neyi tutup neyi atmadığını öğrenin. - [Cline'ı BroadNet Inference ile kullanın (OpenAI Compatible sağlayıcı)](https://broadrouter.com/tr/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Cline VS Code eklentisini OpenAI Compatible sağlayıcısıyla BroadNet Inference'a bağlayın: alanlar, model kimlikleri, bağlam boyutları ve sık sorunların çözümü. - [Roo Code'u BroadNet Inference ile kullanın (OpenAI Compatible)](https://broadrouter.com/tr/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code Mayıs 2026'da kapatıldı. Mevcut kurulumlar ve çatallar için ayarlar ile BroadNet Inference'la nereye geçilir: Cline veya Zoo Code çatalı. - [Continue'yu BroadNet Inference ile kullanın (config.yaml ve apiBase)](https://broadrouter.com/tr/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): BroadNet Inference modellerini Continue IDE eklentisine config.yaml içinde provider openai ve apiBase ile ekleyin; tool_use yeteneği ve bağlam uzunluğu dahil. - [BroadNet Inference'ı opencode'a özel sağlayıcı olarak ekleyin](https://broadrouter.com/tr/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): BroadNet Inference'ı opencode kodlama ajanına opencode.jsonc içinde özel OpenAI uyumlu sağlayıcı olarak ekleyin: model takma adları, yetenekler ve sınırlar. - [Open WebUI'ı BroadNet Inference'a bağlayın](https://broadrouter.com/tr/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): BroadNet Inference'ı Open WebUI'a OpenAI uyumlu bağlantı olarak ekleyin, model kimliklerini izin listesine ekleyin; doğrulama, boş yanıt ve 429 sorunlarını çözün. - [BroadNet Inference'ı LibreChat'e özel uç nokta olarak ekleyin](https://broadrouter.com/tr/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): BroadNet Inference'ı librechat.yaml'da OpenAI uyumlu özel uç nokta olarak ekleyin: anahtar .env'de, model listesi, başlık modeli ve sorun çözümleri. - [n8n'de BroadNet Inference kullanımı (OpenAI kimlik bilgisi, Base URL)](https://broadrouter.com/tr/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): n8n'deki OpenAI kimlik bilgisini Base URL alanıyla BroadNet Inference'a yönlendirin, OpenAI Chat Model düğümünü ID moduyla ve Responses API kapalıyken kullanın. - [Vercel AI SDK'yı BroadNet Inference ile kullanın](https://broadrouter.com/tr/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): Vercel AI SDK'da @ai-sdk/openai-compatible içindeki createOpenAICompatible ile BroadNet Inference'ı çağırın: generateText, streamText, araçlar ve reasoning effort. - [Adım adım Arapça ses metne çevirme API'si](https://broadrouter.com/tr/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): BroadNet Transcribe Arabic ile Arapça sesi OpenAI uyumlu bir uç nokta üzerinden yazıya dökün. Formatlar, language alanı, Python ve curl örnekleri ve fiyat. - [LiteLLM'i BroadNet Inference'a yönlendirin (SDK ve proxy)](https://broadrouter.com/tr/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): BroadNet Inference'ı LiteLLM üzerinden openai/ model önekiyle ve api_base ile çağırın; Python SDK'da ya da proxy config.yaml dosyasında, uygun hız sınırlarıyla. - [OpenAI Agents SDK'yı BroadNet Inference üzerinde çalıştırın](https://broadrouter.com/tr/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Python için OpenAI Agents SDK'yı AsyncOpenAI ve OpenAIChatCompletionsModel ile BroadNet Inference'a bağlayın, OpenAI anahtarı isteyen tracing'i kapatın. - [WhatsApp sesli notlarını API ile metne çevirin](https://broadrouter.com/tr/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): WhatsApp .ogg sesli notlarını veya .m4a telefon kayıtlarını ffmpeg ile WAV'a dönüştürün ve tek API çağrısıyla Arapça ya da 13 başka dilde metne çevirin. - [n8n ile WhatsApp sesli notlarını metne çevirin](https://broadrouter.com/tr/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): WhatsApp sesli notunu alan, Meta'dan indiren, ffmpeg ile dönüştüren ve Arapça ya da 13 başka dilde metne çeviren bir n8n iş akışı. - [Arapça faturaları ve fişleri bir görü modeliyle okuyun](https://broadrouter.com/tr/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Bir fatura veya fiş fotoğrafını BroadNet Inference'taki bir görü modeline gönderin; toplamı, tarihi ve satıcıyı JSON olarak alın; Arapça ya da İngilizce sorun. - [Zed'i OpenAI uyumlu sağlayıcı olarak BroadNet Inference ile kullanın](https://broadrouter.com/tr/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): BroadNet Inference'ı Zed editöründe settings.json üzerinden OpenAI uyumlu sağlayıcı olarak ekleyin: model kimlikleri, token sınırları, yetenekler, anahtar değişkeni. - [Kilo Code'u BroadNet Inference ile kullanın (OpenAI Compatible)](https://broadrouter.com/tr/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Kilo Code'u VS Code iletişim kutusundan veya kilo.json ile BroadNet Inference'a özel OpenAI Compatible sağlayıcı olarak bağlayın: model kimliği, sınırlar, araçlar. - [BroadNet Inference'ı goose'a özel sağlayıcı olarak ekleyin](https://broadrouter.com/tr/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): goose yapay zekâ ajanını BroadNet Inference'a OpenAI Compatible özel sağlayıcı olarak bağlayın: goose Desktop, goose configure ya da bir JSON dosyasıyla. - [Pydantic AI ve BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/tr/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Pydantic AI ajanlarını OpenAIChatModel ve OpenAICompatibleProvider ile BroadNet Inference'ta çalıştırın: araçlar, yapılandırılmış çıktı, model profili. - [CrewAI'ı özel OpenAI uç noktası olarak BroadNet Inference ile kullanın](https://broadrouter.com/tr/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): CrewAI ajanlarını LLM(custom_openai=True, base_url=...) ile BroadNet Inference'a yönlendirin; ardından araçlar, uzun bağlam ve sabit fiyatlı plan için model seçin. - [LlamaIndex'i OpenAILike ile BroadNet Inference'ta kullanın](https://broadrouter.com/tr/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): LlamaIndex'i OpenAILike sınıfıyla BroadNet Inference'a bağlayın; api_base, context_window, is_chat_model ve is_function_calling_model değerlerini doğru ayarlayın. - [BroadNet Inference'ı OpenAI-API-compatible eklentisiyle Dify'a ekleyin](https://broadrouter.com/tr/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): Dify'ı OpenAI-API-compatible model sağlayıcı eklentisiyle BroadNet Inference'a bağlayın: birebir alan değerleri, bağlam boyutları ve function calling ayarları. - [Jan'da BroadNet Inference'ı OpenAI uyumlu özel sağlayıcı yapın](https://broadrouter.com/tr/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): BroadNet Inference'ı Jan masaüstü uygulamasına özel OpenAI uyumlu sağlayıcı olarak ekleyin, model kimliklerini elle girin ve araç ile görsel desteğini açın. - [SillyTavern'i BroadNet Inference'a bağlayın (Custom OpenAI-compatible)](https://broadrouter.com/tr/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): SillyTavern'i Custom (OpenAI-compatible) Chat Completion kaynağıyla BroadNet Inference'a bağlayın: doğru model kimlikleri, istem son işleme ve token sınırları. - [Spring AI'ı BroadNet Inference ile kullanın (OpenAI base-url)](https://broadrouter.com/tr/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Spring AI'ın OpenAI sohbet starter'ını spring.ai.openai.base-url, BroadNet model kimliği ve max-tokens ile BroadNet Inference'a yönlendirin, ChatClient'tan çağırın. - [BroadNet Inference'ı Go'dan openai-go SDK ile çağırın](https://broadrouter.com/tr/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): Resmi openai-go v3 SDK'yı option.WithBaseURL ile BroadNet Inference'a bağlayın: Go'da chat completions, usage bilgili akış ve hata yönetimi. - [BroadNet Inference üzerinde LangGraph ajanları oluşturun](https://broadrouter.com/tr/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): LangGraph StateGraph veya create_agent'ı ChatOpenAI, base_url ve bir BroadNet model kimliğiyle BroadNet Inference'ta çalıştırın: araçlar, akış ve bellek dahil. - [Haystack pipeline'larını ve ajanlarını BroadNet Inference ile kullanın](https://broadrouter.com/tr/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Haystack'in OpenAIChatGenerator bileşenini api_base_url, BroadNet model kimliği ve Secret anahtarla BroadNet Inference'a yönlendirin: pipeline, araç, akış, ajan. - [Semantic Kernel'i BroadNet Inference ile kullanın](https://broadrouter.com/tr/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Microsoft Semantic Kernel'i BroadNet Inference'a bağlayın: Python'da AsyncOpenAI istemcisiyle, C#'ta özel uç noktayla; sohbet, eklentiler ve ajanlar için. - [Mastra ajanlarını BroadNet Inference ile kullanın](https://broadrouter.com/tr/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): Mastra ajanlarını id, url ve apiKey içeren bir model nesnesiyle BroadNet Inference'ta çalıştırın; model kimliğinin gerektirdiği çift broadnet/ önekiyle birlikte. - [BroadNet Inference üzerinde Instructor ile yapılandırılmış çıktı alın](https://broadrouter.com/tr/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Instructor ile BroadNet Inference'tan Pydantic nesneleri çıkarın: özel base_url ile from_openai veya from_provider kullanın ve doğru Mode'u seçin. - [DSPy'ı BroadNet Inference ile kullanın (model_type="chat" ayarlayın)](https://broadrouter.com/tr/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): dspy.LM'i BroadNet Inference için openai/ öneki, api_base ve model_type="chat" ile yapılandırın; çünkü DSPy 3.5 OpenAI modellerini Responses API'ye gönderir. - [BroadNet modellerini promptfoo ile değerlendirin](https://broadrouter.com/tr/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): promptfoo değerlendirmelerini ve yan yana model testlerini openai:chat sağlayıcısı, apiBaseUrl ve apiKeyEnvar ile BroadNet Inference üzerinde çalıştırın. - [smolagents'ı OpenAIModel ile BroadNet Inference'ta çalıştırın](https://broadrouter.com/tr/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Hugging Face smolagents'ı OpenAIModel ve api_base ile BroadNet Inference'a bağlayın: CodeAgent her modelde, ToolCallingAgent GLM 5.3 ve Qwen3.8 27B'de çalışır. - [Flowise'ı BroadNet Inference'a bağlayın (OpenAI Custom Model)](https://broadrouter.com/tr/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): BroadNet Inference'ı Flowise chatflow'larında OpenAI Custom Model düğümüyle kullanın: Base Path, OpenAI API kimlik bilgisi ve elle yazılan model adı. - [Langflow'u BroadNet Inference'a bağlayın](https://broadrouter.com/tr/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): BroadNet Inference'ı Langflow'da OpenAI bileşeninin OpenAI API Base alanı ve elle yazılan model adıyla kullanın; OpenAI Compatible neden listelemez, öğrenin. - [Glossary entries in Turkish](https://broadrouter.com/tr/glossary/) - [Yapay zeka ajanı (AI agent)](https://broadrouter.com/tr/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): Yapay zeka ajanı (AI agent), bir dil modelinin bir hedefe giden adımlarını kendisinin yönettiği bir sistemdir: hangi araçları çağıracağına karar verir, sonuçları okur ve nihai bir yanıt üretene ya da bir sınıra takılana kadar tekrarlar. Modeli belirli noktalarda çağıran sabit bir betiğe ise genellikle iş akışı (workflow) denir. - [API anahtarı (API key)](https://broadrouter.com/tr/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): API anahtarı, bir API'yi çağıran istemciyi tanımlayan ve yetkilendiren gizli bir dizedir. LLM API'leri genellikle anahtarı Authorization: Bearer başlığında bekler; anahtarı elinde tutan herkes, sahibi olan hesaba faturalandırılan istekler yapabilir. - [Arapça harekeler (tashkeel)](https://broadrouter.com/tr/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): Arapça harekeler (tashkeel), kısa ünlüleri, ikizleşen ünsüzleri ve ünlü yokluğunu göstermek için harflerin üstüne veya altına yazılan küçük işaretlerdir. İsteğe bağlıdırlar: haberler dahil Arapça metinlerin çoğu onları yazmaz; okur da yazılım da ünlüleri bağlamdan çıkarır. - [Arapça lehçe tespiti (dialect identification)](https://broadrouter.com/tr/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): Arapça lehçe tespiti, bir metnin veya kaydın hangi Arapça varyantında olduğuna (Mısır, Körfez, Levant, Mağrip ya da Modern Standart Arapça) bölge, ülke veya şehir düzeyinde karar verme görevidir. Mesajları yönlendirmek, istem seçmek ve değerlendirme setlerini bölmek için kullanılır. - [Arapça metin normalizasyonu (normalization)](https://broadrouter.com/tr/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): Arapça metin normalizasyonu, aynı kelimenin yazım varyantlarını eşleştirmeden önce tek bir biçime indiren adımdır: harekeleri ve tatweel'i siler, elif biçimlerini birleştirir ve çoğu zaman ى'yı ي'ye, ة'yı ه'ye çevirir. Gösterilen metne değil, arama anahtarlarına ve değerlendirmeye uygulanır. - [Arabizi (Franco-Arabic)](https://broadrouter.com/tr/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): Arabizi, Latin harfleriyle gayriresmî yazılan Arapçadır; Latin karşılığı olmayan Arapça sesler için ع yerine 3 gibi rakamlar kullanır. Sohbette ve sosyal medyada yaygındır, çoğunlukla lehçelidir, standart bir yazımı yoktur ve sık sık İngilizce veya Fransızcayla karışır. - [AWQ (activation-aware weight quantization)](https://broadrouter.com/tr/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (activation-aware weight quantization, aktivasyona duyarlı ağırlık kuantizasyonu), bir LLM'in ağırlıklarını düşük hassasiyete, genellikle 4 bit tamsayılara sıkıştıran eğitim sonrası bir yöntemdir. Küçük bir kalibrasyon setinin aktivasyon istatistikleriyle en önemli ağırlık kanallarını bulur ve daha az doğruluk kaybetmeleri için ölçekler. Geri yayılım gerektirmez. - [Base URL (temel URL)](https://broadrouter.com/tr/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): Base URL (temel URL), https://api.example.com/v1 gibi, bir istemcinin /chat/completions gibi her uç nokta yolunun önüne eklediği sabit kök adrestir. OpenAI uyumlu SDK'larda base URL, API anahtarı ve model kimliğiyle birlikte değiştirildiğinde aynı kod farklı bir sağlayıcıya yönelir. - [Batch API (toplu işlem API'si)](https://broadrouter.com/tr/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Batch API, çok sayıda isteği genellikle bir dosya olarak tek seferde kabul eden, bunları hemen değil belirli bir süre içinde işleyen ve sonuçları birlikte döndüren asenkron bir LLM uç noktasıdır. OpenAI ve Anthropic gibi sağlayıcılar batch isteklerini standart API'lerine göre indirimli fiyatlandırır. - [BM25](https://broadrouter.com/tr/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25), anahtar kelime araması için bir sıralama fonksiyonudur; bir belgeyi sorgu terimlerinin içinde ne sıklıkla geçtiğine, bu terimlerin koleksiyonda ne kadar nadir olduğuna ve belgenin uzunluğuna göre puanlar. Lucene ve Elasticsearch'te varsayılan benzerlik fonksiyonudur. - [Byte pair encoding (BPE)](https://broadrouter.com/tr/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Byte pair encoding (BPE), tek tek karakterlerden veya byte'lardan başlayıp en sık görülen bitişik çifti hedef sözlük boyutuna ulaşana kadar tekrar tekrar yeni bir sözlük girdisine birleştiren bir alt kelime tokenizasyon algoritmasıdır. Günümüz LLM tokenizer'larının çoğu BPE'yi veya byte düzeyi bir türevini kullanır. - [Önbellekli girdi token'ları (cached input tokens)](https://broadrouter.com/tr/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): Önbellekli girdi token'ları, bir LLM sağlayıcısının kısa süre önce işlediği ve yeniden hesaplamak yerine önbellekten kullanabildiği istem token'larıdır; genellikle sistem istemi veya belge gibi tekrarlanan bir önektir. Sağlayıcılar bunları yeni girdi token'larından daha düşük ücretle faturalandırır. - [Düşünce zinciri (chain-of-thought, CoT)](https://broadrouter.com/tr/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Düşünce zinciri (chain-of-thought, CoT), bir dil modeline nihai cevabından önce ara akıl yürütme adımlarını yazmasının gösterildiği veya istendiği bir istem tekniğidir. Aritmetik ve mantık gibi çok adımlı problemlerde doğruluğu artırır; akıl yürütme modelleri artık bunu kendiliğinden üretir. - [Karakter hata oranı (CER)](https://broadrouter.com/tr/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Karakter hata oranı (CER), transkripsiyon veya OCR doğruluğunu karakter düzeyinde ölçer: çıktıyı referans metne dönüştürmek için gereken değiştirme, silme ve ekleme sayısı, referanstaki karakter sayısına bölünür. Düşük olan daha iyidir; 0 tam eşleşme demektir. - [Chat Completions API](https://broadrouter.com/tr/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Chat Completions API, OpenAI'nin bir model adı ile rollerle etiketlenmiş mesaj listesini alıp modelin sonraki mesajını döndüren HTTP uç noktasıdır: POST /v1/chat/completions. İstek ve yanıt yapısı, OpenAI uyumlu sağlayıcıların ve SDK'ların yeniden kullandığı ortak format hâline geldi. - [Sohbet şablonu (chat template)](https://broadrouter.com/tr/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): Sohbet şablonu (chat template), genellikle Jinja ile yazılan ve modelin tokenizer'ıyla birlikte gelen, rollere sahip sohbet mesajlarından oluşan bir listeyi modelin beklediği tek token dizisine, özel kontrol token'larıyla birlikte dönüştüren biçimlendirme kuralıdır. Yanlış şablon, modelin cevaplarını ciddi biçimde bozabilir. - [Parçalama (chunking, RAG)](https://broadrouter.com/tr/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Parçalama (chunking), bir getirme (retrieval) işlem hattında belgeleri embedding'e çevrilip indekslenmeden önce chunk denen küçük parçalara bölen adımdır. Chunk boyutu her parçanın ne kadar metin tuttuğunu belirler; chunk örtüşmesi ise sınırda kesilen fikirler kaybolmasın diye komşu parçalar arasında bir miktar metni tekrarlar. - [Kod değiştirme (code-switching)](https://broadrouter.com/tr/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Kod değiştirme (code-switching), konuşmada veya yazıda aynı sohbet ya da cümle içinde iki veya daha fazla dilin ya da bir dilin varyantlarının kullanılmasıdır. Örnekler arasında Arapça konuşanların lehçeye İngilizce kelimeler karıştırması veya Modern Standart Arapça ile kendi lehçeleri arasında gidip gelmesi vardır. - [Yapay zekâ kodlama ajanı (AI coding agent)](https://broadrouter.com/tr/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): Yapay zekâ kodlama ajanı, bir dil modelinin bir kod tabanı üzerinde döngü içinde çalıştığı bir yazılım aracıdır: dosyaları okur, düzenler, testler gibi komutları çalıştırır, çıktıyı okur ve görev bitene ya da bir sınıra ulaşılana kadar bir sonraki adıma karar verir. - [Eşzamanlılık (concurrency)](https://broadrouter.com/tr/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): Eşzamanlılık (concurrency), bir istemcinin aynı anda işlemde olan istek sayısıdır. LLM API'leri bunu genellikle dakikadaki istek sayısından ayrı olarak anahtar veya plan başına sınırlar, çünkü açık her istek yanıtı bitene kadar GPU kapasitesi tutar. - [Bağlam sıkıştırma (context compaction)](https://broadrouter.com/tr/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): Bağlam sıkıştırma (context compaction), bağlam penceresi sınırına yaklaşan bir sohbeti özetleyip oturuma bu özetten devam etme yöntemidir; böylece ajan, eski araç çıktıları gibi ham geçmişi atarken önemli kararları ve durumu korur. - [Bağlam mühendisliği (context engineering)](https://broadrouter.com/tr/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): Bağlam mühendisliği (context engineering), bir dil modelinin her istekte gördüğü token kümesini, yani sistem istemini, araç tanımlarını, geçmişi, getirilen belgeleri ve notları seçme ve sürdürme işidir; amaç, modelin görevi iyi yapması için gereken en küçük, en yüksek sinyalli bilgiyi vermektir. - [Bağlam çürümesi (context rot)](https://broadrouter.com/tr/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): Bağlam çürümesi (context rot), dil modellerinin girdisi uzadıkça, görev aynı kalsa ve girdi bağlam penceresinin rahatça içinde olsa bile, doğruluk ve güvenilirlikte yaşadığı kademeli kayıptır. Terimi, 18 modeli test eden 2025 tarihli bir Chroma teknik raporu yaygınlaştırdı. - [Bağlam penceresi (context window)](https://broadrouter.com/tr/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): Bağlam penceresi, bir dil modelinin tek bir istekte işleyebileceği en fazla token sayısıdır. Sistem istemini, konuşma geçmişini, araç tanımlarını, görselleri ve akıl yürütme token'ları dahil üretilen çıktıyı kapsar. Sığmayan her şey kesilmeli, özetlenmeli veya birden fazla isteğe bölünmelidir. - [Veri yerleşimi (data residency)](https://broadrouter.com/tr/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): Veri yerleşimi (data residency), verilerin belirli bir ülke veya bölgede saklanıp işlenmesine yönelik bir gereklilik ya da taahhüttür. Bir LLM API'sinde istemlerin çıkarım sırasında nerede işlendiğini ve saklanan içeriğin nerede durduğunu kapsar; hesap ve faturalandırma verilerini çoğu zaman dışarıda bırakır. - [Arapça lehçeler (dialectal Arabic)](https://broadrouter.com/tr/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): Arapça lehçeler (dialectal Arabic), Mısır, Körfez, Levant veya Mağrip Arapçası gibi bir bölgenin ya da şehrin gündelik konuşulan Arapçasıdır. Lehçeler Modern Standart Arapçadan ses, kelime ve dilbilgisi bakımından ayrılır, standart bir yazımları yoktur ve internette giderek daha çok yazılırlar; bu da onları dil modelleri için zorlaştırır. - [Konuşmacı ayrımı (speaker diarization)](https://broadrouter.com/tr/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): Konuşmacı ayrımı (speaker diarization), bir ses kaydını konuşmacılara göre bölümlere ayırma ve "kim ne zaman konuştu" sorusunu yanıtlama görevidir. Her konuşma bölümünü SPEAKER_00 gibi anonim bir konuşmacı kimliğiyle etiketler ve genellikle her satırı kimin söylediğini göstermek için transkriptle birleştirilir. - [Damıtma (knowledge distillation)](https://broadrouter.com/tr/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): Damıtma (distillation) ya da bilgi damıtma (knowledge distillation), daha küçük bir öğrenci modelin daha büyük bir öğretmen modelin olasılık dağılımları veya ürettiği yanıtlar gibi çıktılarını yeniden üretmeyi öğrendiği bir eğitim yöntemidir; böylece öğrenci, öğretmenin kalitesinin büyük kısmını daha düşük çalıştırma maliyetiyle korur. - [Veri işleme sözleşmesi (DPA)](https://broadrouter.com/tr/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): Veri işleme sözleşmesi (DPA, data processing agreement), veri sorumlusu ile veri işleyen arasında, veri işleyenin kişisel verileri nasıl ele alacağını düzenleyen sözleşmedir. GDPR Madde 28(3) bunu zorunlu kılar ve içeriğini sayar: belgelenmiş talimatlar, gizlilik, güvenlik, alt işleyen kuralları, verilerin silinmesi veya iadesi ve denetimler. - [Embedding (gömme vektörü)](https://broadrouter.com/tr/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): Embedding'ler (gömme vektörleri), bir modelin bir metin parçasını, görseli veya başka bir girdiyi temsil etmek için ürettiği kayan noktalı sayı listeleridir (vektörler). Anlamca benzer girdiler birbirine yakın vektörler alır; bu yüzden vektörler arasındaki mesafe arama sonuçlarını sıralamak, belgeleri kümelemek veya ilgili öğeleri bulmak için kullanılabilir. - [Eval (LLM değerlendirmesi)](https://broadrouter.com/tr/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Eval (evaluation, yani değerlendirme kelimesinin kısaltması), bir dil modelinin veya istemin tekrarlanabilir bir testidir: sabit bir girdi seti, iyi bir çıktının neye benzediğinin tarifi ve sürümlerin karşılaştırılabilmesi için her çıktıyı puanlayan bir değerlendirici (kod, bir insan veya başka bir model). - [Üstel geri çekilme (exponential backoff)](https://broadrouter.com/tr/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): Üstel geri çekilme (exponential backoff), istemcinin her başarısız denemeden sonra tekrar denemeden önce daha uzun beklediği, genellikle gecikmeyi bir üst sınıra kadar ikiye katladığı bir yeniden deneme stratejisidir. Her beklemeye rastgele jitter eklemek, birçok istemcinin yeniden denemelerini yayar; böylece sunucuya aynı anda yüklenmezler. - [Few-shot prompting (az örnekli istem)](https://broadrouter.com/tr/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): Few-shot prompting (az örnekli istem), istemin gerçek girdiden önce az sayıda çözülmüş girdi-çıktı örneği içerdiği bir tekniktir; dil modeli görevi ve biçimi bu örneklerden çıkarır. Modelin ağırlıkları değişmez; zero-shot hiç örnek kullanmaz, one-shot ise bir örnek kullanır. - [İnce ayar (fine-tuning)](https://broadrouter.com/tr/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): İnce ayar (fine-tuning), önceden eğitilmiş bir dil modelini örnek sohbetler gibi daha küçük, göreve özgü bir veri setiyle eğitmeye devam etme işlemidir; böylece ağırlıkları değişir ve model, her istemde uzun talimatlara veya örneklere ihtiyaç duymadan o görevi ya da üslubu daha güvenilir biçimde yerine getirir. - [finish_reason (bitiş nedeni)](https://broadrouter.com/tr/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason, bir chat completion yanıtındaki her seçimde bulunan ve modelin üretimi neden durdurduğunu söyleyen alandır. OpenAI formatında değerler stop, length, tool_calls, content_filter ve kullanımdan kaldırılmış function_call'dır. Anthropic'in Messages API'si aynı bilgiyi stop_reason olarak bildirir. - [Sabit fiyatlı LLM planı (flat-rate)](https://broadrouter.com/tr/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): Sabit fiyatlı LLM planı, bir dil modeli API'sine erişim için token başına faturalandırma yerine ay gibi bir dönem için sabit bir ücret alır. Bu planlar genellikle eşzamanlılık, hız, model listesi veya kullanım pencereleri gibi başka bir şeyi sınırlar; bu yüzden sınırlar en az fiyat kadar önemlidir. - [FP8](https://broadrouter.com/tr/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8, derin öğrenmede kullanılan, iki kodlamayla tanımlanmış 8 bitlik bir kayan noktalı sayı biçimidir: 4 üs ve 3 mantis bitli E4M3 ile 5 üs ve 2 mantis bitli E5M2. 16 bitlik biçimlere göre belleği yarıya indirir ve yeni GPU'larda yerel olarak çalışır. - [Frekans ve varlık cezası (frequency/presence penalty)](https://broadrouter.com/tr/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): Frekans ve varlık cezaları, çıktıda zaten bulunan token'ların puanını düşüren örnekleme ayarlarıdır. Varlık cezası (presence penalty), bir token bir kez göründüğünde uygulanır; frekans cezası (frequency penalty) her tekrarda artar. OpenAI'nin API'sinde ikisinin de aralığı -2.0 ile 2.0 arasıdır ve varsayılan 0'dır. - [GDPR (Genel Veri Koruma Tüzüğü)](https://broadrouter.com/tr/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): GDPR (General Data Protection Regulation, Genel Veri Koruma Tüzüğü, 2016/679 sayılı AB Tüzüğü), AB'nin kişisel verilerin işlenmesine ilişkin yasasıdır. 25 Mayıs 2018'den beri uygulanır; AB'de kurulu kuruluşları ve AB dışındaki bazılarını kapsar. LLM API'lerinde veri işleyen sözleşmelerini, uluslararası aktarımları ve veri minimizasyonunu düzenler. - [GGUF](https://broadrouter.com/tr/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF, dil modellerini llama.cpp gibi GGML tabanlı çalışma ortamlarında çıkarım için saklayan ikili bir dosya biçimidir. Tek bir GGUF dosyası, çoğu zaman kuantize edilmiş ağırlıkları ve modeli yüklemek için gereken tokenizer ve mimari ayrıntıları gibi meta verileri barındırır; hızlı yükleme için belleğe eşlenebilir. - [GPU saati (GPU-hour)](https://broadrouter.com/tr/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): GPU saati (GPU-hour), bir grafik işlemcinin bir saat boyunca kullanılmasıdır. Bulut sağlayıcıları kiralık GPU örneklerini saat veya saniye bazında faturalar; araştırmacılar da bir işin ne kadar hesaplama gerektirdiğini GPU saatiyle ifade eder. Makine meşgul olsun olmasın, onun süresi için ödeme yaparsınız. - [Temellendirme (grounding)](https://broadrouter.com/tr/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): Temellendirme (grounding), bir dil modelinin çıktısını sağlanan belgeler, arama sonuçları veya veritabanı kayıtları gibi belirli, doğrulanabilir bilgi kaynaklarına bağlama ve modele bunlardan yanıt vermesini söyleme pratiğidir. Temellendirilmiş yanıtlar bir kaynağa kadar izlenebilir ve uydurma bilgi içerme olasılıkları daha düşüktür. - [Guardrails (koruma bariyerleri)](https://broadrouter.com/tr/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): Guardrails (koruma bariyerleri), bir uygulamanın dil modeli çağrısını belirlenen sınırlar içinde tutmak için çağrının etrafında çalıştırdığı denetimlerdir. Girdi guardrail'leri modele gireni, çıktı guardrail'leri modelden çıkanı denetler, araç guardrail'leri ise eylemleri çalışmadan önce doğrular. Başarısız bir denetim isteği engeller, yeniden yazar veya işaretler. - [Halüsinasyon (yapay zeka halüsinasyonu)](https://broadrouter.com/tr/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): Halüsinasyon, bir dil modelinin kulağa makul gelen ama yanlış, uydurma ya da kendisine verilen girdiyle desteklenmeyen çıktısıdır; örneğin uydurma bir kaynak, yanlış bir sayı veya var olmayan bir API. Modelin gerçekleri kontrol etmek yerine olası metni tahmin etmesinden kaynaklanır. - [Hemze (hamza)](https://broadrouter.com/tr/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): Hemze (ء), Arapçada gırtlak durmasını gösteren harftir. Tek başına ya da bir taşıyıcı üzerinde yazılır: elif (أ إ), vav (ؤ) veya ye (ئ); آ ise elifi medde ile birleştirir. Klavyeyle yazılan metin bu konuda tutarsız olduğundan, arama motorları ve konuşma puanlaması elif biçimlerini çoğu zaman ا harfine indirger. - [Hibrit arama (hybrid search)](https://broadrouter.com/tr/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): Hibrit arama, aynı sorgu üzerinde hem BM25 gibi bir anahtar kelime aramasını hem de embedding'ler üzerinde vektör aramasını çalıştıran, ardından iki sıralı listeyi genellikle reciprocal rank fusion (RRF) ya da normalize edilmiş skorların ağırlıklı toplamıyla tek listede birleştiren bir getirme (retrieval) yöntemidir. - [Idempotency anahtarı (idempotency key)](https://broadrouter.com/tr/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Idempotency anahtarı, istemcinin ödeme veya satın alma gibi durum değiştiren bir isteğe eklediği, genellikle UUID gibi rastgele bir kimlik olan benzersiz bir değerdir. Aynı istek aynı anahtarla yeniden gönderilirse sunucu işlemi ikinci kez yapmak yerine ilk sonucu döndürür. - [Görsel token'ları (image tokens)](https://broadrouter.com/tr/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): Görsel token'ları, bir görsel dil modelinin istemdeki bir resmi temsil etmek için kullandığı girdi token'larıdır. Sağlayıcılar bunları resmin boyutundan, genellikle yeniden boyutlandırdıktan sonra sabit boyutlu yamalara veya karolara bölerek hesaplar ve metin girdi token'ları gibi faturalandırır; bu yüzden büyük resimler daha pahalıdır. - [Çıkarım sağlayıcısı (inference provider)](https://broadrouter.com/tr/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): Çıkarım sağlayıcısı, makine öğrenimi modellerini, çoğu zaman açık ağırlıklı LLM'leri, sahip olduğu, uzun süreli kiraladığı veya saatlik kiraladığı GPU'larda çalıştıran ve bunlara genellikle token, istek veya zaman birimi başına fiyatlandırılan bir API üzerinden erişim satan şirkettir. Sunumu, ölçeklemeyi ve donanımı o üstlenir. - [Girdi ve çıktı token'ları](https://broadrouter.com/tr/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): Girdi token'ları, bir dil modelinin istekte okuduğu token'lardır: sistem istemi, konuşma geçmişi, araç tanımları ve görseller dahil. Çıktı token'ları ise modelin yanıt olarak ürettikleridir, akıl yürütme dahil. LLM API'leri bunları genellikle ayrı fiyatlandırır ve çıktı genellikle daha pahalıdır. - [INT4](https://broadrouter.com/tr/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4, 16 düzeyi temsil edebilen 4 bitlik bir tamsayı biçimidir. LLM çıkarımında genellikle 4 bit yalnızca ağırlık kuantizasyonu (W4A16) anlamına gelir: ağırlıklar ölçek katsayılarıyla INT4 olarak saklanır, hesaplama sırasında 16 bite açılır ve 16 bitlik ağırlıkların yaklaşık dörtte biri kadar bellek kullanır. - [IP izin listesi (IP allow list)](https://broadrouter.com/tr/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): IP izin listesi, yalnızca listelenen IP adreslerinden veya CIDR aralıklarından gelen istekleri kabul edip diğer tümünü reddeden bir güvenlik denetimidir. Bir API'de sızdırılmış bir anahtarın vereceği zararı sınırlar, çünkü anahtar listede olmayan ağlardan çalışmaz. - [Jailbreak (LLM)](https://broadrouter.com/tr/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): Jailbreak, bir dil modelinin güvenlik eğitimini veya onu kullanan uygulamanın koyduğu kuralları yok saymasını sağlamak için hazırlanmış bir girdidir; amaç modelin normalde reddedeceği içeriği üretmesi veya eylemleri yapmasıdır. OWASP, jailbreak'i bir prompt injection (istem enjeksiyonu) biçimi olarak ele alır. - [JSON modu (JSON mode)](https://broadrouter.com/tr/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): JSON modu, OpenAI tarzı API'lerde response_format {"type": "json_object"} ile açılan ve modelin yanıtını geçerli, ayrıştırılabilir JSON yapan bir LLM API ayarıdır. JSON Schema kullanan yapılandırılmış çıktıların aksine, nesnenin hangi anahtarları, türleri veya değerleri içereceğini garanti etmez. - [JSON Schema (JSON şeması)](https://broadrouter.com/tr/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema, JSON verisinin yapısını tanımlamak ve doğrulamak için JSON ile yazılan bir sözlüktür: türler, zorunlu alanlar, izin verilen değerler ve iç içe yapı. Güncel sürüm 2020-12'dir. LLM API'leri onu araç parametrelerini ve yapılandırılmış çıktının izlemesi gereken kesin yapıyı tanımlamak için kullanır. - [KV önbelleği (KV cache)](https://broadrouter.com/tr/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): KV önbelleği, bir transformer dil modelinin daha önce işlediği her token için hesapladığı dikkat anahtarlarını ve değerlerini sakladığı bellektir; böylece her yeni token tüm diziyi yeniden hesaplamak yerine bunları yeniden kullanır. Bağlam uzunluğuyla büyür ve çoğu zaman bir GPU'nun kaç isteğe hizmet verebileceğini sınırlar. - [Gecikme (latency)](https://broadrouter.com/tr/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): Gecikme (latency), bir dil modeli API'sinde, isteğin gönderilmesinden son çıktı token'ının alınmasına kadar geçen süredir. İlk token süresi ile üretim süresinin toplamıdır ve genellikle ortalama yerine p50 ve p95 gibi yüzdeliklerle raporlanır. - [Hakem olarak LLM (LLM-as-a-judge)](https://broadrouter.com/tr/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge (hakem olarak LLM), bir dil modelinin diğer model çıktılarını notlandırdığı bir değerlendirme yöntemidir: ya tek bir yanıtı bir rubriğe göre puanlar ya da iki yanıttan daha iyisini seçer. İnsan yargısına çok daha düşük maliyetle yaklaşır, ancak ilk veya daha uzun yanıtı kayırmak gibi bilinen yanlılıkları vardır. - [LLM benchmark (kıyaslama testi)](https://broadrouter.com/tr/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): LLM benchmark, dil modellerini tek bir beceride, örneğin bilgi soruları, kodlama veya konuşma tanımada ölçmek ve karşılaştırmak için kullanılan, puanlama yöntemi olan sabit ve genellikle herkese açık bir görev setidir. Benchmark skoru o testteki göreli yeteneği gösterir; modelin sizin iş yükünüzde nasıl performans göstereceğini değil. - [LLM yönlendirici (LLM router)](https://broadrouter.com/tr/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): LLM yönlendirici (router), bir uygulama ile bir veya daha fazla dil modeli API'si arasında duran ve her istek için maliyet, gecikme, erişilebilirlik veya görev türü gibi kurallara göre hangi modelin veya sağlayıcının isteği karşılayacağına karar veren bir yazılım ya da hizmettir. - [Logprobs (log olasılıkları)](https://broadrouter.com/tr/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Logprobs, bir dil modelinin ürettiği token'lara atadığı log olasılıklarıdır. Bunları döndüren bir API, her çıktı token'ı için modelin onu ne kadar olası bulduğunu ve isteğe bağlı olarak o konumdaki en olası alternatif token'ları gösterir. - [LoRA (düşük ranklı uyarlama)](https://broadrouter.com/tr/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation, düşük ranklı uyarlama), önceden eğitilmiş bir modelin ağırlıklarını dondurup seçilen katmanlar için bir çift küçük düşük ranklı matris eğiten, parametre açısından verimli bir ince ayar yöntemidir. Öğrenilen güncelleme küçük bir adaptör olarak saklanabilir veya ağırlıklarla birleştirilebilir; çıkarıma gecikme eklemez. - [Lost in the middle (ortada kaybolma)](https://broadrouter.com/tr/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle (ortada kaybolma), dil modellerinin ilgili bilgiyi uzun bir girdinin başında veya sonunda olduğunda en iyi, ortasında olduğunda ise en kötü kullanma eğilimidir. Adını Liu vd.'nin 2023 tarihli çalışması koymuştur. - [max_tokens](https://broadrouter.com/tr/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens, bir modelin tek bir yanıtta üretebileceği en yüksek token sayısını belirleyen bir LLM API parametresidir. Üretim bu sınıra ulaştığında durur ve yanıt kesilmiş olarak işaretlenir; örneğin finish_reason değeri "length" olur. - [MCP (Model Context Protocol)](https://broadrouter.com/tr/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol, Model Bağlam Protokolü), LLM uygulamalarının harici araçlara ve verilere nasıl bağlanacağını standartlaştıran, JSON-RPC 2.0 üzerine kurulu açık bir protokoldür. Bir MCP sunucusu araçlar, kaynaklar ve istemler sunar; host uygulamanın içindeki MCP istemcisi bunları keşfeder ve modelin araç çağrılarını sunucuya iletir. - [Uzmanlar karışımı (mixture of experts, MoE)](https://broadrouter.com/tr/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Uzmanlar karışımı (MoE), bir katmanın çok sayıda uzman alt ağ barındırdığı ve öğrenilmiş bir yönlendiricinin her token'ı bunların yalnızca birkaçına gönderdiği bir sinir ağı tasarımıdır. Bir MoE dil modelinin toplam parametre sayısı büyüktür, ancak her token için küçük bir aktif alt küme kullanır. - [Model kullanımdan kaldırma (model deprecation)](https://broadrouter.com/tr/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): Model kullanımdan kaldırma (deprecation), bir API sağlayıcısının bir modeli artık sunmayacağını, genellikle bir emeklilik veya kapatma tarihi ve önerilen bir yerine geçen modelle birlikte duyurmasıdır. Bu tarihten sonra modele yapılan istekler başarısız olur; bu yüzden uygulamaların önceden başka bir model kimliğine geçmesi gerekir. - [Yedek modele geçiş (model fallback)](https://broadrouter.com/tr/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): Yedek modele geçiş (model fallback), başarısız bir LLM isteğinin sıralı bir listedeki farklı bir model veya sağlayıcıda yeniden denendiği bir güvenilirlik kalıbıdır. Tetikleyici bir hata, hız sınırı, bağlam uzunluğu hatası veya içerik politikası reddi olabilir. İstemcinizde ya da API'lerin önündeki bir gateway'de çalışır. - [Model lisansı (model license)](https://broadrouter.com/tr/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): Model lisansı, bir modelin ağırlıklarının, kodunun ve çıktılarının hangi yasal koşullarla kullanılabileceğini, değiştirilebileceğini ve paylaşılabileceğini belirleyen şartlardır. Açık ağırlıklı modeller Apache 2.0 veya MIT gibi standart yazılım lisanslarını ya da büyük ticari hizmetlere sınırlar gibi ek koşullar getirebilen özel model lisanslarını kullanır. - [Modern Standart Arapça (MSA, Fusha)](https://broadrouter.com/tr/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): Modern Standart Arapça (MSA), Arap dünyasında yazıda, haberlerde, eğitimde ve resmi konuşmada kullanılan standartlaşmış Arapçadır. Kimsenin günlük ana dili değildir: insanlar Mısır veya Körfez Arapçası gibi bölgesel lehçeleri konuşur ve resmi ortamlarda MSA'ya geçer. - [Çoklu ajan sistemi (multi-agent system)](https://broadrouter.com/tr/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): Çoklu ajan sistemi, her birinin kendi istemi, araçları ve bağlamı olan birkaç ajanın tek bir görevin parçaları üzerinde çalıştığı bir LLM uygulamasıdır. Yaygın bir tasarım orkestratör-işçi düzenidir: bir lider ajan plan yapar, işçi ajanları paralel başlatır ve sonuçlarını birleştirir. - [Çok modlu yapay zeka (multimodal AI)](https://broadrouter.com/tr/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): Çok modlu yapay zeka, metin, görsel, video ve ses gibi birden fazla veri türüyle çalışan bir model veya sistemdir. Çok modlu bir model bir fotoğrafı ve bir soruyu birlikte okuyabilir ya da konuşmayı girdi olarak alıp metinle yanıt verebilir. En yaygın türü görsel-dil modelidir. - [Varlık ismi tanıma (named entity recognition, NER)](https://broadrouter.com/tr/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): Varlık ismi tanıma (NER), metindeki adları bulup her birini kişi, kurum veya konum gibi bir türle etiketleme görevidir. Klasik sistemler her token'ı etiketler; LLM'lerden ise genellikle varlıkların bir listesini JSON olarak döndürmeleri istenir ve bu liste daha sonra kaynak metne karşı kontrol edilir. - [Optik karakter tanıma (OCR)](https://broadrouter.com/tr/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): Optik karakter tanıma (OCR), bir görsel, tarama veya PDF içindeki basılı ya da el yazısı metni tespit edip makinenin okuyabileceği karakterlere dönüştüren teknolojidir. Klasik OCR metni ve sayfadaki konumunu verir; görsel-dil modelleri ise metni okuyup onunla ilgili soruları doğrudan yanıtlayabilir. - [Ogg Opus](https://broadrouter.com/tr/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus, Opus kodekiyle (RFC 6716) kodlanıp bir Ogg kapsayıcısında (RFC 7845) saklanan sestir. Düşük bit hızlarında konuşma ve etkileşimli ses için tasarlanmıştır ve WhatsApp'ın sesli mesajlar için kullandığı formattır; dosyalar genellikle .ogg veya .opus uzantılıdır. - [Açık ağırlıklı model (open-weight model)](https://broadrouter.com/tr/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): Açık ağırlıklı model (open-weight model), eğitilmiş ağırlıkları herkesin indirip çalıştırabileceği şekilde, genellikle koşullar içeren bir lisansla yayımlanan bir makine öğrenimi modelidir. Eğitim kodunu ve eğitim verileri hakkında ayrıntılı bilgiyi de yayımlayan, böylece başkalarının modeli incelemesine ve yeniden üretmesine imkân veren tam açık kaynak yapay zekadan farklıdır. - [OpenAI uyumlu API](https://broadrouter.com/tr/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): OpenAI uyumlu API, POST /v1/chat/completions gibi OpenAI API'siyle aynı istek ve yanıt biçimlerini kabul eden bir HTTP API'sidir; böylece OpenAI SDK'ları ve araçları yalnızca base URL, API anahtarı ve model adı değiştirilerek onu kullanabilir. - [Paralel araç çağrıları (parallel tool calls)](https://broadrouter.com/tr/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Paralel araç çağrıları, bir modelin tek bir yanıtta birden fazla araç çağrısı döndürmesidir; örneğin aynı anda üç dosya okuma. OpenAI formatında bunlar tool_calls içinde birkaç girdi olarak gelir ve parallel_tool_calls değerini false yapmak modeli tur başına sıfır veya bir çağrıyla sınırlar. - [Kullandıkça öde (pay-as-you-go)](https://broadrouter.com/tr/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): Kullandıkça öde, abonelik veya asgari taahhüt olmadan yalnızca kullandığınız kadar ödediğiniz, token, istek veya ses dakikası gibi birimler üzerinden ölçülen bir fiyatlandırma modelidir. LLM API'lerinin çoğu bu şekilde, ön ödemeli bir bakiyeden veya dönemsel faturayla ücretlendirir. - [PDPL (Suudi Arabistan ve BAE veri koruma yasaları)](https://broadrouter.com/tr/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law, Kişisel Verilerin Korunması Kanunu), Suudi Arabistan'ın M/19 sayılı Kraliyet Kararnamesiyle çıkarılan ve SDAIA tarafından uygulanan kişisel veri yasası ile BAE'nin 2021 tarihli 45 sayılı Federal Kanun Hükmünde Kararnamesi için kullanılan ortak addır. İkisi de kişisel verilerin işlenmesini düzenler ve yabancı LLM API'lerine gönderilen istemler dahil yurt dışına aktarımlara koşullar koyar. - [PII (kişisel olarak tanımlanabilir bilgi)](https://broadrouter.com/tr/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): PII (kişisel olarak tanımlanabilir bilgi), bir kişinin kimliğini ayırt etmek veya izini sürmek için kullanılabilen ad, kimlik numarası veya biyometrik kayıt gibi bilgilerle, o kişiye bağlı ya da bağlanabilir diğer bilgilerdir. GDPR'daki daha geniş terim kişisel veridir. LLM çalışmalarında PII maskeleme, bu bilgileri istemler sistemlerinizden çıkmadan önce kaldırır. - [Prefill ve decode (LLM çıkarım aşamaları)](https://broadrouter.com/tr/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill ve decode, LLM çıkarımının iki aşamasıdır. Prefill aşamasında model istemin tamamını paralel olarak işler ve KV önbelleğini doldurur; decode aşamasında bu önbelleği yeniden kullanarak yanıtı token token üretir. İlk token süresini çoğunlukla prefill, çıktı hızını ise decode belirler. - [Ön ödemeli kredi (prepaid credit)](https://broadrouter.com/tr/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): Ön ödemeli kredi, kullanımdan önce bir API hesabına yüklenen paradır. Her istek kendi maliyetini bakiyeden düşer; bakiye bittiğinde ölçülen istekler siz yeniden yükleme yapana kadar durur. Böylece harcama, yatırdığınız tutarla sınırlı kalır. - [Prompt mühendisliği (istem mühendisliği)](https://broadrouter.com/tr/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): Prompt mühendisliği (istem mühendisliği), bir dil modeline verilen talimatları, gereksinimlerinizi karşılayan çıktıyı tutarlı biçimde üretecek şekilde yazma, yapılandırma ve test etme sürecidir. Rol mesajlarını, açık talimatları, örnekleri, biçimlendirmeyi ve sonuçların test vakalarıyla ölçülmesini kapsar. - [Prompt injection (istem enjeksiyonu)](https://broadrouter.com/tr/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): Prompt injection (istem enjeksiyonu), bir dil modeline kullanıcı tarafından verilen ya da bir web sayfasına, e-postaya, dosyaya veya araç sonucuna gizlenen metnin, modelin davranışını geliştiricinin niyetine aykırı biçimde değiştirdiği bir saldırıdır. OWASP, LLM Uygulamaları için Top 10 listesinde onu birinci sıraya (LLM01:2025) koyar. - [Kuantizasyon (quantization)](https://broadrouter.com/tr/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): Kuantizasyon (quantization), bir sinir ağının sayılarını, özellikle ağırlıklarını, 16 bit yerine 8 bit veya 4 bit gibi daha düşük hassasiyetli bir biçimde saklama işlemidir; böylece model daha az bellek kullanır ve çoğu zaman daha hızlı çalışır. İyi yöntemler doğruluk kaybını küçük tutar. - [RAG (erişimle artırılmış üretim)](https://broadrouter.com/tr/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG (retrieval-augmented generation, erişimle artırılmış üretim), bir uygulamanın önce bir soruyla ilgili pasajları arama veya vektör dizini gibi harici bir depodan getirdiği, ardından bunları isteme eklediği bir yöntemdir; böylece dil modeli yalnızca belleğine değil, bu metne dayanarak yanıt verebilir. - [Hız sınırı (rate limit)](https://broadrouter.com/tr/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): Hız sınırı (rate limit), bir API'nin bir istemcinin belirli bir zaman aralığındaki kullanımına koyduğu üst sınırdır; örneğin dakika başına istek veya dakika başına token. Sınırı aşan istekler, kullanım tekrar sınırın altına inene kadar genellikle HTTP 429 Too Many Requests ile reddedilir. - [ReAct (reason and act, akıl yürüt ve eyleme geç)](https://broadrouter.com/tr/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (reason and act), Yao ve arkadaşlarının 2022'de tanıttığı ve bir dil modelinin akıl yürütme izlerini arama veya araç çağrısı gibi eylemlerle iç içe yürüttüğü bir istem düzenidir. Model her gözlemi okur, planını günceller ve nihai bir yanıt verebilene kadar bunu tekrarlar. - [Akıl yürütme token'ları (reasoning tokens)](https://broadrouter.com/tr/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): Akıl yürütme token'ları, bir akıl yürütme modelinin görünür yanıtını yazmadan önce problemi çözmek için ürettiği token'lardır. Genellikle çıktı token'ı olarak faturalandırılır, çıktı sınırına ve bağlam penceresine sayılır; API tarafından gizlenebilir veya özetlenebilir. - [İstek zaman aşımı (request timeout)](https://broadrouter.com/tr/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): İstek zaman aşımı (request timeout), bir HTTP istemcisinin isteği bırakmadan önce sunucuyu bekleyeceği en uzun süredir. İstemciler bunu genellikle bağlantıyı açmak için bir bağlantı zaman aşımına ve veri beklemek için bir okuma zaman aşımına ayırır. LLM çağrıları cömert okuma zaman aşımları gerektirir, çünkü uzun yanıtlar zaman alır. - [Yeniden sıralama (reranking)](https://broadrouter.com/tr/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Yeniden sıralama (reranking), daha isabetli ama daha yavaş bir modelin, genellikle bir cross-encoder'ın, ilk aramadan gelen her sorgu ve aday belge çiftini puanlayıp yeniden sıraladığı ikinci erişim aşamasıdır; böylece dil modeline yalnızca en alakalı birkaç sonuç gönderilir. - [Sağdan sola (RTL) metin](https://broadrouter.com/tr/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): Sağdan sola (RTL) metin, Arapça veya İbranice gibi sağdan sola gösterilen yazıdır. Bilgisayarlar bu metni mantıksal okuma sırasıyla saklar; RTL metin soldan sağa metin, rakamlar, model adları veya kodla karıştığında gösterim sırasına Unicode Çift Yönlü Algoritması (bidi) karar verir. - [Örnekleme hızı (sample rate)](https://broadrouter.com/tr/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): Örnekleme hızı (sample rate), saniyede yakalanan ses örneği sayısıdır ve hertz (Hz) ile ölçülür. Konuşma tanıma modelleri genellikle 16.000 Hz (16 kHz) mono ses bekler. Örnekleme hızı, bit derinliği ve kanal sayısıyla birlikte sıkıştırılmamış bir ses dosyasının dakika başına boyutunu belirler. - [Seed parametresi (rastgelelik tohumu)](https://broadrouter.com/tr/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): Seed parametresi, aynı istek tekrarlandığında API'nin aynı token'ları örneklemeye çalışması için istekle gönderdiğiniz bir tam sayıdır. OpenAI bunu en iyi çaba (best effort) olarak tanımlar: determinizm garanti edilmez ve system_fingerprint alanı arka uç yapılandırmasının ne zaman değiştiğini gösterir. - [Sunucusuz çıkarım (serverless inference)](https://broadrouter.com/tr/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): Sunucusuz çıkarım (serverless inference), platformun bir uç noktanın arkasındaki hesaplama kaynağını otomatik olarak başlattığı, ölçeklediği ve durdurduğu, boştayken çoğu zaman sıfıra indirdiği bir model çalıştırma yöntemidir. Ayrılmış makineler yerine kullanım faturalanır. Bedeli olası soğuk başlatma gecikmesidir. - [SLA (hizmet seviyesi sözleşmesi)](https://broadrouter.com/tr/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): SLA (hizmet seviyesi sözleşmesi), bir sağlayıcının ölçülebilir bir hizmet seviyesine, genellikle aylık erişilebilirlik yüzdesine verdiği sözleşmesel taahhüttür; hedef tutmazsa tanımlı bir telafi, çoğunlukla hizmet kredisi uygulanır. SLA, neyin kesinti sayıldığını, nelerin hariç tutulduğunu ve talebin nasıl yapılacağını tanımlar. - [Spekülatif çözümleme (speculative decoding)](https://broadrouter.com/tr/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): Spekülatif çözümleme (speculative decoding), küçük ve hızlı bir taslak modelin sonraki birkaç token'ı önerdiği, büyük hedef modelin bunları tek bir ileri geçişte doğrulayıp kabul edilenleri tuttuğu bir çıkarım tekniğidir. Çıktının dağılımını değiştirmeden üretimi hızlandırır. - [Konuşmadan metne (speech-to-text)](https://broadrouter.com/tr/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): Konuşmadan metne (speech-to-text), otomatik konuşma tanıma (ASR) olarak da bilinir; konuşulan sesi yazılı metne dönüştüren teknolojidir. Bir konuşmadan metne API'si ses dosyası veya akışı alıp transkript döndürür ve genellikle ses dakikası başına fiyatlandırılır. Doğruluk kelime hata oranıyla ölçülür; aksana, lehçeye ve kayıt kalitesine göre değişir. - [Harcama limiti (spend limit)](https://broadrouter.com/tr/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): Harcama limiti (spend limit), bir API hesabının, projesinin veya anahtarının bir dönemde, genellikle bir ayda, harcayabileceği tutara konan tavandır. Kullanım bu tavana ulaştığında sağlayıcı, dönem sıfırlanana veya limit yükseltilene kadar yeni istekleri çoğu zaman HTTP 429 veya 400 hatasıyla reddeder. - [SRT altyazı](https://broadrouter.com/tr/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip), düz metin bir altyazı formatıdır. Her altyazı bir sıra numarası, hh:mm:ss,mmm --> hh:mm:ss,mmm biçiminde yazılmış başlangıç ve bitiş zamanı, bir veya daha fazla metin satırı ve bir boş satırdan oluşur. WebVTT, HTML videoda kullanılan benzer web formatıdır. - [Server-sent events (SSE, sunucu gönderimli olaylar)](https://broadrouter.com/tr/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Server-sent events (SSE), bir sunucunun text/event-stream içerik türüne sahip, uzun süre açık kalan tek bir HTTP yanıtı üzerinden istemciye metin olayları akışı göndermesini sağlayan bir web standardıdır. Her olay, data: gibi bir veya daha fazla alan satırından oluşur ve boş bir satırla biter. - [Durdurma dizileri (stop sequences)](https://broadrouter.com/tr/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): Durdurma dizileri (stop sequences), istekle birlikte gönderdiğiniz ve model bunlardan birini ürettiği anda üretimi durduran metinlerdir. OpenAI bu alana stop der ve en fazla 4 dizi kabul eder; Anthropic ise stop_sequences adını kullanır. Eşleşen metin normalde döndürülen çıktıya dahil edilmez. - [Akış (streaming)](https://broadrouter.com/tr/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): Akış (streaming), bir dil modeli yanıtını sonda tek parça halinde değil, hâlâ üretilirken küçük parçalar halinde döndürme yöntemidir. OpenAI uyumlu API'ler, istekte stream değeri true olduğunda server-sent events (SSE) üzerinden akış yapar. - [Yapılandırılmış çıktı (structured output)](https://broadrouter.com/tr/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): Yapılandırılmış çıktı, modelin yanıtını geliştiricinin belirlediği bir biçimle, genellikle bir JSON Schema ile sınırlayan bir LLM API özelliğidir; böylece kod yanıtı güvenilir şekilde ayrıştırabilir. Geçerli JSON'u garanti eden ama belirli bir yapıyı garanti etmeyen JSON modundan daha katıdır. - [Alt ajan (subagent)](https://broadrouter.com/tr/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): Alt ajan (subagent), ana ajanın odaklı tek bir görevi yürütmek için başlattığı ayrı bir ajandır. Kendi bağlam penceresi, sistem istemi ve araç erişimiyle çalışır, işi yapar ve ana ajana yalnızca bir sonuç veya özet döndürür; böylece ana konuşma küçük kalır. - [Alt işleyen (subprocessor)](https://broadrouter.com/tr/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): Alt işleyen (subprocessor), bir veri işleyenin veri sorumlusu adına kişisel veri işlemek için görevlendirdiği üçüncü taraftır. Bir LLM API'si için GPU barındırıcıları, iş ortağı çıkarım sağlayıcıları ve yedek kapasite alt işleyen olabilir. GDPR Madde 28'e göre bunlar veri sorumlusunun yetkisini ve aynı veri koruma yükümlülüklerini gerektirir. - [Sistem istemi (system prompt)](https://broadrouter.com/tr/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): Sistem istemi (system prompt), bir uygulamanın konuşmadan önce dil modeline gönderdiği talimat metnidir; genellikle system veya developer rolündeki bir mesajdır. Modelin rolünü, kurallarını, üslubunu ve çıktı biçimini belirler; model, çelişen kullanıcı mesajlarına karşı buna uymak üzere eğitilir. - [Tâ merbûta (taa marbuta, ة)](https://broadrouter.com/tr/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): Tâ merbûta (ة, U+0629), yalnızca kelime sonunda görülen ve çoğu zaman dişil isim ve sıfatları işaretleyen bir Arapça harftir. İki noktalı he (ه, U+0647) harfine benzediği için yazılan metinde ikisi karışır; birçok arama normalleştiricisi ة'yi ه'ye çevirir. - [Sıcaklık (temperature)](https://broadrouter.com/tr/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): Sıcaklık (temperature), bir dil modelinin çıktısının ne kadar rastgele olacağını belirleyen bir örnekleme parametresidir. Model, token puanlarını olasılığa çevirmeden önce sıcaklığa böler: düşük değerlerde en olası token baskın çıkar, yüksek değerler dağılımı düzleştirir ve daha az olası token'lar daha sık seçilir. - [Metinden sese (text-to-speech, TTS)](https://broadrouter.com/tr/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): Metinden sese (text-to-speech, TTS) veya ses sentezi, yazılı metni konuşma sesine dönüştürür. Bir TTS API'si metni, bir sesi ve MP3 veya WAV gibi bir çıktı formatını alır ve üretilmiş konuşmayı döndürür. Sesi metne dönüştüren konuşmadan metnenin tersidir. - [İşlem hacmi (throughput)](https://broadrouter.com/tr/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): İşlem hacmi (throughput), LLM sunumunda bir sistemin tüm eşzamanlı isteklerde birim zamanda tamamladığı toplam iş miktarıdır; genellikle saniyedeki çıktı token'ı veya saniyedeki istek olarak ölçülür. Tek bir isteğin token'larını ne kadar hızlı aldığını gösteren akış başına hızdan farklıdır. - [İlk token süresi (TTFT)](https://broadrouter.com/tr/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): İlk token süresi (time to first token, TTFT), bir dil modeline istek gönderildiği andan ilk çıktı token'ının geldiği ana kadar geçen süredir. Ağ süresini, sunucudaki kuyruk beklemesini ve modelin tüm istemi okuduğu ön doldurma (prefill) adımını içerir; bu yüzden istem uzunluğu ve yükle birlikte artar. - [Zaman damgaları (konuşmadan metne)](https://broadrouter.com/tr/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): Konuşmadan metnede zaman damgaları, transkriptin parçalarına eklenen saniye cinsinden başlangıç ve bitiş zamanlarıdır. Segment zaman damgaları bir ifadeyi veya cümleyi, kelime zaman damgaları her kelimeyi kapsar. Altyazılar, kayıt içinde arama ve metni konuşmacılarla hizalamak için gereklidir. - [Token](https://broadrouter.com/tr/glossary/token/) (translation of https://broadrouter.com/glossary/token/): Token, bir dil modelinin okuduğu ve yazdığı metin birimidir: modelin tokenizer'ının ürettiği bir kelime, kelime parçası, noktalama işareti veya bayt dizisi. LLM API'leri kullanımı token cinsinden sayar, sınırlar ve genellikle faturalandırır. - [Tokenizer](https://broadrouter.com/tr/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): Tokenizer, metni bir dil modelinin okuduğu tam sayı token kimlikleri dizisine dönüştüren ve üretilen kimlikleri tekrar metne çeviren bileşendir. Her model kendi tokenizer'ıyla gelir; bu yüzden aynı metin farklı modellerde farklı sayıda token üretebilir. - [Saniyedeki token (tokens per second, TPS)](https://broadrouter.com/tr/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): Saniyedeki token (TPS), bir dil modelinin çıktı token'larını üretme hızıdır ve genellikle tek bir istek için ilk token geldikten sonra ölçülür. Başlayan bir yanıtın ne kadar hızlı akacağını belirler. Tüm eşzamanlı isteklerde toplandığında sistem işlem hacmine dönüşür; ilk token'dan önceki bekleme ayrıca ölçülür. - [Araç çağırma (tool calling)](https://broadrouter.com/tr/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): Araç çağırma (tool calling), fonksiyon çağırma (function calling) olarak da bilinir; modelin, geliştiricinin tanımladığı bir fonksiyonu çalıştırmak için JSON argümanlı yapılandırılmış bir istek döndürdüğü bir LLM API özelliğidir. Fonksiyonu uygulama çalıştırır ve sonucu geri gönderir; model hiçbir zaman kodu kendisi çalıştırmaz. - [tool_choice](https://broadrouter.com/tr/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice, araç kullanımını kontrol eden bir istek parametresidir. OpenAI'nin Chat Completions API'sinde auto (karar modelde), required (bir veya daha fazla araç çağırmalı), none (araç yok) ya da zorlanacak adlandırılmış bir fonksiyon olabilir. Anthropic'teki karşılıkları auto, any, tool ve none seçenekleridir. - [Top-p (nucleus sampling, çekirdek örnekleme)](https://broadrouter.com/tr/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Top-p, diğer adıyla nucleus sampling (çekirdek örnekleme), bir dil modelinin sonraki token'ı yalnızca toplam olasılığı 0,9 gibi bir p eşiğine ulaşan en olası token'ların en küçük kümesinden seçtiği ve olasılığı düşük kuyruğu attığı bir kod çözme yöntemidir. - [Unicode normalizasyonu (Unicode normalization)](https://broadrouter.com/tr/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Unicode normalizasyonu, aynı görünen dizelerin aynı kod noktalarını kullanması için metni standart bir biçime dönüştürür. Dört biçim NFD, NFC, NFKD ve NFKC'dir; K biçimleri ayrıca Arapça sunum biçimleri ve bitişik harfler (ligatür) gibi uyumluluk karakterlerini sıradan harflere katlar. - [Ses etkinliği algılama (VAD)](https://broadrouter.com/tr/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): Ses etkinliği algılama (voice activity detection, VAD), sesin kısa karelerini konuşma veya konuşma dışı olarak sınıflandıran bir sinyal işleme adımıdır. Konuşma sistemleri bunu sessizliği atlamak, uzun kayıtları duraklamalardan bölmek, konuşmacının sözünü bitirdiğini anlamak ve arka plan gürültüsünü yazıya dökmemek için kullanır. - [Vektör veritabanı (vector database)](https://broadrouter.com/tr/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): Vektör veritabanı, embedding vektörlerini çoğu zaman kaynak metin ve meta verileriyle birlikte tutan ve en yakın komşu sorgularını yanıtlayan bir veri deposudur: bir sorgu vektörü verildiğinde, genellikle kosinüs mesafesi, Öklid mesafesi veya iç çarpıma göre ona en çok benzeyen kayıtlı vektörleri döndürür. - [Görsel-dil modeli (VLM)](https://broadrouter.com/tr/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): Görsel-dil modeli (vision-language model, VLM), girdi olarak görselleri ve metni birlikte alıp metin üreten bir modeldir; böylece resimleri betimleyebilir, onlarla ilgili soruları yanıtlayabilir, belgeleri, grafikleri veya ekran görüntülerini okuyabilir. VLM'lere görselden metne (image-text-to-text) veya çok kipli (multimodal) model de denir. - [vLLM](https://broadrouter.com/tr/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM, ilk olarak UC Berkeley'deki Sky Computing Lab'de geliştirilen, LLM çıkarımı ve sunumu (serving) için açık kaynak bir kütüphanedir. KV önbelleği belleğini PagedAttention ile yönetir, gelen istekleri sürekli olarak toplu işler, birçok kuantizasyon formatını destekler ve OpenAI uyumlu bir HTTP sunucusu sunar. - [Sözlük boyutu (vocabulary size)](https://broadrouter.com/tr/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): Sözlük boyutu (vocabulary size), bir dil modelinin tokenizer'ındaki, her birinin kendi kimliği (ID) ve embedding'i olan farklı token'ların sayısıdır. Daha büyük bir sözlük yaygın metni daha az token'la kodlayabilir, ancak embedding ve çıktı katmanlarını büyütür ve yalnızca sözlüğün eğitildiği türden metinde işe yarar. - [WAV](https://broadrouter.com/tr/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format veya WAVE), Microsoft ve IBM'in RIFF formatı üzerine kurulu bir ses kapsayıcısıdır. Genellikle sıkıştırılmamış doğrusal PCM örnekleri taşır; bu yüzden boyutu öngörülebilirdir: saniye başına örnekleme hızı çarpı örnek başına bayt çarpı kanal sayısı. - [Kelime hata oranı (WER)](https://broadrouter.com/tr/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): Kelime hata oranı (word error rate, WER), konuşma tanımada standart doğruluk ölçüsüdür: bir transkripti doğru referansa dönüştürmek için gereken kelime değiştirme, silme ve ekleme sayısının referanstaki kelime sayısına bölümüdür. Düşük olan daha iyidir. - [Sıfır veri saklama (ZDR)](https://broadrouter.com/tr/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): Sıfır veri saklama (ZDR), bir yapay zeka sağlayıcısının istemleri ve çıktıları yalnızca isteği karşılamak için işlediği ve bu içeriği sonrasında loglarda, önbelleklerde veya eğitim verilerinde saklamadığı bir veri işleme politikasıdır. Token sayıları ve süreler gibi içerik barındırmayan kayıtlar genellikle yine tutulur. ### Urdu (ur, اردو): 294 translated pages - [Articles in Urdu](https://broadrouter.com/ur/blog/) - [اکتوبر 2026 میں LLM API قیمتوں میں تبدیلیاں: 3 اضافے، کٹوتی کوئی نہیں](https://broadrouter.com/ur/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): اکتوبر 2026 میں 7 میں سے 5 LLM API فراہم کنندگان کے قیمتوں کے صفحات پر 23 تبدیلیاں: 3 اضافے، کوئی کٹوتی نہیں، 9 نئے ماڈلز کی قیمت آئی، 11 ہٹے۔ آرکائیو لنکس سمیت۔ - [AI کوڈنگ ایجنٹ کے ایک رن کی لاگت: مراحل، کانٹیکسٹ اور کیشنگ](https://broadrouter.com/ur/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): کوڈنگ ایجنٹ کا رن جتنا لمبا چلتا ہے، ہر مرحلہ اتنا مہنگا کیوں پڑتا ہے، کیشنگ کل لاگت کیسے بدلتی ہے، اور مقررہ سیشن کب جیتتا ہے۔ ٹوکنز کا مکمل حساب۔ - [LLM سے عربی متن پر اعراب (تشکیل) لگانا: پرامپٹ، جانچ، DER](https://broadrouter.com/ur/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): LLM API سے عربی اعراب (tashkeel) بحال کریں: ہر حرف محفوظ رکھنے والا پرامپٹ، تبدیلیاں پکڑنے والی جانچ، DER اور WER اسکورنگ، اور CATT کب منتخب کریں۔ - [LLM سے عربی گرامر اور املا کی تصحیح: ایک عملی رہنما](https://broadrouter.com/ur/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): LLM API سے عربی املا اور گرامر درست کریں: کم سے کم تبدیلی والا پرامپٹ، دوبارہ لکھنے سے حفاظت، غلطی کی قسم کے لحاظ سے edits اور M2 جیسی اسکورنگ۔ - [عربی ریکارڈنگز سے میٹنگ منٹس: ایک API پائپ لائن](https://broadrouter.com/ur/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): عربی میٹنگ ریکارڈنگز کو فیصلوں اور ایکشن آئٹمز والے منٹس میں بدلیں: ٹرانسکرائب کریں، صاف کریں، LLM سے JSON میں خلاصہ لیں، اور نام و تاریخیں چیک کریں۔ - [LLM سے عربی named entity extraction: JSON، تاریخیں اور جانچ](https://broadrouter.com/ur/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): LLM سے عربی نام، ادارے، مقامات اور ہجری تاریخیں JSON میں نکالیں: ایک پرامپٹ، اعراب نظرانداز کرنے والی span جانچ، تاریخ کی تبدیلی اور اسکورنگ اسکرپٹ۔ - [عربی OCR API کا موازنہ: Google، Azure، Textract، Mistral، LLMs](https://broadrouter.com/ur/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): کون سے OCR API سرکاری طور پر چھپی اور ہاتھ سے لکھی عربی سپورٹ کرتے ہیں، ہر ایک کیا لوٹاتا ہے اور فی 1,000 صفحات کتنا لیتا ہے، vision LLM API کے مقابلے میں۔ - [speech-to-text API سے عربی سب ٹائٹلز (SRT): آپ کو کیا چاہیے](https://broadrouter.com/ur/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): SRT فائل کو کیا چاہیے، کون سے speech-to-text APIs عربی کے لیے ٹائم اسٹیمپس یا SRT دیتے ہیں، اور صرف متن والے API سے سب ٹائٹلز بنانے والی ٹیسٹ شدہ پائپ لائن۔ - [LLM سے عربی متن کی classification: sentiment، intent، triage](https://broadrouter.com/ur/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): LLM API سے عربی ریویوز، پیغامات اور سپورٹ ٹکٹس classify کریں: لیبل ڈیزائن، لہجوں سے باخبر پرامپٹ، JSON آؤٹ پٹ، macro F1 ٹیسٹنگ اور بڑے والیوم پر لاگت۔ - [سرچ اور RAG کے لیے عربی متن کی normalization: کیا بدلیں، کیا نہیں](https://broadrouter.com/ur/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): تشکیل، الف کی شکلیں، ta marbuta، ہمزہ، فارسی حروف، ہندسے اور Unicode: عربی normalization کے کون سے اصول سرچ میں مدد دیتے ہیں، کون نقصان، اور ٹیسٹ شدہ Python۔ - [LLM سے عربی متن کا خلاصہ: پرامپٹس، لمبائی اور اسکورنگ](https://broadrouter.com/ur/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): LLM API سے عربی مضامین، رپورٹس اور لہجے والے پیغامات کا خلاصہ: پرامپٹ پیٹرنز، الفاظ کی گنتی، نمبروں کی جانچ، اور ڈیفالٹ ROUGE عربی کو 0 کیوں دیتا ہے۔ - [Artificial Analysis کے متبادل: 2026 میں LLMs کا موازنہ کہاں کریں](https://broadrouter.com/ur/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Artificial Analysis کے متبادلات کا موازنہ: Arena، Epoch AI، Vals AI، OpenRouter رینکنگ، Hugging Face اور قیمتوں کا ڈیٹا، ہر ایک کیا ناپتا ہے اور کب کام آتا ہے۔ - [BroadNet کا GPU ہارڈویئر اور سروِنگ کا طریقہ، BroadNet کی زبانی](https://broadrouter.com/ur/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): BroadNet Inference کے پیچھے کے ہارڈویئر کے بارے میں BroadNet کیا کہتا ہے: B300 سے کلائنٹ appliances تک تین GPU درجے، فی GPU throughput، اور ٹیسٹس سے ملنے والے سبق۔ - [LLM API جوابات کیش کر کے لاگت کم کریں: Python میں ایگزیکٹ میچ کیش](https://broadrouter.com/ur/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Python میں LLM API جوابات کے لیے ٹیسٹ شدہ ایگزیکٹ میچ کیش: کلید میں کیا جاتا ہے، کیا کبھی کیش نہ کریں، TTL، پرائیویسی اور بنانے سے پہلے ہٹ ریٹ کیسے ناپیں۔ - [اوپن ماڈلز کے ساتھ Claude Code: پراکسی، نیٹو اینڈپوائنٹس اور حدیں](https://broadrouter.com/ur/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code، Anthropic Messages API بولتا ہے۔ LiteLLM پراکسی یا نیٹو اینڈپوائنٹ سے اسے اوپن ماڈلز پر کیسے چلائیں، اور راستے میں کیا کیا ٹوٹتا ہے۔ - [دستاویزی AI ورک فلو میں انسانی جائزے کو شامل کریں](https://broadrouter.com/ur/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): دستاویزی AI پائپ لائن کا انسانی جائزے والا مرحلہ کیسے بنائیں: توثیقی اصول، کیا انسان تک بھیجیں، رینڈم آڈٹ سیمپل، آڈٹ ٹریل اور جائزے کی شرح۔ - [مصر کا ڈیٹا تحفظ کا قانون (Law 151 of 2020) اور LLM API](https://broadrouter.com/ur/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): مصریوں کا ذاتی ڈیٹا بیرونِ ملک LLM یا speech-to-text API کو بھیجنا: Law 151 of 2020 اور Decision 816 of 2025 منتقلی، لائسنس اور بریچ پر کیا کہتے ہیں۔ - [فائن ٹیوننگ سے پہلے: حل کرنے والا مسئلہ طے کریں](https://broadrouter.com/ur/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): کب پرامپٹنگ، RAG یا evaluation مسئلہ حل کر دیتے ہیں اور فائن ٹیوننگ ضروری نہیں، کیسا ٹریننگ ڈیٹا چاہیے، اور فائن ٹیون مانگنے سے پہلے اسے کیسے جانچیں۔ - [LLM کا غلط JSON کیسے ٹھیک کریں: وجوہات، مرمت، دوبارہ کوشش](https://broadrouter.com/ur/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): LLM ٹوٹا JSON کیوں لوٹاتے ہیں (کوڈ فینس، نثر، max_tokens پر کٹاؤ، ٹریلنگ کوما) اور ایک ٹیسٹ شدہ Python فنکشن جو محفوظ چیزیں ٹھیک کرے، باقی رد کرے۔ - [LLM API کی قیمتیں کتنی بار بدلتی ہیں؟ آرکائیو صفحات کے 3 سال](https://broadrouter.com/ur/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): ہم نے 2023 سے 2026 تک 11 LLM API فراہم کنندگان کے 308 آرکائیو شدہ قیمتوں کے صفحات پڑھے۔ 80% ماڈلز کی قیمت کبھی نہیں بدلی؛ 75% تبدیلیاں DeepInfra نے کیں۔ - [LLM API کی لاگت کیسے کم کریں: 10 مراحل کی چیک لسٹ، درست ترتیب میں](https://broadrouter.com/ur/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): LLM API کا بل درست ترتیب سے کم کریں: پہلے پیمائش، پھر پری فکس کیش، پرامپٹ مختصر، آؤٹ پٹ اور ریزننگ محدود، مناسب ماڈل، بیچ، اور مقررہ قیمت کا حساب۔ - [LLM API فیل اوور: اپنے کوڈ میں دوسرے فراہم کنندہ پر فال بیک](https://broadrouter.com/ur/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): OpenAI سے مطابقت رکھنے والے LLM فراہم کنندگان کے بیچ فیل اوور کیسے کریں: کن ایررز پر سوئچ کریں، سرکٹ بریکر، ماڈل ID میپنگ، اسٹریمنگ، لاگت اور پرائیویسی کی برابری۔ - [LLM API قیمتوں میں کٹوتیاں 2023–2026: سب سے بڑی کمی، ذرائع کے ساتھ](https://broadrouter.com/ur/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): 2023–2026 میں LLM API کی سب سے بڑی قیمت کٹوتیاں، پہلے اور بعد کے Internet Archive لنکس کے ساتھ: o3، Codestral، Gemini 1.5، اور کم زیر بحث آنے والے اضافے بھی۔ - [کیا LLM API کی قیمتیں صرف کم ہوتی ہیں؟ 52 اضافے اور بجٹ کا طریقہ](https://broadrouter.com/ur/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): LLM API کی قیمتیں بڑھتی بھی ہیں۔ 2024 سے 2026 کے آرکائیو شدہ قیمتوں کے صفحات میں ایک ہی ماڈل پر 52 اضافے ملے، میڈین 40%۔ تمام ذرائع اور بجٹ کا اصول۔ - [فراہم کنندہ کے لحاظ سے LLM API ریٹ لمٹس: RPM، TPM اور کنکرنسی](https://broadrouter.com/ur/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): 13 فراہم کنندگان کی ابتدائی ٹائر LLM API ریٹ لمٹس، ان کی دستاویزات سے: OpenAI، Anthropic، Gemini، Groq، DeepSeek، Fireworks وغیرہ، اور RPM و TPM کا مطلب۔ - [LLM API SLA موازنہ: Azure، Bedrock، Google، OpenAI کے وعدے](https://broadrouter.com/ur/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): ہر LLM API فراہم کنندہ تحریری طور پر کیا وعدہ کرتا ہے: اپ ٹائم فیصد، ڈاؤن ٹائم کسے مانا جاتا ہے، سروس کریڈٹس اور لیٹنسی اہداف، SLA صفحات سے پڑھ کر۔ - [Batch API کے بغیر LLM بیچ پروسیسنگ: دوبارہ شروع ہونے والا runner](https://broadrouter.com/ur/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Batch API کے بغیر chat-completions API سے 10,000 سے 1,000,000 پرامپٹس چلائیں: بیک وقت درخواستوں کی حد، رفتار کا کنٹرول، دوبارہ کوشش، چیک پوائنٹس اور وقت کا حساب۔ - [Safety Map: 39 LLMs حساس ڈیٹا کو کیسے سنبھالتے ہیں](https://broadrouter.com/ur/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet نے 14 لیبز کے 39 LLMs کو credentials، PII، صحت اور مالی ڈیٹا پر پرکھا: کیا منتقل ہوا، کیا نہیں، اور tool calls اور ریزننگ سے ہونے والے لیکس۔ - [حساس ڈیٹا کی حفاظت کے لیے LLM کو کیسے پرکھیں: BroadNet کا طریقہ](https://broadrouter.com/ur/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): ڈیپلائے کرنے سے پہلے LLM، credentials، PII، صحت یا مالی ڈیٹا لیک کرتا ہے یا نہیں، یہ پرکھنے کا BroadNet کا شائع شدہ طریقہ: کیٹیگریز، leak بمقابلہ miss، 3+ رنز۔ - [OpenAI API سے منتقلی: Python، Node اور cURL میں کوڈ واک تھرو](https://broadrouter.com/ur/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): OpenAI سے OpenAI سے مطابقت رکھنے والے API پر جانے کے لیے کوڈ کی تبدیلیاں: کلائنٹ سیٹ اپ، Responses سے Chat Completions، ٹولز، اسٹریمنگ، embeddings اور اسموک ٹیسٹ۔ - [BroadNet نے کون سے ماڈلز پرکھ کر مسترد کیے، اور کیوں](https://broadrouter.com/ur/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet اپنے ٹیسٹ کیے اور مسترد ماڈلز وجوہات سمیت شائع کرتا ہے: کریڈنشل سیفٹی جانچ میں ناکام چیٹ ماڈلز، اور ہارنے والے اسپیچ، OCR اور ایمبیڈنگ ماڈلز۔ - [کئی زبانوں کے لیے AI پروڈکٹ کیسے ڈیزائن کریں](https://broadrouter.com/ur/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): عربی، انگریزی اور ہسپانوی صارفین کے لیے AI پروڈکٹ کا ڈیزائن: UI لوکیل بمقابلہ ماڈل کی زبان، RTL اور ملا جلا متن، اعداد، تاریخیں، زبان کی شناخت اور لاگت۔ - [اوپن ویٹ بمقابلہ کلوزڈ ماڈل API لاگت: قیمتوں کا ڈیٹا کیا بتاتا ہے](https://broadrouter.com/ur/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): کیا API کے ذریعے اوپن ویٹ ماڈلز GPT، Claude یا Gemini سے سستے ہیں؟ فراہم کنندگان کے اپنے صفحات سے فی ٹوکن قیمتیں، ہوسٹس کا فرق اور فی صارف لاگت کی مثال۔ - [OpenAI، Anthropic اور Google: API قیمتوں کی تاریخ، 2023–2026](https://broadrouter.com/ur/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): آرکائیو شدہ قیمتوں کے صفحات سے GPT، Claude اور Gemini API کی قیمتیں ساتھ ساتھ، 2023 سے 2026: کون سی لائن سستی ہوئی، کون سی مہنگی، ہر ایک آرکائیو لنک کے ساتھ۔ - [OpenAI-compatible API کے base URL: 14 فراہم کنندگان، جانچ شدہ](https://broadrouter.com/ur/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): 14 OpenAI-compatible LLM API کا base_url، ہر ایک فراہم کنندہ کی اپنی دستاویزات سے لیا گیا، ان کے بتائے اینڈپوائنٹس اور وہ باریکیاں جو سیدھی تبدیلی کو توڑ دیتی ہیں۔ - [سستے ماڈل پر OpenClaw اور Hermes Agent: سیٹ اپ اور ٹوکن کا حساب](https://broadrouter.com/ur/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): OpenClaw یا Hermes Agent کو OpenAI-compatible اینڈپوائنٹ سے جوڑیں، سمجھیں کہ شیڈولڈ چیکس ٹوکن کیوں کھاتے ہیں، اور ہلکے کانٹیکسٹ و سستے ماڈلز سے بل کم کریں۔ - [پرائیویٹ اینڈپوائنٹ کا آغاز واضح ورک لوڈ سے ہوتا ہے](https://broadrouter.com/ur/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): پرائیویٹ LLM اینڈپوائنٹ مانگنے سے پہلے کیا طے کریں: ورک لوڈ، فی سیکنڈ ٹوکنز، بیک وقت درخواستیں، کانٹیکسٹ، ریجن، لائسنس اور بجٹ، پھر اسے ٹیسٹ کیسے کریں۔ - [کوڈنگ ایجنٹس میں prompt injection: دفاع جو واقعی کام کرتے ہیں](https://broadrouter.com/ur/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): prompt injection کوڈنگ ایجنٹ تک کہاں سے پہنچتا ہے، اور وینڈرز اور محققین کون سے تہہ دار دفاع بتاتے ہیں: سینڈ باکس، نیٹ ورک پر پابندی، allow-lists اور منظوری۔ - [RAG چنک سائز اور اوورلیپ: بازیافت اور لاگت پر ان کا اثر](https://broadrouter.com/ur/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): چنک سائز اور اوورلیپ RAG ریکال، انڈیکس سائز، ایمبیڈنگ لاگت اور ہر کوئری میں LLM کو جانے والے ٹوکن کیسے بدلتے ہیں، شائع شدہ نتائج اور حساب کے ساتھ۔ - [ریزننگ ماڈلز کے لیے max_tokens کیسے طے کریں: بجٹ کا طریقہ](https://broadrouter.com/ur/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): کم max_tokens پر ریزننگ ماڈلز خالی جواب کیوں دیتے ہیں، اور ناپے گئے percentiles، کانٹیکسٹ ونڈو اور بڑھتے retry سے بجٹ کیسے طے کریں۔ - [براؤزر میں آڈیو ریکارڈ کریں اور speech-to-text API کو بھیجیں](https://broadrouter.com/ur/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): MediaRecorder سے مائیکروفون ریکارڈ کریں، WebM یا Safari کا MP4 16 kHz WAV میں بدلیں، اور کلید رکھنے والے چھوٹے سرور کے ذریعے ٹرانسکرپشن API کو بھیجیں۔ - [پرامپٹ ٹوکن کیسے کم کریں: سسٹم پرامپٹ، ٹولز، ڈیٹا کی ناپی گئی کٹوتی](https://broadrouter.com/ur/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): سسٹم پرامپٹس، few-shot مثالیں، ٹول اسکیما، ڈیٹا فارمیٹس، چیٹ ہسٹری اور عربی اعراب کم کرنے سے ٹوکن کی ناپی گئی بچت، اصل ٹوکنائزرز سے گنی گئی۔ - [ایک ماڈل، مختلف ہوسٹ: LLM API کی قیمتیں کتنی مختلف ہوتی ہیں](https://broadrouter.com/ur/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): کئی ہوسٹس پر بکنے والے 119 ماڈلز میں سب سے مہنگی فہرستی قیمت سب سے سستی کی میڈین 1.76 گنا ہے۔ 41 فراہم کنندگان کے اپنے قیمتوں کے صفحات سے ڈیٹا، نومبر 2026۔ - [سعودی اور UAE PDPL اور LLM API: سرحد پار ڈیٹا منتقلی کی چیک لسٹ](https://broadrouter.com/ur/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): سعودی عرب یا UAE سے LLM API کو ذاتی ڈیٹا بھیجنا: PDPL کے منتقلی کے قواعد کیا کہتے ہیں، فراہم کنندہ سے کیا پوچھیں، اور ملک کے اندر inference کب ضروری ہے۔ - [متن، آواز یا وژن؟ پہلے ٹاسک دیکھیں](https://broadrouter.com/ur/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): ٹاسک کے حساب سے متنی چیٹ، تصویری ان پٹ یا آڈیو ٹرانسکرپشن چنیں، BroadNet کے دستاویزی ماڈلز اور حدوں کے ساتھ، اور وہ بھی جو وہ نہیں دیتا، جیسے TTS۔ - [25 MB سے بڑا لمبا آڈیو ٹرانسکرائب کریں: تقسیم، ارسال، جوڑ](https://broadrouter.com/ur/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): speech-to-text API کی فائل حد سے بڑی ریکارڈنگ کیسے ٹرانسکرائب کریں: فائل چھوٹی کریں، ffmpeg سے خاموشی پر تقسیم کریں، ٹکڑے متوازی بھیجیں اور متن جوڑیں۔ - [LLM سے ایپ strings کا عربی ترجمہ: placeholders اور plurals](https://broadrouter.com/ur/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): LLM API سے i18n JSON فائلوں کا عربی ترجمہ کریں: placeholders منجمد کریں، عربی کی چھ plural شکلیں بنوائیں، CI میں ICU syntax جانچیں اور شامل کیے گئے نام isolate کریں۔ - [API سے عربی آڈیو کا انگریزی متن میں ترجمہ: ٹیسٹ شدہ پائپ لائن](https://broadrouter.com/ur/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): دو API کالز سے عربی ریکارڈنگز کو انگریزی متن بنائیں: language=ar سے ٹرانسکرائب کریں، پھر LLM سے ترجمہ، نمبروں کی جانچ، chunking، لاگت اور اسکورنگ کے ساتھ۔ - [Live، on request، research: ماڈل کیٹلاگ کو واضح طور پر کیسے پڑھیں](https://broadrouter.com/ur/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): کیسے پہچانیں کہ کیٹلاگ کے کون سے ماڈل آج کال ہو سکتے ہیں، کن کے لیے استفسار چاہیے اور کون صرف ٹیسٹ ہوئے، BroadNet کے کیٹلاگ اور /v1/models کی مثال سے۔ - [vLLM میں Uno: BroadNet کی سرونگ اسٹڈی نے کیا ناپا](https://broadrouter.com/ur/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet نے Uno self-drafting adapter کو vLLM میں شامل کیا اور H100 پر ناپا: ایک stream پر سادہ vLLM سے 2.58x تھروپٹ، 32 پر 1.74x۔ طریقہ، DOI اور حدیں۔ - [Gemma 4 26B پر Uno: BroadNet کی mixture-of-experts اسٹڈی](https://broadrouter.com/ur/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet نے mixture-of-experts ماڈل Gemma 4 26B A4B کے لیے پہلا Uno self-drafting adapter ٹرین کیا۔ کیا ناپا گیا، self-drafting کب فائدہ دیتا ہے، اور DOI۔ - [ایجنٹ ٹوکن بجٹ: ایجنٹ لوپ کے ٹوکن کہاں جاتے ہیں](https://broadrouter.com/ur/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): سسٹم پرامپٹ، ٹول اسکیما، فائلیں پڑھنا، ہسٹری اور ریزننگ: ایجنٹ لوپ ٹوکن کیسے خرچ کرتا ہے، usage ڈیٹا سے ہر حصہ کیسے ناپیں، اور پہلے کیا کم کریں۔ - [عربی ASR میں word error rate (WER) کیا ہے، آسان الفاظ میں](https://broadrouter.com/ur/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): word error rate کا مطلب کیا ہے، اسے کیسے نکالیں، اور عربی WER اعراب، املا کی نارملائزیشن اور لہجوں پر اتنا منحصر کیوں ہے۔ Python اور اصل نمبروں کے ساتھ۔ - [عربی لہجوں کی speech recognition: خلیجی، مصری، شامی، مغربی](https://broadrouter.com/ur/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): خلیجی، مصری، شامی اور مغربی عربی پر speech-to-text کیوں لڑکھڑاتا ہے، اور لہجوں کا ایسا ٹیسٹ سیٹ کیسے بنائیں جو دکھائے کہ فراہم کنندہ آپ کے آڈیو پر کیسا کرے گا۔ - [LLMs سے عربی–انگریزی ترجمہ: پرامپٹس، اصطلاحات، جائزہ](https://broadrouter.com/ur/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): LLM API سے عربی سے انگریزی اور واپس ترجمہ کریں: لہجہ برقرار رکھنے والے پرامپٹس، مقررہ اصطلاحات کی فہرست، نمبر اور نام، اور معیار کا جائزہ کیسے لیں۔ - [خلیجی، مصری اور شامی عربی: LLMs لہجوں کو کیسے سنبھالتے ہیں](https://broadrouter.com/ur/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): LLMs خلیجی، مصری اور شامی عربی کو کیسے سنبھالتے ہیں: کہاں ناکام ہوتے ہیں، لہجے کے لیے پرامپٹ اور ٹیسٹ کیسے کریں، اور Qwen3.8 27B پر BroadNet کی رپورٹ۔ - [عربی OCR بمقابلہ vision LLMs: عربی دستاویزات اور انوائسز پڑھنا](https://broadrouter.com/ur/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): عربی انوائسز، رسیدوں اور فارمز کے لیے کب روایتی OCR اور کب vision LLM: درستگی کے جال، JSON آؤٹ پٹ، فی صفحہ لاگت اور ایک validation پائپ لائن۔ - [عربی پرامپٹ انجینئرنگ: register، RTL، ہندسے اور ملا جلا متن](https://broadrouter.com/ur/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): ایسے پرامپٹس لکھیں جو LLM سے اچھی عربی نکلوائیں: register اور لہجہ چنیں، ہندسے اور تاریخیں قابو کریں، RTL فارمیٹنگ صاف رکھیں، عربی–انگریزی امتزاج سنبھالیں۔ - [عربی RAG: normalisation، chunking اور retrieval جو واقعی کام کرے](https://broadrouter.com/ur/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): عربی دستاویزات پر RAG: الف، یاء اور تاء مربوطہ normalise کریں، سمجھداری سے chunk کریں، کثیر لسانی embeddings چنیں اور keyword search شامل کریں۔ - [عربی speech-to-text API: 2026 میں درست API کیسے منتخب کریں](https://broadrouter.com/ur/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): عربی ٹرانسکرپشن API کیسے چنیں: word error rate، لہجے، فائل کی حدیں، فی آڈیو گھنٹہ قیمت اور پرائیویسی، ساتھ میں آپ کی ریکارڈنگز کے لیے ٹیسٹ پلان۔ - [عربی WhatsApp chatbot کا آرکیٹیکچر: وائس نوٹس سے جوابات تک](https://broadrouter.com/ur/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): عربی WhatsApp اسسٹنٹ کا کام کرتا آرکیٹیکچر: webhook، queue، وائس نوٹ ٹرانسکرپشن، LLM جواب، پالیسی کے اصول اور لاگت، Python کوڈ کے ساتھ۔ - [Arabizi اور LLMs: 3 اور 7 جیسے ہندسوں والی Franco-Arabic سنبھالنا](https://broadrouter.com/ur/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): Arabizi عربی کو لاطینی حروف میں لکھتی ہے، ع کے لیے 3 اور ح کے لیے 7 جیسے ہندسوں کے ساتھ۔ اسے کیسے پہچانیں، LLM سے عربی رسم الخط میں بدلیں، اور درست جواب دیں۔ - [Aider کے لیے بہترین LLM: ماڈل کیسے چنیں اور جوڑیں](https://broadrouter.com/ur/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Aider کے لیے بہترین LLM کیسے چنیں: ترمیمی فارمیٹ کی پابندی، کانٹیکسٹ سائز اور فی گھنٹہ لاگت، اپنی ریپو پر ایک گھنٹے کا ٹیسٹ، اور OPENAI_API_BASE کا درست سیٹ اپ۔ - [کوڈنگ ایجنٹس کے لیے بہترین اوپن ماڈل: ٹولز، کانٹیکسٹ اور لاگت](https://broadrouter.com/ur/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): آپ کے کوڈنگ ایجنٹ کے لیے GLM 5.3، DeepSeek V4.1 Flash، Qwen3.8 27B یا Gemma 4؟ ٹول کالنگ، کانٹیکسٹ، ریزننگ اور فی ٹاسک لاگت سے منتخب کریں، مکمل حساب کے ساتھ۔ - [کیش شدہ ان پٹ ٹوکن کیا ہیں: LLM API کا بل کیسے کم کریں](https://broadrouter.com/ur/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): کیش شدہ ان پٹ عام ان پٹ کی قیمت کے چھوٹے حصے پر بل ہوتا ہے۔ پرامپٹ کیشنگ کا بل، usage میں cached_tokens پڑھنا، اور پرامپٹ کو کیش کے قابل بنانا۔ - [سب سے سستا وژن LLM API: فی تصویر اور فی صفحہ لاگت کا حساب](https://broadrouter.com/ur/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): وژن LLM کال کی فی تصویر اور فی دستاویزی صفحہ لاگت: Gemma 4، Qwen3.8 27B اور DeepSeek V4.1 Flash کے ٹوکن کا حساب، اور DeepInfra کہاں سستا ہے۔ - [LLM API فراہم کنندہ کیسے چنیں: 20 سوالات کی چیک لسٹ](https://broadrouter.com/ur/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): LLM API چننے سے پہلے پوچھنے کے 20 سوالات: قیمت، حدیں، OpenAI مطابقت، پرائیویسی، ماڈلز، سپورٹ اور اپ ٹائم، ساتھ میں اسکورنگ شیٹ اور ٹیسٹ پلان۔ - [عربی کے لیے LLM API کیسے چنیں: کیا کیا ٹیسٹ کریں](https://broadrouter.com/ur/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): عربی ایپس کے لیے LLM کیسے چنیں: لہجوں کی سمجھ، تحریر کا معیار، ترجمہ، عربی دستاویزات اور تصاویر، لاگت، اور ایک دن میں بننے والا ٹیسٹ سیٹ۔ - [Cline بمقابلہ Roo Code بمقابلہ Aider: اوپن ماڈلز والی ٹیموں کے لیے](https://broadrouter.com/ur/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline، Roo Code (اب آرکائیو، Zoo Code کے طور پر جاری) اور Aider کا موازنہ، اوپن ماڈلز کے لیے اہم باتوں پر: ٹول کالز، کانٹیکسٹ، MCP، git اور منظوریاں۔ - [LLM درخواست بھیجنے سے پہلے ٹوکن کیسے گنیں](https://broadrouter.com/ur/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): LLM API کال سے پہلے زبان اور ماڈل کے حساب سے پرامپٹ ٹوکن کا اندازہ: ٹوکنائزر کیوں مختلف ہیں، پانچ زبانوں کی ناپی گئی مثالیں، اور محفوظ max_tokens مارجن۔ - [کوڈنگ ایجنٹ کی LLM لاگت کم کرنے کے 10 طریقے، اندازاً بچت کے ساتھ](https://broadrouter.com/ur/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Cline، Aider، opencode یا آپ کا اپنا کوڈنگ ایجنٹ ٹوکنز پر جو خرچ کرتا ہے، اسے کم کرنے کے دس ٹھوس طریقے۔ ہر ایک کے ساتھ بچت کا حساب اور اسے ناپنے کا طریقہ۔ - [DeepSeek API کے متبادل: DeepSeek ماڈلز کال کرنے کے 5 طریقے](https://broadrouter.com/ur/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): DeepSeek ماڈلز کو آفیشل API، ایگریگیٹر، کلاؤڈ کیٹلاگ، اپنے GPU یا BroadNet Inference سے کال کرنا: قیمت، پرائیویسی اور حدوں کے لحاظ سے فائدے اور نقصانات۔ - [DeepSeek V4.1 Flash API کی قیمت: فی ٹوکن، کیش شدہ اور پلانز](https://broadrouter.com/ur/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): DeepSeek V4.1 Flash API کی قیمت فی 1M ان پٹ، کیش شدہ اور آؤٹ پٹ ٹوکن، ماہانہ، ہفتہ وار اور گھنٹے والے مقررہ قیمت پلانز، اور DeepSeek، DeepInfra، Fireworks سے موازنہ۔ - [فیچر لانچ سے پہلے LLM API کی لاگت کا اندازہ کیسے لگائیں](https://broadrouter.com/ur/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): نئے AI فیچر کے لیے مرحلہ وار لاگت ماڈل: صارفین × ایکشنز × کالز × ٹوکن، کیش، ریزننگ، دوبارہ کوششوں، ریٹ لمٹس اور ایک حل شدہ ماہانہ جدول کے ساتھ۔ - [فراہم کنندہ بدلنے سے پہلے LLM کی جانچ کیسے کریں](https://broadrouter.com/ur/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): اصل ٹریفک سے ایک چھوٹا eval سیٹ بنائیں، کوڈ چیکس اور تعصب سے پاک LLM جج سے اسکور کریں، اور مائیگریٹ کرنے سے پہلے معیار، لاگت اور لیٹنسی کا موازنہ کریں۔ - [وژن LLM سے انوائس، رسید اور فارم کا ڈیٹا JSON میں نکالیں](https://broadrouter.com/ur/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): انوائسز، رسیدوں اور فارمز کے لیے پروڈکشن پائپ لائن: تصویر کی تیاری، JSON اسکیما والا پرامپٹ، Pydantic توثیق، جمع کی جانچ، retries اور انسانی جائزے کی قطار۔ - [مقررہ قیمت بمقابلہ فی ٹوکن ادائیگی والا LLM API: بریک ایون کا حساب](https://broadrouter.com/ur/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): مقررہ قیمت والا LLM پلان فی ٹوکن ادائیگی سے کب سستا پڑتا ہے؟ DeepSeek V4.1 Flash کے چار حل شدہ ماہانہ بل، بریک ایون حجم، اور اپنا استعمال ناپنے کا طریقہ۔ - [مفت LLM API کریڈٹ: ابتدائی کریڈٹ اصل میں کتنا چلتا ہے](https://broadrouter.com/ur/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): مفت LLM API کریڈٹ سے اصل میں کیا ملتا ہے: BroadNet کے ابتدائی کریڈٹ پر ٹوکن، آڈیو منٹ اور گھنٹے والی کلیدیں، Gemini اور OpenRouter کے مفت ٹائرز سے موازنہ۔ - [GDPR اور LLM API: اپنے فراہم کنندہ سے پوچھنے کے 15 سوال](https://broadrouter.com/ur/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): LLM API کے لیے GDPR چیک لسٹ: DPA، سب پروسیسرز، ڈیٹا ریٹینشن، ٹریننگ، بین الاقوامی منتقلی اور بریچ نوٹس، ہر سوال کے پیچھے GDPR کے آرٹیکل کے ساتھ۔ - [GLM 5.3 بمقابلہ DeepSeek V4.1 Flash: کون سا چنیں اور کب](https://broadrouter.com/ur/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 بمقابلہ DeepSeek V4.1 Flash: قیمت، کانٹیکسٹ، ٹولز، اسٹرکچرڈ آؤٹ پٹ، تصاویر اور ریزننگ کا موازنہ، اصل کاموں کی لاگت اور کب کسے چنیں، اس کا سیدھا اصول۔ - [LLM API میں 429 ایرر سے کیسے نمٹیں: بیک آف، جِٹر اور لمٹرز](https://broadrouter.com/ur/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 بمقابلہ 5xx، مکمل جِٹر کے ساتھ exponential backoff، LLM APIs کے لیے بیک وقت درخواستوں کے لمٹرز اور قطاریں، ساتھ میں پروڈکشن کے لیے Python اور Node کوڈ۔ - [گھنٹے والی LLM API کلید: فی گھنٹہ مقررہ قیمت کی وضاحت](https://broadrouter.com/ur/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): BroadNet کی گھنٹے والی کلید مقررہ قیمت پر ایک اسٹریم پر Gemma 4 VL 26B کا ایک گھنٹہ دیتی ہے۔ اس میں کیا شامل ہے، حدیں، یہ کس کے لیے ہے اور کب فائدہ دیتی ہے۔ - [LLM API ڈیٹا ریٹینشن کا موازنہ: OpenAI، Anthropic، Gemini، Mistral](https://broadrouter.com/ur/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): OpenAI، Anthropic، Gemini API، Mistral اور BroadNet، API پرامپٹس محفوظ کرنے اور ان پر ٹریننگ کے بارے میں سرکاری طور پر کیا کہتے ہیں، نومبر 2026 کے صفحات سے۔ - [LLM API کلید کو محفوظ کیسے رکھیں: اسٹوریج، روٹیشن اور IP allow-list](https://broadrouter.com/ur/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): LLM API کلیدیں کوڈ اور لاگز سے دور رکھیں، ہر ورک لوڈ کو الگ کلید دیں، مقررہ وقت پر روٹیٹ کریں، کلیدوں کو سرور IPs سے باندھیں اور لیک شدہ کلید سیکنڈوں میں منسوخ کریں۔ - [LLM API کی قیمت کیسے بنتی ہے: ان پٹ، کیش شدہ، آؤٹ پٹ اور ریزننگ](https://broadrouter.com/ur/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): فی ٹوکن LLM API کی قیمت کیسے بنتی ہے: ان پٹ، کیش شدہ ان پٹ، آؤٹ پٹ، ریزننگ ٹوکن، تصاویر اور فی منٹ آڈیو، ایک شائع شدہ ریٹ کارڈ کی مثالوں کے ساتھ۔ - [LLM API کو حساس ڈیٹا بھیجنا: redact کریں، کم بھیجیں، کم لاگ کریں](https://broadrouter.com/ur/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): گاہک، صحت یا مالی ڈیٹا LLM API کو محفوظ طریقے سے کیسے بھیجیں: فیلڈز کی درجہ بندی، واپس بدلے جا سکنے والے placeholders، زیرو ریٹینشن اور صاف لاگز۔ - [LLM API خرچ پر کنٹرول: بجٹ، کلید کی حدیں اور لوپ سے بچاؤ](https://broadrouter.com/ur/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): LLM API کا بل بے قابو ہونے سے روکیں: پری پیڈ بیلنس، ہر ایپ کی الگ کلید، IP allow-list، usage لاگنگ، کوڈ میں بجٹ، لوپ کی حد اور مقررہ قیمت والی کلیدیں۔ - [LLM API usage لاگنگ: ہر صارف اور فیچر کی لاگت، PII ہٹا کر](https://broadrouter.com/ur/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): ہر LLM API کال کو ٹوکنز، لاگت، لیٹنسی اور درخواست ID کے ساتھ لاگ کریں، خرچ کو صارفین اور فیچرز سے جوڑیں، اور لکھنے سے پہلے PII ہٹائیں۔ Python کوڈ کے ساتھ۔ - [فی صارف LLM لاگت: SaaS کے AI فیچر کی یونٹ اکنامکس](https://broadrouter.com/ur/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): حساب لگائیں کہ ایک AI فیچر فی فعال صارف فی مہینہ کتنے کا پڑتا ہے، بھاری صارفین اوسط سے زیادہ اہم کیوں ہیں، اور SaaS ٹائرز کی قیمت کیسے طے کریں۔ - [انفرنس فراہم کنندگان اوپن ماڈلز کی قیمت کیسے لگاتے ہیں: 6 کا موازنہ](https://broadrouter.com/ur/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): OpenRouter، Together AI، Fireworks، Groq، DeepInfra اور BroadNet اوپن ماڈلز کی قیمت کیسے لیتے ہیں: فی ٹوکن، کیش شدہ، بیچ، GPU گھنٹے اور مقررہ قیمت والے پلانز۔ - [LLM لیٹنسی: TTFT، فی سیکنڈ ٹوکن اور مکمل جواب کا وقت](https://broadrouter.com/ur/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): ٹائم ٹو فرسٹ ٹوکن، فی سیکنڈ ٹوکن اور اینڈ ٹو اینڈ لیٹنسی کیا ناپتے ہیں، کلائنٹ سے ہر ایک کو کوڈ کے ذریعے کیسے ناپیں، اور کون سی اصلاح کون سا عدد بدلتی ہے۔ - [LLM اسٹریمنگ کی وضاحت: SSE چنکس، ڈیلٹا، usage اور منسوخی](https://broadrouter.com/ur/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Chat completion اسٹریمنگ وائر پر کیسے چلتی ہے، ڈیلٹا سے جواب کیسے جوڑیں، آخری چنک سے usage پڑھیں اور Python، Node اور curl میں صاف طریقے سے منسوخ کریں۔ - [لانگ کانٹیکسٹ بمقابلہ RAG: کب 1M ٹوکن ویکٹر DB کی جگہ لیتے ہیں](https://broadrouter.com/ur/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): 1M ٹوکن کانٹیکسٹ ونڈو کب ویکٹر ڈیٹا بیس کی جگہ لے سکتی ہے اور کب RAG ہی بہتر ہے: فی کوئری لاگت، لیٹنسی، ریکال اور دیکھ بھال پر موازنہ۔ - [طویل چلنے والے ایجنٹس میں 1M ونڈو کے ساتھ کانٹیکسٹ مینجمنٹ](https://broadrouter.com/ur/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): 1M ٹوکن ونڈو کے ساتھ طویل ایجنٹ سیشن کو درست اور سستا کیسے رکھیں: append-only ہسٹری، کیش کے موافق پرامپٹ، کب compact کریں اور پہلے کیا ہٹائیں۔ - [لمبے چلنے والے ایجنٹس: پائیدار timeouts، retries اور checkpoints](https://broadrouter.com/ur/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): LLM ایجنٹ کو گھنٹوں چلاتے رہیں: ہر کال پر timeout، اٹکنے کا پتا لگانے کے لیے اسٹریمنگ، محفوظ retries، idempotent ٹول side effects اور ہر مرحلے کے بعد checkpoint۔ - [OpenAI سے مطابقت والے ماڈلز کے ساتھ MCP سرورز (کسی بھی فراہم کنندہ پر)](https://broadrouter.com/ur/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP کسی بھی LLM کے ساتھ چلتا ہے: کلائنٹ MCP ٹولز کو فنکشن کالز میں بدلتا ہے۔ Cline، Continue، opencode، n8n یا اپنے Python کوڈ میں اوپن ماڈلز سے MCP کیسے چلائیں۔ - [1M ٹوکن کانٹیکسٹ ونڈو کب فائدہ مند ہے (اور کب نہیں)](https://broadrouter.com/ur/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): 1,048,576 ٹوکن میں کیا سماتا ہے، DeepSeek V4.1 Flash پر اس کی لاگت کیا ہے، ریزننگ اور کیشنگ حساب کیسے بدلتے ہیں، اور کب RAG اب بھی بہتر ہے۔ - [ملٹی ایجنٹ سیٹ اپس کی اصل لاگت: فی ٹوکن بمقابلہ مقررہ سیشن](https://broadrouter.com/ur/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): ملٹی ایجنٹ سسٹمز ایک ایجنٹ سے کئی گنا ٹوکن خرچ کرتے ہیں۔ مقررہ پلانز بمقابلہ فی ٹوکن بلنگ کا سیشن والا حساب، چھپے ضارب، اور کب کون جیتتا ہے۔ - [کثیر لسانی speech-to-text API: 14 زبانیں، ایک اینڈپوائنٹ](https://broadrouter.com/ur/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): انگریزی، ہسپانوی، فرانسیسی، چینی، عربی اور 9 مزید زبانوں کو فی منٹ قیمت والے ایک OpenAI-compatible اینڈپوائنٹ سے ٹرانسکرائب کریں، اور جانیں کہ عربی ماڈل کب چنیں۔ - [سستے LLM پر n8n AI Agent: سیٹ اپ، ماڈل کا انتخاب اور فی رن لاگت](https://broadrouter.com/ur/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): کم لاگت والے OpenAI-compatible ماڈل پر n8n AI Agent ورک فلوز بنائیں: ٹولز کے لیے کون سا ماڈل، کون سی سیٹنگز ٹوکن کم کرتی ہیں، اور ہر 10,000 رنز کا حساب۔ - [کاروبار کے لیے open-weight بمقابلہ closed AI ماڈلز: لاگت، کنٹرول، خطرہ](https://broadrouter.com/ur/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): آپ کے کاروبار کے لیے open-weight یا closed LLMs؟ لاگت، کنٹرول، پرائیویسی، معیار کا فرق اور lock-in کا موازنہ، open ماڈلز چلانے کے تین طریقوں کے ساتھ۔ - [OpenAI-compatible API پر منتقلی: مکمل چیک لسٹ](https://broadrouter.com/ur/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): base_url بدلنا آسان حصہ ہے۔ ٹوٹتے ہیں ماڈل IDs، ریزننگ ٹوکنز، اسٹریمنگ usage، ایررز، ریٹ لمٹس اور کلیدیں۔ کوڈ کے ساتھ مکمل چیک لسٹ۔ - [AI ایجنٹس کے لیے OpenRouter بمقابلہ مقررہ قیمت والا پلان: کون سستا؟](https://broadrouter.com/ur/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter فی ٹوکن اور کریڈٹ پر فیس لیتا ہے؛ مقررہ قیمت والا پلان فی سیشن۔ ہر وقت چلنے والے ایجنٹس کی لاگت، OpenRouter کی فیس اور حدیں، اور کون کس کے لیے موزوں ہے۔ - [کوڈنگ ایجنٹس کو رات بھر بغیر نگرانی محفوظ طریقے سے کیسے چلائیں](https://broadrouter.com/ur/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): بغیر نگرانی کوڈنگ ایجنٹ رنز کا عملی سیٹ اپ: سینڈ باکس، git چیک پوائنٹس، ٹاسک فائلیں، مراحل کی حد، صبح کا جائزہ، اور رات بھر کے لیے مقررہ قیمت پلان سیشنز۔ - [پرامپٹ کی ساخت کیسے بنائیں تاکہ پری فکس کیش ہو](https://broadrouter.com/ur/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): LLM پرامپٹس کو ایسے ترتیب دیں کہ کیش شدہ پری فکس یکساں رہے: سسٹم پرامپٹ، ٹول اسکیما، دستاویزات، صرف آگے جڑتی ہسٹری، اور اسے ثابت کرنے والا ٹیسٹ۔ - [ریزننگ ٹوکنز: نظر آنے سے زیادہ آؤٹ پٹ کی ادائیگی کیوں ہوتی ہے](https://broadrouter.com/ur/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): ریزننگ ماڈلز چھپی سوچ کا بل آؤٹ پٹ ٹوکنز کے طور پر بناتے ہیں۔ اسے کیسے ناپیں، ہر ماڈل پر اس کی لاگت، اور reasoning_effort، max_tokens اور روٹنگ سے اسے کیسے کم کریں۔ - [مقررہ قیمت والے LLM API پر متوازی کوڈنگ ایجنٹس کیسے چلائیں](https://broadrouter.com/ur/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): لامحدود DeepSeek V4.1 Flash پر سب ایجنٹس، آرکسٹریٹرز اور بیچ ورکرز۔ کتنے پلان سیشن چاہییں، انہیں کلیدوں میں کیسے بانٹیں، اور اوورلیپ پر حد کیسے لگائیں۔ - [LLM خود ہوسٹ کریں یا API لیں: اصل لاگت اور بریک ایون](https://broadrouter.com/ur/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): سرکاری صفحات سے GPU کرائے کی قیمتیں، اپنے GPU پر فی دس لاکھ ٹوکن لاگت کا فارمولا، API کو مات دینے کے لیے درکار استعمال کی شرح، اور آپریشنز کی بھولی ہوئی لاگتیں۔ - [اسپیچ ٹو ٹیکسٹ API کی قیمتوں کا موازنہ: 2026 میں فی منٹ قیمت](https://broadrouter.com/ur/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): OpenAI، Google، AWS، Deepgram، AssemblyAI، Speechmatics، Groq اور BroadNet Transcribe کی فی منٹ اور فی 1,000 گھنٹے قیمتیں، آفیشل صفحات سے، ایڈ آنز سمیت۔ - [JSON mode بمقابلہ json_schema: قابل اعتماد اسٹرکچرڈ آؤٹ پٹ](https://broadrouter.com/ur/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON mode درست JSON دیتا ہے؛ json_schema آپ کی ساخت۔ OpenAI سے مطابقت والے ماڈلز پر دونوں کیسے مانگیں، Pydantic سے تصدیق کریں، اور غلط آؤٹ پٹ لوپ میں ٹھیک کریں۔ - [عربی بمقابلہ انگریزی میں فی لفظ ٹوکن: عربی مہنگی کیوں پڑتی ہے](https://broadrouter.com/ur/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): ایک ہی مطلب کے لیے عربی میں اکثر انگریزی سے زیادہ LLM ٹوکن کیوں لگتے ہیں، فی لفظ ٹوکن کیوں گمراہ کن ہے، اور usage فیلڈ سے اپنا تناسب ناپنے کی اسکرپٹ۔ - [اوپن ماڈلز کے ساتھ ٹول کالنگ: اسکیما، متوازی کالز، دوبارہ کوشش](https://broadrouter.com/ur/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): OpenAI سے مطابقت رکھنے والے اوپن ماڈلز پر فنکشن کالنگ: اسکیما کے مشورے، متوازی ٹول کالز، ٹول کالز کے ناکام ہونے کے پانچ طریقے، اور تصدیق و ری ٹرائی والا Python لوپ۔ - [کال سینٹر ریکارڈنگز کیسے ٹرانسکرائب کریں: عربی کے لیے تیار پائپ لائن](https://broadrouter.com/ur/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): اسٹیریو کالز کو چینل کے حساب سے الگ کریں، بولنے کے turns کاٹیں، عربی اور انگریزی ٹرانسکرائب کریں، PII ہٹائیں اور LLM سے خلاصہ بنائیں۔ لاگت کے ساتھ Python پائپ لائن۔ - [کوڈنگ ایجنٹس میں اوپن ماڈلز: Aider، Cline، opencode، Codex](https://broadrouter.com/ur/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Aider، Cline، Roo Code، opencode، Continue اور Codex CLI کو کسی بھی OpenAI-compatible اینڈپوائنٹ سے جوڑیں۔ ہر ایک کے فیلڈز کے درست نام، اور کون سی سیٹنگز ٹوٹتی ہیں۔ - [کاروبار کے لیے وائس نوٹ سے متن: WhatsApp اور Telegram سے ٹکٹس](https://broadrouter.com/ur/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): WhatsApp Business اور Telegram بوٹس پر آنے والے صارفین کے وائس نوٹس کو ٹرانسکرپٹ اور سپورٹ ٹکٹس میں بدلیں: ڈاؤن لوڈ، ٹرانسکرائب، LLM سے فیلڈز نکالیں، روٹ کریں۔ - [مختصر پروجیکٹس کا ہفتہ وار LLM پلان: Orbit، گھنٹہ یا مہینہ](https://broadrouter.com/ur/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): ایک ہفتے کا لامحدود DeepSeek V4.1 Flash پلان کب گھنٹے والے سیشن، ماہانہ پلان یا فی ٹوکن ادائیگی سے بہتر ہے: ہیکاتھون، اسپرنٹ اور مائیگریشن کا حساب۔ - [LLM API پلان میں ”لامحدود“ کا اصل مطلب کیا ہے؟](https://broadrouter.com/ur/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): ٹوکن میٹر، 5 گھنٹے کی رولنگ ونڈو، ہفتہ وار حد، ریٹ لمٹ اور کنکرنسی۔ سات سوال جو بتاتے ہیں کہ ”لامحدود“ AI پلان آپ کے ایجنٹ کے لیے واقعی لامحدود ہے یا نہیں۔ - [Whisper API کا متبادل: کب بدلیں، اور اسے کیسے ٹیسٹ کریں](https://broadrouter.com/ur/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): OpenAI Whisper API کا ہوسٹڈ متبادل کب درست ہے: آپ کی آڈیو پر word error rate، زبانیں، فی گھنٹہ قیمت، حدیں اور پرائیویسی، ساتھ میں ٹیسٹ کوڈ۔ - [زیرو ڈیٹا ریٹینشن والا LLM API: 10 سوالوں کی چیک لسٹ](https://broadrouter.com/ur/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): LLM API کے لیے ”زیرو ڈیٹا ریٹینشن“ کا کیا مطلب ہونا چاہیے: پرامپٹس، آؤٹ پٹس، لاگز، ٹریننگ، ذیلی پروسیسرز اور آڈٹس۔ ڈیٹا بھیجنے سے پہلے پوچھنے کے دس سوال۔ - [Guides in Urdu](https://broadrouter.com/ur/guides/) - [BroadNet Inference کے ساتھ OpenAI Python SDK استعمال کریں](https://broadrouter.com/ur/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): صرف base_url بدل کر سرکاری OpenAI Python SDK کو BroadNet Inference سے جوڑیں۔ انسٹال کریں، کلید سیٹ کریں، پہلی درخواست بھیجیں اور جواب اسٹریم کریں۔ - [BroadNet Inference کے ساتھ OpenAI Node.js SDK استعمال کریں](https://broadrouter.com/ur/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): baseURL سیٹ کر کے سرکاری OpenAI Node.js اور TypeScript SDK کو BroadNet Inference سے جوڑیں: انسٹال، پہلی درخواست، ٹوکن اسٹریمنگ اور ریکویسٹ ID لاگ کرنا۔ - [BroadNet Inference کو براؤزر میں آزمائیں، اکاؤنٹ کے بغیر](https://broadrouter.com/ur/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): BroadNet کے براؤزر پلے گراؤنڈ میں سائن اپ یا کارڈ کے بغیر عربی اسپیچ ٹو ٹیکسٹ، دستاویز ریڈنگ اور Gemma 4 چیٹ ٹیسٹ کریں، پھر وہی درخواست API پر لے جائیں۔ - [curl کے ساتھ BroadNet Inference کو پہلی درخواست](https://broadrouter.com/ur/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): curl سے BroadNet Inference کی پہلی کالز کریں: ایک chat completion، ایک اسٹریم شدہ جواب اور ایک آڈیو ٹرانسکرپشن، پھر ریسپانس، usage اور ایررز پڑھنا سیکھیں۔ - [chat completions اسٹریم کریں اور ٹوٹے کنکشن سنبھالیں](https://broadrouter.com/ur/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): BroadNet Inference پر اسٹریمنگ کیسے کام کرتی ہے: Server-Sent Events، [DONE] مارکر، آخری چنک میں ٹوکن usage، اور اسٹریم درمیان میں ختم ہو جائے تو کیا کریں۔ - [ریزننگ ماڈل خالی جواب کیوں لوٹاتا ہے](https://broadrouter.com/ur/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash، GLM 5.3 اور Qwen3.8 27B جواب سے پہلے سوچتے ہیں، اور یہ ریزننگ max_tokens میں گنی جاتی ہے۔ خالی جوابات سے بچنے کا طریقہ۔ - [BroadNet Inference کے ایرر کوڈز اور دوبارہ کوشش کب کریں](https://broadrouter.com/ur/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): BroadNet Inference API کے ہر HTTP اسٹیٹس کا مطلب، کن پر بیک آف کے ساتھ دوبارہ کوشش کریں، کنہیں پہلے ٹھیک کریں، اور سپورٹ کو ریکویسٹ ID کے ساتھ کیا بھیجیں۔ - [IP allow-list سے API کلید کو اپنے سرورز تک محدود کریں](https://broadrouter.com/ur/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): IP allow-list سے BroadNet Inference کی کلیدیں محفوظ کریں، انفرنس کلیدیں پورٹل کریڈنشلز سے الگ رکھیں اور لیک شدہ کلید سیکنڈوں میں منسوخ کریں۔ curl مثالوں کے ساتھ۔ - [ایجنٹس کے لیے لامحدود LLM API پلانز کیسے کام کرتے ہیں](https://broadrouter.com/ur/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): لامحدود DeepSeek V4.1 Flash پلانز میں نہ ٹوکن میٹر ہے نہ ہفتہ وار حد۔ جانیں کہ سیشنز، متوازی ایجنٹس، کلیدیں اور بلنگ اصل میں کیسے کام کرتے ہیں۔ - [Aider کو لامحدود DeepSeek V4.1 Flash پر چلائیں](https://broadrouter.com/ur/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Aider کوڈنگ ایجنٹ کو اس کی OpenAI-compatible سیٹنگ کے ذریعے BroadNet Inference سے جوڑیں اور مقررہ قیمت والے لامحدود DeepSeek V4.1 Flash پلان پر چلائیں۔ - [BroadNet Inference کے ساتھ LangChain استعمال کریں](https://broadrouter.com/ur/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): base_url اور api_key کے ذریعے LangChain کے ChatOpenAI کو BroadNet Inference سے جوڑیں، اور جانیں کہ LangChain ریزننگ ماڈلز کے جواب میں سے کیا رکھتا ہے اور کیا نہیں۔ - [Cline میں BroadNet Inference (OpenAI Compatible فراہم کنندہ)](https://broadrouter.com/ur/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Cline VS Code ایکسٹینشن کو اس کے OpenAI Compatible فراہم کنندہ کے ذریعے BroadNet Inference پر سیٹ کریں: درست فیلڈز، ماڈل IDs، کانٹیکسٹ سائز اور عام مسائل کا حل۔ - [Roo Code میں BroadNet Inference (OpenAI Compatible)](https://broadrouter.com/ur/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code مئی 2026 میں بند ہو گیا۔ موجودہ انسٹالز اور فورکس کی سیٹنگز، اور BroadNet Inference کے ساتھ آگے کہاں جائیں (Cline یا Zoo Code فورک)۔ - [Continue میں BroadNet Inference (config.yaml اور apiBase)](https://broadrouter.com/ur/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): config.yaml میں provider openai اور apiBase کے ذریعے Continue IDE ایکسٹینشن میں BroadNet Inference کے ماڈلز شامل کریں، tool_use صلاحیت اور کانٹیکسٹ لمبائی سمیت۔ - [opencode میں BroadNet Inference کو کسٹم پرووائیڈر کے طور پر شامل کریں](https://broadrouter.com/ur/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): opencode.jsonc میں BroadNet Inference کو opencode کوڈنگ ایجنٹ کے کسٹم OpenAI-compatible پرووائیڈر کے طور پر شامل کریں، ماڈل aliases، capabilities اور limits کے ساتھ۔ - [Open WebUI کو BroadNet Inference سے جوڑیں](https://broadrouter.com/ur/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Open WebUI میں BroadNet Inference کو OpenAI-compatible کنکشن کی طرح جوڑیں، ماڈل IDs کو allowlist میں ڈالیں، اور verify، خالی جواب اور ریٹ لمٹ کے مسائل سلجھائیں۔ - [LibreChat میں BroadNet Inference بطور کسٹم اینڈ پوائنٹ شامل کریں](https://broadrouter.com/ur/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): librechat.yaml میں BroadNet Inference کو OpenAI-compatible کسٹم اینڈ پوائنٹ بنائیں: .env میں کلید، ماڈلز کی فہرست، ٹائٹل ماڈل اور عام مسائل کا حل۔ - [n8n میں BroadNet Inference استعمال کریں (OpenAI credential، Base URL)](https://broadrouter.com/ur/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): n8n کے OpenAI credential میں Base URL فیلڈ سے BroadNet Inference سیٹ کریں، پھر OpenAI Chat Model نوڈ کو ID سے چلائیں اور Responses API ٹوگل بند رکھیں۔ - [BroadNet Inference کے ساتھ Vercel AI SDK استعمال کریں](https://broadrouter.com/ur/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): @ai-sdk/openai-compatible کے createOpenAICompatible سے Vercel AI SDK میں BroadNet Inference کال کریں: generateText، streamText، ٹولز اور reasoning effort۔ - [عربی speech-to-text API، قدم بہ قدم](https://broadrouter.com/ur/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): OpenAI-compatible اینڈپوائنٹ کے ذریعے BroadNet Transcribe Arabic سے عربی آڈیو ٹرانسکرائب کریں۔ فارمیٹس، language فیلڈ، Python اور curl کی مثالیں، اور قیمت۔ - [LiteLLM کو BroadNet Inference پر روٹ کریں (SDK اور پراکسی)](https://broadrouter.com/ur/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): openai/ ماڈل پری فکس اور api_base کے ساتھ LiteLLM سے BroadNet Inference کال کریں، Python SDK میں یا پراکسی کی config.yaml میں، مطابقت رکھنے والی ریٹ لمٹس کے ساتھ۔ - [BroadNet Inference پر OpenAI Agents SDK چلائیں](https://broadrouter.com/ur/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): AsyncOpenAI اور OpenAIChatCompletionsModel کے ذریعے Python والے OpenAI Agents SDK کو BroadNet Inference سے جوڑیں، اور OpenAI کلید مانگنے والی ٹریسنگ بند کریں۔ - [API کے ذریعے WhatsApp وائس نوٹس کو متن میں بدلیں](https://broadrouter.com/ur/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): WhatsApp کے .ogg وائس نوٹس یا فون کی .m4a ریکارڈنگز کو ffmpeg سے WAV میں بدلیں اور ایک API کال سے عربی یا 13 دوسری زبانوں میں ٹرانسکرائب کریں۔ - [n8n میں WhatsApp وائس نوٹس کو متن میں بدلیں](https://broadrouter.com/ur/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): ایک n8n ورک فلو جو WhatsApp وائس نوٹ وصول کرتا ہے، Meta سے ڈاؤن لوڈ کرتا ہے، ffmpeg سے تبدیل کرتا ہے اور عربی یا 13 دوسری زبانوں میں ٹرانسکرائب کرتا ہے۔ - [وژن ماڈل سے عربی انوائسز اور رسیدیں پڑھیں](https://broadrouter.com/ur/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): انوائس یا رسید کی تصویر BroadNet Inference کے وژن ماڈل کو بھیجیں اور کل رقم، تاریخ اور فروخت کنندہ JSON میں حاصل کریں، سوال عربی یا انگریزی میں پوچھیں۔ - [Zed میں BroadNet Inference کو OpenAI-compatible فراہم کنندہ بنائیں](https://broadrouter.com/ur/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Zed ایڈیٹر کی settings.json میں BroadNet Inference کو OpenAI-compatible فراہم کنندہ کے طور پر شامل کریں: ماڈل IDs، ٹوکن حدیں، capabilities اور کلید کا ویری ایبل۔ - [Kilo Code میں BroadNet Inference (OpenAI Compatible فراہم کنندہ)](https://broadrouter.com/ur/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Kilo Code کو BroadNet Inference سے کسٹم OpenAI Compatible فراہم کنندہ کے طور پر جوڑیں، VS Code ڈائیلاگ یا kilo.json میں: ماڈل IDs، حدوں اور ٹول فلیگز کے ساتھ۔ - [goose میں BroadNet Inference کو کسٹم فراہم کنندہ کے طور پر شامل کریں](https://broadrouter.com/ur/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): goose AI ایجنٹ کو BroadNet Inference سے OpenAI Compatible کسٹم فراہم کنندہ کے طور پر جوڑیں: goose Desktop، goose configure یا ایک JSON فائل سے۔ - [BroadNet Inference کے ساتھ Pydantic AI (OpenAICompatibleProvider)](https://broadrouter.com/ur/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): OpenAIChatModel اور OpenAICompatibleProvider سے BroadNet Inference پر Pydantic AI ایجنٹس چلائیں، ساتھ میں ٹول کالز، اسٹرکچرڈ آؤٹ پٹ اور ماڈل پروفائل کے حل۔ - [CrewAI میں BroadNet Inference کو کسٹم OpenAI اینڈ پوائنٹ بنائیں](https://broadrouter.com/ur/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): LLM(custom_openai=True, base_url=...) سے CrewAI ایجنٹس کو BroadNet Inference پر لائیں، پھر ٹولز، لمبے کانٹیکسٹ اور مقررہ قیمت والے پلان کے لیے ماڈل چنیں۔ - [OpenAILike کے ذریعے BroadNet Inference کے ساتھ LlamaIndex](https://broadrouter.com/ur/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): OpenAILike کلاس سے LlamaIndex کو BroadNet Inference سے جوڑیں، اور api_base، context_window، is_chat_model اور is_function_calling_model درست سیٹ کریں۔ - [OpenAI-API-compatible پلگ اِن سے Dify میں BroadNet Inference شامل کریں](https://broadrouter.com/ur/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): OpenAI-API-compatible ماڈل فراہم کنندہ پلگ اِن سے Dify کو BroadNet Inference سے جوڑیں: درست فیلڈ ویلیوز، کانٹیکسٹ سائز اور function calling سیٹنگز کے ساتھ۔ - [Jan میں BroadNet Inference: کسٹم OpenAI-compatible فراہم کنندہ](https://broadrouter.com/ur/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Jan ڈیسک ٹاپ ایپ میں BroadNet Inference کو کسٹم OpenAI-compatible فراہم کنندہ کے طور پر شامل کریں، ماڈل IDs ہاتھ سے ڈالیں، اور ہر ماڈل کے لیے ٹولز اور ویژن آن کریں۔ - [SillyTavern کو BroadNet Inference سے جوڑیں (Custom OpenAI-compatible)](https://broadrouter.com/ur/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): SillyTavern کو Custom (OpenAI-compatible) Chat Completion سورس سے BroadNet Inference سے جوڑیں: درست ماڈل IDs، پرامپٹ پوسٹ پروسیسنگ اور ٹوکن کی حدیں۔ - [Spring AI کو BroadNet Inference کے ساتھ چلائیں (OpenAI base-url)](https://broadrouter.com/ur/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Spring AI کے OpenAI چیٹ اسٹارٹر کو spring.ai.openai.base-url، BroadNet ماڈل ID اور max-tokens سے BroadNet Inference پر لگائیں، پھر ChatClient سے کال کریں۔ - [openai-go SDK سے Go میں BroadNet Inference کال کریں](https://broadrouter.com/ur/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): option.WithBaseURL سے آفیشل openai-go v3 SDK کو BroadNet Inference سے جوڑیں: Go میں چیٹ کمپلیشنز، usage کے ساتھ اسٹریمنگ اور ایرر ہینڈلنگ۔ - [BroadNet Inference پر LangGraph ایجنٹس بنائیں](https://broadrouter.com/ur/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): ChatOpenAI، base_url اور BroadNet ماڈل ID سے LangGraph کا StateGraph یا create_agent، BroadNet Inference پر چلائیں، ساتھ میں ٹولز، اسٹریمنگ اور میموری۔ - [BroadNet Inference کے ساتھ Haystack پائپ لائنز اور ایجنٹس چلائیں](https://broadrouter.com/ur/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Haystack کے OpenAIChatGenerator کو api_base_url، BroadNet ماڈل ID اور Secret کلید سے BroadNet Inference پر سیٹ کریں: پائپ لائنز، ٹولز، اسٹریمنگ اور ایجنٹس کے لیے۔ - [Semantic Kernel کو BroadNet Inference کے ساتھ استعمال کریں](https://broadrouter.com/ur/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Microsoft Semantic Kernel کو BroadNet Inference سے جوڑیں: Python میں AsyncOpenAI کلائنٹ سے، یا C# میں کسٹم endpoint سے، چیٹ، پلگ انز اور ایجنٹس کے لیے۔ - [BroadNet Inference کے ساتھ Mastra ایجنٹس استعمال کریں](https://broadrouter.com/ur/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): id، url اور apiKey والے model آبجیکٹ سے Mastra ایجنٹس BroadNet Inference پر چلائیں، ساتھ میں وہ دوہرا broadnet/ پری فکس جو ماڈل ID کو چاہیے۔ - [BroadNet Inference پر Instructor سے اسٹرکچرڈ آؤٹ پٹ حاصل کریں](https://broadrouter.com/ur/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Instructor سے BroadNet Inference پر Pydantic آبجیکٹس نکالیں: کسٹم base_url کے ساتھ from_openai یا from_provider استعمال کریں، اور درست Mode منتخب کریں۔ - [DSPy میں BroadNet Inference (model_type="chat" سیٹ کریں)](https://broadrouter.com/ur/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): openai/ پری فکس، api_base اور model_type="chat" سے BroadNet Inference کے لیے dspy.LM سیٹ کریں، کیونکہ DSPy 3.5 OpenAI ماڈلز کو Responses API پر بھیجتا ہے۔ - [promptfoo سے BroadNet ماڈلز کی ایویلیوایشن کریں](https://broadrouter.com/ur/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): openai:chat پرووائیڈر، apiBaseUrl اور apiKeyEnvar سے BroadNet Inference پر promptfoo evals اور ماڈلز کے آمنے سامنے ٹیسٹ چلائیں، گریڈرز کے ساتھ۔ - [OpenAIModel کے ساتھ smolagents کو BroadNet Inference پر چلائیں](https://broadrouter.com/ur/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Hugging Face smolagents کو OpenAIModel اور api_base سے BroadNet Inference سے جوڑیں: کسی بھی ماڈل پر CodeAgent، یا GLM 5.3 اور Qwen3.8 27B پر ToolCallingAgent۔ - [Flowise کو BroadNet Inference سے جوڑیں (OpenAI Custom Model)](https://broadrouter.com/ur/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Flowise چیٹ فلوز میں OpenAI Custom Model نوڈ سے BroadNet Inference استعمال کریں: Base Path، OpenAI API کریڈینشل اور ہاتھ سے ٹائپ کیا گیا BroadNet ماڈل نام۔ - [Langflow کو BroadNet Inference سے جوڑیں](https://broadrouter.com/ur/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Langflow کے OpenAI کمپوننٹ میں OpenAI API Base فیلڈ اور ٹائپ کیے گئے ماڈل نام سے BroadNet Inference استعمال کریں، اور جانیں OpenAI Compatible پرووائیڈر کیوں نہیں۔ - [Glossary entries in Urdu](https://broadrouter.com/ur/glossary/) - [AI ایجنٹ (AI agent)](https://broadrouter.com/ur/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): AI ایجنٹ ایسا نظام ہے جس میں لینگویج ماڈل کسی ہدف کی طرف اپنے قدم خود طے کرتا ہے: کون سے ٹولز کال کرنے ہیں، یہ چنتا ہے، نتائج پڑھتا ہے اور حتمی جواب یا کسی حد تک دہراتا رہتا ہے۔ مقررہ مقامات پر ماڈل کو کال کرنے والی اسکرپٹ کو عموماً workflow کہا جاتا ہے۔ - [API کلید (API key)](https://broadrouter.com/ur/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): API کلید ایک خفیہ اسٹرنگ ہے جو API کال کرنے والے کلائنٹ کی شناخت کرتی اور اسے اجازت دیتی ہے۔ LLM API عموماً اسے Authorization: Bearer ہیڈر میں چاہتے ہیں، اور جس کے پاس بھی یہ کلید ہو وہ ایسی درخواستیں کر سکتا ہے جن کا بل کلید والے اکاؤنٹ پر بنتا ہے۔ - [عربی اعراب (tashkeel)](https://broadrouter.com/ur/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): عربی اعراب، یا tashkeel، حروف کے اوپر یا نیچے لگنے والے چھوٹے نشانات ہیں جو مختصر حرکات، مشدّد حروف اور حرکت کی عدم موجودگی دکھاتے ہیں۔ یہ اختیاری ہیں: خبروں سمیت زیادہ تر عربی متن انہیں چھوڑ دیتا ہے، اس لیے قاری اور سافٹ ویئر سیاق و سباق سے حرکات کا اندازہ لگاتے ہیں۔ - [عربی لہجے کی شناخت (Arabic dialect identification)](https://broadrouter.com/ur/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): عربی لہجے کی شناخت وہ کام ہے جس میں طے کیا جاتا ہے کہ کوئی متن یا ریکارڈنگ کس عربی شکل میں ہے، جیسے مصری، خلیجی، شامی (لیونٹائن)، مغربی یا جدید معیاری عربی، اور یہ علاقے، ملک یا شہر کی سطح پر ہوتا ہے۔ اسے پیغامات روٹ کرنے، پرامپٹ چننے اور جانچ کے سیٹ بانٹنے میں استعمال کیا جاتا ہے۔ - [عربی متن کی نارملائزیشن (Arabic text normalization)](https://broadrouter.com/ur/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): عربی متن کی نارملائزیشن وہ مرحلہ ہے جو میچنگ سے پہلے ایک ہی لفظ کی مختلف املاؤں کو ایک شکل میں بدلتا ہے: اعراب اور tatweel ہٹانا، الف کی شکلیں ایک کرنا، اور اکثر ى کو ي اور ة کو ه کرنا۔ یہ سرچ کی اور جانچ پر لگتی ہے، دکھائے جانے والے متن پر نہیں۔ - [Arabizi (عربیزی)](https://broadrouter.com/ur/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): Arabizi لاطینی حروف میں غیر رسمی طور پر لکھی گئی عربی ہے، جس میں لاطینی متبادل سے محروم عربی آوازوں کے لیے ہندسے آتے ہیں، جیسے ع کے لیے 3۔ یہ چیٹ اور سوشل میڈیا میں عام ہے، زیادہ تر بول چال کی ہے، اس کی طے شدہ املا نہیں، اور اکثر انگریزی یا فرانسیسی سے ملی ہوتی ہے۔ - [AWQ (activation-aware weight quantization)](https://broadrouter.com/ur/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (activation-aware weight quantization) ایک post-training طریقہ ہے جو LLM کے weights کو کم پریسیژن، عموماً 4-bit integers، میں دباتا ہے۔ یہ چھوٹے calibration سیٹ کے activation اعداد سے اہم ترین weight چینلز پہچان کر انہیں scale کرتا ہے، تاکہ ان کی درستی کم گھٹے۔ اسے backpropagation نہیں چاہیے۔ - [base URL](https://broadrouter.com/ur/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): base URL کسی API کا مقررہ بنیادی پتہ ہے، جیسے https://api.example.com/v1، جسے کلائنٹ /chat/completions جیسے ہر اینڈپوائنٹ پاتھ کے آگے لگاتا ہے۔ OpenAI-compatible SDK میں base URL کو API کلید اور ماڈل ID کے ساتھ بدلنے سے وہی کوڈ کسی دوسرے فراہم کنندہ سے جڑ جاتا ہے۔ - [Batch API (بیچ API)](https://broadrouter.com/ur/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Batch API ایک asynchronous LLM اینڈپوائنٹ ہے جو ایک ساتھ کئی درخواستیں لیتا ہے، عموماً ایک فائل کی صورت میں، انہیں فوراً نہیں بلکہ ایک طے شدہ مدت میں پروسیس کرتا ہے، اور نتائج اکٹھے لوٹاتا ہے۔ OpenAI اور Anthropic جیسے فراہم کنندگان batch درخواستوں پر اپنے عام API سے کم قیمت لیتے ہیں۔ - [BM25](https://broadrouter.com/ur/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) کی ورڈ سرچ کا ایک رینکنگ فنکشن ہے جو دستاویز کو تین باتوں سے اسکور کرتا ہے: کوئری کے الفاظ اس میں کتنی بار آتے ہیں، وہ پورے مجموعے میں کتنے نایاب ہیں، اور دستاویز کتنی لمبی ہے۔ یہ Lucene اور Elasticsearch کی ڈیفالٹ similarity ہے۔ - [بائٹ پیئر انکوڈنگ (byte pair encoding، BPE)](https://broadrouter.com/ur/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Byte pair encoding (BPE) ایک subword tokenization الگورتھم ہے جو اکیلے کیریکٹرز یا bytes سے شروع ہوتا ہے اور سب سے عام ساتھ ساتھ والے جوڑے کو بار بار ملا کر نئی ذخیرۂ الفاظ اندراج بناتا ہے، جب تک مطلوبہ سائز نہ آ جائے۔ آج کے زیادہ تر LLM tokenizer BPE یا اس کی byte-level شکل استعمال کرتے ہیں۔ - [کیش شدہ ان پٹ ٹوکن (cached input tokens)](https://broadrouter.com/ur/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): کیش شدہ ان پٹ ٹوکن وہ پرامپٹ ٹوکن ہیں جنہیں LLM فراہم کنندہ نے حال ہی میں پروسیس کیا ہو اور دوبارہ حساب کرنے کے بجائے کیش سے پھر استعمال کر سکے، عموماً سسٹم پرامپٹ یا دستاویز جیسا دہرایا گیا پریفکس۔ فراہم کنندگان ان کا بل نئے ان پٹ ٹوکنز سے کم شرح پر بناتے ہیں۔ - [چین آف تھاٹ (chain-of-thought، CoT)](https://broadrouter.com/ur/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Chain-of-thought (CoT) ایک پرامپٹنگ تکنیک ہے جس میں لینگویج ماڈل کو حتمی جواب سے پہلے درمیانی ریزننگ مراحل لکھ کر دکھائے جاتے ہیں یا لکھنے کو کہا جاتا ہے۔ اس سے حساب اور منطق جیسے کئی مرحلوں والے مسائل پر درستی بڑھتی ہے، اور ریزننگ ماڈلز اب اسے خود بخود بناتے ہیں۔ - [کیریکٹر ایرر ریٹ (character error rate، CER)](https://broadrouter.com/ur/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Character error rate (CER) ٹرانسکرپشن یا OCR کی درستی کیریکٹر کی سطح پر ناپتا ہے: آؤٹ پٹ کو ریفرنس متن بنانے کے لیے درکار substitutions، deletions اور insertions، تقسیم ریفرنس کے کیریکٹرز کی تعداد۔ کم بہتر ہے، اور 0 مکمل مطابقت ہے۔ - [Chat Completions API](https://broadrouter.com/ur/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Chat Completions API، OpenAI کا HTTP اینڈپوائنٹ POST /v1/chat/completions ہے، جو ایک ماڈل کا نام اور رول والے پیغامات کی فہرست لیتا ہے اور ماڈل کا اگلا پیغام لوٹاتا ہے۔ اس کی درخواست اور جواب کی ساخت وہ عام فارمیٹ بن گئی جسے OpenAI-compatible فراہم کنندگان اور SDKs اپناتے ہیں۔ - [چیٹ ٹیمپلیٹ (chat template)](https://broadrouter.com/ur/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): چیٹ ٹیمپلیٹ وہ فارمیٹنگ اصول ہے، عموماً Jinja میں لکھا اور ماڈل کے tokenizer کے ساتھ آنے والا، جو roles والے چیٹ پیغامات کی فہرست کو اس ایک ٹوکن ترتیب میں بدلتا ہے جس کی ماڈل توقع کرتا ہے، اس کے special control ٹوکنز سمیت۔ غلط ٹیمپلیٹ سے جوابات بہت بگڑ سکتے ہیں۔ - [چنکنگ (chunking، RAG)](https://broadrouter.com/ur/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Chunking retrieval پائپ لائن کا وہ مرحلہ ہے جو دستاویزات کو embed اور انڈیکس کرنے سے پہلے چھوٹے ٹکڑوں، یعنی chunks، میں بانٹتا ہے۔ Chunk size طے کرتا ہے کہ ہر ٹکڑے میں کتنا متن ہو، اور chunk overlap ساتھ والے chunks میں کچھ متن دہراتا ہے تاکہ حد پر کٹے خیالات ضائع نہ ہوں۔ - [کوڈ سوئچنگ (code-switching)](https://broadrouter.com/ur/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Code-switching ایک ہی گفتگو یا جملے میں، بول کر یا لکھ کر، دو یا زیادہ زبانوں، یا ایک زبان کی مختلف شکلوں، کا استعمال ہے۔ مثالیں: عربی بولنے والے اپنے لہجے میں انگریزی الفاظ ملاتے ہیں، یا جدید معیاری عربی اور اپنے لہجے کے درمیان آتے جاتے ہیں۔ - [AI کوڈنگ ایجنٹ (AI coding agent)](https://broadrouter.com/ur/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): AI کوڈنگ ایجنٹ ایک سافٹ ویئر ٹول ہے جس میں لینگویج ماڈل کوڈ بیس پر لوپ میں کام کرتا ہے: فائلیں پڑھتا ہے، انہیں بدلتا ہے، ٹیسٹ جیسی کمانڈز چلاتا ہے، آؤٹ پٹ پڑھتا ہے اور اگلا قدم طے کرتا ہے، جب تک کام مکمل نہ ہو یا کوئی حد نہ آ جائے۔ - [کنکرنسی (concurrency)](https://broadrouter.com/ur/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): کنکرنسی، کسی کلائنٹ کی بیک وقت چل رہی (in-flight) درخواستوں کی تعداد ہے۔ LLM APIs اکثر اس کی حد ہر کلید یا پلان پر، فی منٹ درخواستوں سے الگ رکھتے ہیں، کیونکہ ہر کھلی درخواست جواب مکمل ہونے تک GPU کی گنجائش گھیرے رکھتی ہے۔ - [کانٹیکسٹ کمپیکشن (context compaction)](https://broadrouter.com/ur/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): Context compaction وہ طریقہ ہے جس میں کانٹیکسٹ ونڈو کی حد کے قریب پہنچی گفتگو کا خلاصہ بنایا جاتا ہے اور سیشن اسی خلاصے سے آگے چلتا ہے، تاکہ ایجنٹ اہم فیصلے اور حالت برقرار رکھے اور پرانے ٹول آؤٹ پٹ جیسی خام ہسٹری چھوڑ دے۔ - [کانٹیکسٹ انجینئرنگ (context engineering)](https://broadrouter.com/ur/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): کانٹیکسٹ انجینئرنگ وہ کام ہے جس میں طے کیا اور سنبھالا جاتا ہے کہ لینگویج ماڈل ہر درخواست میں کون سے ٹوکن دیکھے: سسٹم پرامپٹ، ٹول ڈیفینیشنز، تاریخ، ریٹریو کی گئی دستاویزات اور نوٹس، تاکہ ماڈل کو کام کے لیے ضروری، سب سے مختصر اور سب سے کارآمد معلومات ملیں۔ - [کانٹیکسٹ روٹ (context rot)](https://broadrouter.com/ur/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): کانٹیکسٹ روٹ درستگی اور بھروسے کا وہ بتدریج نقصان ہے جو لینگویج ماڈلز میں ان پٹ لمبا ہونے پر نظر آتا ہے، تب بھی جب کام وہی رہے اور ان پٹ کانٹیکسٹ ونڈو کے اندر ہی ہو۔ یہ اصطلاح 2025 کی Chroma تکنیکی رپورٹ سے مشہور ہوئی، جس نے 18 ماڈلز ٹیسٹ کیے۔ - [کانٹیکسٹ ونڈو (context window)](https://broadrouter.com/ur/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): کانٹیکسٹ ونڈو ٹوکنز کی وہ زیادہ سے زیادہ تعداد ہے جس کے ساتھ لینگویج ماڈل ایک درخواست میں کام کر سکتا ہے۔ اس میں پرامپٹ، گفتگو کی تاریخ، ٹول ڈیفینیشنز، تصاویر اور بننے والا آؤٹ پٹ، ریزننگ ٹوکنز سمیت، سب شامل ہیں۔ جو اس میں نہ سمائے اسے کاٹنا، خلاصہ کرنا یا کئی درخواستوں میں بانٹنا پڑتا ہے۔ - [ڈیٹا ریزیڈنسی (data residency)](https://broadrouter.com/ur/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): ڈیٹا ریزیڈنسی وہ شرط یا وعدہ ہے کہ ڈیٹا کسی مخصوص ملک یا خطے میں ہی محفوظ اور پروسیس ہو گا۔ LLM API کے لیے یہ بتاتی ہے کہ انفرینس کے دوران پرامپٹس کہاں پروسیس ہوتے ہیں اور محفوظ کیا گیا مواد کہاں رہتا ہے؛ اکثر اکاؤنٹ اور بلنگ ڈیٹا اس میں شامل نہیں ہوتا۔ - [عربی لہجے (dialectal Arabic)](https://broadrouter.com/ur/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): عربی لہجے کسی خطے یا شہر میں روزمرہ بولی جانے والی عربی ہیں، جیسے مصری، خلیجی، شامی (لیونٹائن) یا مغربی عربی۔ یہ جدید معیاری عربی سے آوازوں، الفاظ اور گرامر میں مختلف ہیں، ان کا کوئی معیاری املا نہیں، اور آن لائن یہ بڑھتی تعداد میں لکھے جا رہے ہیں، جس سے لینگویج ماڈلز کے لیے یہ مشکل ہیں۔ - [اسپیکر ڈائرائزیشن (speaker diarization)](https://broadrouter.com/ur/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): اسپیکر ڈائرائزیشن آڈیو ریکارڈنگ کو بولنے والوں کے لحاظ سے حصوں میں بانٹنے کا کام ہے، جو بتاتا ہے کہ "کون کب بولا"۔ یہ گفتگو کے ہر حصے پر SPEAKER_00 جیسی گمنام اسپیکر ID لگاتی ہے، اور عموماً ٹرانسکرپٹ کے ساتھ ملا کر دکھاتی ہے کہ ہر سطر کس نے کہی۔ - [ڈسٹلیشن (knowledge distillation)](https://broadrouter.com/ur/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): ڈسٹلیشن، یا نالج ڈسٹلیشن، تربیت کا وہ طریقہ ہے جس میں ایک چھوٹا اسٹوڈنٹ ماڈل بڑے ٹیچر ماڈل کے آؤٹ پٹ، جیسے اس کی احتمالی تقسیمیں یا بنائے گئے جوابات، دہرانا سیکھتا ہے، تاکہ اسٹوڈنٹ کم لاگت پر چلتے ہوئے ٹیچر کا کافی معیار برقرار رکھے۔ - [ڈیٹا پروسیسنگ معاہدہ (DPA)](https://broadrouter.com/ur/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): ڈیٹا پروسیسنگ معاہدہ (DPA) کنٹرولر اور پروسیسر کے درمیان وہ معاہدہ ہے جو طے کرتا ہے کہ پروسیسر ذاتی ڈیٹا کیسے سنبھالے گا۔ GDPR آرٹیکل 28(3) اسے لازمی کرتا ہے اور اس کے مندرجات گنواتا ہے: تحریری ہدایات، رازداری، سیکیورٹی، سب پروسیسر کے اصول، ڈیٹا مٹانا یا واپس کرنا، اور آڈٹ۔ - [ایمبیڈنگز (embeddings)](https://broadrouter.com/ur/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): ایمبیڈنگز فلوٹنگ پوائنٹ اعداد کی فہرستیں (ویکٹرز) ہیں جو کوئی ماڈل متن، تصویر یا دوسرے ان پٹ کی نمائندگی کے لیے بناتا ہے۔ ملتے جلتے معنی والے ان پٹس کے ویکٹرز ایک دوسرے کے قریب ہوتے ہیں، اس لیے ویکٹرز کا فاصلہ سرچ نتائج کی درجہ بندی، دستاویزات کی گروہ بندی یا متعلقہ چیزوں کی تلاش میں کام آتا ہے۔ - [Eval (LLM ایویلیوایشن)](https://broadrouter.com/ur/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Eval (evaluation کا مخفف) کسی لینگویج ماڈل یا پرامپٹ کا دہرایا جا سکنے والا ٹیسٹ ہے: ان پٹس کا ایک مقررہ سیٹ، اچھے آؤٹ پٹ کی وضاحت، اور ایک گریڈر (کوڈ، انسان یا دوسرا ماڈل) جو ہر آؤٹ پٹ کو اسکور کرتا ہے، تاکہ مختلف ورژنز کا موازنہ ہو سکے۔ - [ایکسپونینشل بیک آف (exponential backoff)](https://broadrouter.com/ur/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): ایکسپونینشل بیک آف دوبارہ کوشش کی ایک حکمتِ عملی ہے جس میں کلائنٹ ہر ناکام کوشش کے بعد، دوبارہ کوشش سے پہلے، زیادہ دیر انتظار کرتا ہے، عموماً تاخیر کو ایک حد تک دگنا کرتے ہوئے۔ ہر انتظار میں رینڈم jitter شامل کرنے سے کئی کلائنٹس کی دوبارہ کوششیں بکھر جاتی ہیں تاکہ وہ بیک وقت سرور تک نہ پہنچیں۔ - [فیو شاٹ پرامپٹنگ (few-shot prompting)](https://broadrouter.com/ur/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): فیو شاٹ پرامپٹنگ وہ تکنیک ہے جس میں پرامپٹ اصل ان پٹ سے پہلے ان پٹ اور آؤٹ پٹ کی چند حل شدہ مثالیں دیتا ہے، تاکہ لینگویج ماڈل ان سے کام اور فارمیٹ سمجھ لے۔ ماڈل کے ویٹس نہیں بدلتے؛ زیرو شاٹ میں کوئی مثال نہیں ہوتی اور ون شاٹ میں ایک۔ - [فائن ٹیوننگ (fine-tuning)](https://broadrouter.com/ur/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): فائن ٹیوننگ پہلے سے تربیت یافتہ لینگویج ماڈل کو کسی چھوٹے، کام کے لحاظ سے مخصوص ڈیٹاسیٹ، جیسے مثالی گفتگو، پر مزید تربیت دینے کا عمل ہے، تاکہ اس کے ویٹس بدلیں اور وہ ہر پرامپٹ میں لمبی ہدایات یا مثالوں کے بغیر وہ کام یا انداز زیادہ بھروسے سے انجام دے۔ - [finish_reason (جواب رکنے کی وجہ)](https://broadrouter.com/ur/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason، chat completion جواب کی ہر choice میں ایک فیلڈ ہے جو بتاتا ہے کہ ماڈل نے جنریٹ کرنا کیوں روکا۔ OpenAI فارمیٹ میں اس کی قدریں stop، length، tool_calls، content_filter اور متروک (deprecated) function_call ہیں۔ Anthropic کا Messages API یہی بات stop_reason میں بتاتا ہے۔ - [مقررہ قیمت والا LLM پلان (flat-rate LLM plan)](https://broadrouter.com/ur/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): مقررہ قیمت والا LLM پلان کسی لینگویج ماڈل API تک رسائی کے لیے فی ٹوکن بل بنانے کے بجائے ایک مدت، جیسے ایک مہینے، کی مقررہ قیمت لیتا ہے۔ ایسے پلان عموماً کچھ اور محدود کرتے ہیں، جیسے concurrency، رفتار، ماڈلز یا استعمال کی ونڈوز، اس لیے حدیں قیمت جتنی ہی اہم ہیں۔ - [FP8 (8 بٹ فلوٹنگ پوائنٹ)](https://broadrouter.com/ur/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 ڈیپ لرننگ میں استعمال ہونے والا 8 بٹ فلوٹنگ پوائنٹ نمبر فارمیٹ ہے، جس کی دو انکوڈنگز ہیں: E4M3، جس میں 4 ایکسپوننٹ اور 3 مینٹیسا بٹس ہیں، اور E5M2، جس میں 5 ایکسپوننٹ اور 2 مینٹیسا بٹس ہیں۔ یہ 16 بٹ فارمیٹس کے مقابلے میں میموری آدھی کرتا ہے اور نئے GPUs پر براہِ راست چلتا ہے۔ - [فریکوئنسی اور پریزنس پینلٹی](https://broadrouter.com/ur/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): فریکوئنسی اور پریزنس پینلٹی سیمپلنگ سیٹنگز ہیں جو آؤٹ پٹ میں پہلے آ چکے ٹوکنز کا اسکور کم کرتی ہیں۔ پریزنس پینلٹی ٹوکن کے ایک بار بھی آنے پر لگتی ہے؛ فریکوئنسی پینلٹی ہر تکرار کے ساتھ بڑھتی ہے۔ OpenAI کے API میں دونوں کی رینج -2.0 سے 2.0 ہے، ڈیفالٹ 0۔ - [GDPR (جنرل ڈیٹا پروٹیکشن ریگولیشن)](https://broadrouter.com/ur/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): GDPR (General Data Protection Regulation، Regulation (EU) 2016/679) ذاتی ڈیٹا کی پروسیسنگ سے متعلق یورپی یونین کا قانون ہے۔ یہ 25 مئی 2018 سے نافذ ہے اور یورپی یونین میں قائم اور اس سے باہر کے کچھ اداروں پر لاگو ہوتا ہے۔ LLM APIs کے لیے یہ پروسیسر معاہدے، بین الاقوامی منتقلی اور ڈیٹا منیمائزیشن طے کرتا ہے۔ - [GGUF](https://broadrouter.com/ur/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF لینگویج ماڈلز محفوظ کرنے کا ایک بائنری فائل فارمیٹ ہے، جو llama.cpp جیسے GGML پر مبنی رن ٹائمز پر انفرینس کے لیے بنا ہے۔ ایک GGUF فائل میں ویٹس، اکثر کوانٹائزڈ، اور ماڈل لوڈ کرنے کا ضروری میٹا ڈیٹا ہوتا ہے، جیسے ٹوکنائزر اور آرکیٹیکچر کی تفصیلات، اور اسے تیز لوڈنگ کے لیے میموری میپ کیا جا سکتا ہے۔ - [GPU گھنٹہ (GPU-hour)](https://broadrouter.com/ur/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): GPU گھنٹہ ایک گرافکس پروسیسر کا ایک گھنٹے تک استعمال ہے۔ کلاؤڈ فراہم کنندگان کرائے کے GPU انسٹینسز کا بل گھنٹے یا سیکنڈ کے حساب سے بناتے ہیں، اور محققین کسی کام کا کمپیوٹ GPU گھنٹوں میں بتاتے ہیں۔ آپ مشین کے وقت کی ادائیگی کرتے ہیں، چاہے وہ مصروف ہو یا نہیں۔ - [گراؤنڈنگ (grounding)](https://broadrouter.com/ur/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): گراؤنڈنگ لینگویج ماڈل کے آؤٹ پٹ کو معلومات کے مخصوص، قابلِ تصدیق ذرائع، جیسے فراہم کردہ دستاویزات، سرچ نتائج یا ڈیٹابیس ریکارڈز، سے جوڑنے اور ماڈل کو انہی سے جواب دینے کی ہدایت کرنے کا طریقہ ہے۔ گراؤنڈڈ جواب کا ماخذ تلاش کیا جا سکتا ہے اور اس میں گھڑے ہوئے حقائق کم ہوتے ہیں۔ - [گارڈریلز (LLM guardrails)](https://broadrouter.com/ur/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): گارڈریلز وہ جانچیں ہیں جو کوئی ایپلیکیشن لینگویج ماڈل کال کے گرد چلاتی ہے تاکہ وہ مقررہ حدود میں رہے۔ ان پٹ گارڈریلز ماڈل میں جانے والی چیز پرکھتے ہیں، آؤٹ پٹ گارڈریلز باہر آنے والی، اور ٹول گارڈریلز کارروائی چلنے سے پہلے اسے جانچتے ہیں۔ جانچ ناکام ہو تو درخواست رکتی، بدلتی یا نشان زد ہوتی ہے۔ - [ہیلوسینیشن (hallucination)](https://broadrouter.com/ur/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): ہیلوسینیشن، لینگویج ماڈل کا ایسا آؤٹ پٹ ہے جو سننے میں درست لگتا ہے مگر غلط، گھڑا ہوا یا دیے گئے ان پٹ سے غیر ثابت شدہ ہوتا ہے، جیسے جعلی حوالہ، غلط عدد یا من گھڑت API۔ یہ اس لیے ہوتا ہے کہ ماڈل حقائق جانچنے کے بجائے ممکنہ متن کی پیش گوئی کرتا ہے۔ - [ہمزہ (hamza)](https://broadrouter.com/ur/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): ہمزہ (ء) گلوٹل اسٹاپ کے لیے عربی حرف ہے۔ یہ اکیلا یا کسی کرسی پر لکھا جاتا ہے: الف (أ إ)، واؤ (ؤ) یا یاء (ئ)، اور آ الف کو مد کے ساتھ جوڑتا ہے۔ ٹائپ شدہ متن میں یہ یکساں نہیں لکھا جاتا، سو سرچ انجن اور اسپیچ اسکورنگ اکثر الف والی شکلوں کو ا میں بدل دیتے ہیں۔ - [ہائبرڈ سرچ (hybrid search)](https://broadrouter.com/ur/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): ہائبرڈ سرچ ایک ریٹریول طریقہ ہے جو ایک ہی کوئری پر BM25 جیسی کی ورڈ سرچ اور embeddings پر ویکٹر سرچ دونوں چلاتا ہے، پھر دونوں رینک شدہ فہرستوں کو ایک میں ملاتا ہے، اکثر reciprocal rank fusion (RRF) یا normalized اسکورز کے weighted sum سے۔ - [Idempotency key (آئیڈیمپوٹینسی کلید)](https://broadrouter.com/ur/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Idempotency key ایک منفرد قدر ہے، عموماً UUID جیسی رینڈم ID، جسے کلائنٹ اسٹیٹ بدلنے والی درخواست، جیسے ادائیگی یا خریداری، کے ساتھ لگاتا ہے۔ اگر وہی درخواست اسی کلید کے ساتھ دوبارہ آئے تو سرور کام دو بار کرنے کے بجائے پہلا نتیجہ لوٹاتا ہے۔ - [امیج ٹوکنز (image tokens)](https://broadrouter.com/ur/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): امیج ٹوکنز وہ ان پٹ ٹوکنز ہیں جن سے کوئی vision language ماڈل پرامپٹ میں ایک تصویر کو ظاہر کرتا ہے۔ فراہم کنندگان انہیں تصویر کے سائز سے نکالتے ہیں، عموماً ری سائز کے بعد مقررہ patches یا tiles میں بانٹ کر، اور ٹیکسٹ ان پٹ ٹوکنز کی طرح بل کرتے ہیں، اس لیے بڑی تصویر مہنگی پڑتی ہے۔ - [انفرنس فراہم کنندہ (inference provider)](https://broadrouter.com/ur/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): انفرنس فراہم کنندہ وہ کمپنی ہے جو مشین لرننگ ماڈلز، اکثر اوپن ویٹ LLM، اپنے، لیز پر لیے یا کرائے کے GPUs پر چلاتی ہے اور API کے ذریعے ان تک رسائی بیچتی ہے، عموماً فی ٹوکن، فی درخواست یا فی وقت کی اکائی قیمت پر۔ سرونگ، اسکیلنگ اور ہارڈ ویئر وہی سنبھالتی ہے۔ - [ان پٹ اور آؤٹ پٹ ٹوکن (input and output tokens)](https://broadrouter.com/ur/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): ان پٹ ٹوکن وہ ٹوکن ہیں جو لینگویج ماڈل کسی درخواست میں پڑھتا ہے، جن میں سسٹم پرامپٹ، گفتگو کی تاریخ، ٹول ڈیفینیشنز اور تصاویر شامل ہیں۔ آؤٹ پٹ ٹوکن وہ ہیں جو وہ جواب میں بناتا ہے، ریزننگ سمیت۔ LLM API عموماً دونوں کی قیمت الگ رکھتے ہیں، اور آؤٹ پٹ عموماً مہنگا ہوتا ہے۔ - [INT4 (4 بٹ کوانٹائزیشن)](https://broadrouter.com/ur/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 ایک 4-bit integer نمبر فارمیٹ ہے جو 16 سطحیں ظاہر کر سکتا ہے۔ LLM انفرنس میں اس کا مطلب عموماً 4-bit weight-only کوانٹائزیشن (W4A16) ہوتا ہے: weights کو scale factors کے ساتھ INT4 میں رکھا جاتا ہے اور حساب کے دوران 16-bit میں پھیلایا جاتا ہے، جس سے 16-bit weights کی تقریباً ایک چوتھائی میموری لگتی ہے۔ - [IP allow list (اجازت یافتہ IP فہرست)](https://broadrouter.com/ur/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): IP allow list ایک سیکیورٹی کنٹرول ہے جو صرف فہرست میں دیے گئے IP پتوں یا CIDR رینجز سے آئی درخواستیں قبول کرتا ہے اور باقی سب کو رد کر دیتا ہے۔ کسی API پر یہ لیک ہوئی کلید سے ہونے والا نقصان محدود کرتا ہے، کیونکہ فہرست سے باہر کے نیٹ ورکس سے کلید کام نہیں کرتی۔ - [جیل بریک (LLM jailbreak)](https://broadrouter.com/ur/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): جیل بریک ایک ایسا ان پٹ ہے جو کسی لینگویج ماڈل سے اس کی سیفٹی ٹریننگ یا اسے چلانے والی ایپ کے اصولوں کو نظرانداز کروانے کے لیے بنایا جاتا ہے، تاکہ وہ ایسا مواد دے یا ایسے کام کرے جن سے وہ عموماً انکار کرتا۔ OWASP جیل بریکنگ کو prompt injection کی ایک شکل مانتا ہے۔ - [JSON موڈ (JSON mode)](https://broadrouter.com/ur/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): JSON موڈ، LLM API کی ایک سیٹنگ ہے، OpenAI طرز کے APIs میں response_format {"type": "json_object"}، جو ماڈل کے جواب کو درست اور پارس ہونے کے قابل JSON بناتی ہے۔ JSON Schema والے اسٹرکچرڈ آؤٹ پٹ کے برعکس، یہ ضمانت نہیں دیتی کہ آبجیکٹ میں کون سی keys، ٹائپس یا قدریں ہوں گی۔ - [JSON Schema (JSON اسکیما)](https://broadrouter.com/ur/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema، JSON میں لکھی ایک لغت ہے جو JSON ڈیٹا کی ساخت بیان کرتی اور جانچتی ہے: ٹائپس، لازمی فیلڈز، قابلِ قبول قدریں اور نیسٹنگ۔ موجودہ ورژن 2020-12 ہے۔ LLM APIs اس سے ٹول پیرامیٹرز طے کرتے ہیں، اور وہ عین شکل بھی جس کی اسٹرکچرڈ آؤٹ پٹ کو پابندی کرنی ہے۔ - [KV cache (کی ویلیو کیش)](https://broadrouter.com/ur/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): KV cache وہ میموری ہے جہاں ٹرانسفارمر لینگویج ماڈل پہلے سے پروسیس شدہ ہر ٹوکن کے لیے نکالی گئی attention keys اور values رکھتا ہے، تاکہ ہر نیا ٹوکن پوری sequence دوبارہ گننے کے بجائے انہیں استعمال کرے۔ یہ context کی لمبائی کے ساتھ بڑھتا ہے اور اکثر طے کرتا ہے کہ GPU کتنی درخواستیں سنبھالے۔ - [لیٹنسی (latency)](https://broadrouter.com/ur/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): لینگویج ماڈل API کے لیے لیٹنسی وہ وقت ہے جو کسی درخواست کے بھیجے جانے سے آخری آؤٹ پٹ ٹوکن موصول ہونے تک لگتا ہے۔ یہ پہلے ٹوکن تک کے وقت اور جنریشن ٹائم کا مجموعہ ہے، اور عموماً اوسط کے بجائے p50 اور p95 جیسے پرسنٹائلز میں بتائی جاتی ہے۔ - [LLM-as-a-judge (LLM جج)](https://broadrouter.com/ur/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge ایک جانچ کا طریقہ ہے جس میں ایک لینگویج ماڈل دوسرے ماڈلز کے آؤٹ پٹس کو گریڈ کرتا ہے، ایک جواب کو rubric پر اسکور دے کر یا دو جوابوں میں سے بہتر چن کر۔ یہ کم لاگت پر انسانی فیصلے کے قریب پہنچتا ہے، مگر اس میں معروف biases ہیں، جیسے پہلے یا لمبے جواب کو ترجیح دینا۔ - [LLM بینچ مارک (LLM benchmark)](https://broadrouter.com/ur/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): LLM بینچ مارک کاموں کا ایک مقررہ، عموماً پبلک سیٹ ہے جس کا اسکورنگ طریقہ طے ہوتا ہے، اور جس سے لینگویج ماڈلز کو کسی ایک مہارت، جیسے معلوماتی سوالات، کوڈنگ یا اسپیچ ریکگنیشن، پر ناپا اور موازنہ کیا جاتا ہے۔ بینچ مارک اسکور اس ٹیسٹ پر نسبتی صلاحیت دکھاتا ہے، آپ کے اپنے ورک لوڈ پر کارکردگی نہیں۔ - [LLM راؤٹر (LLM router)](https://broadrouter.com/ur/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): LLM راؤٹر ایک سافٹ ویئر یا سروس ہے جو ایپ اور ایک یا زیادہ لینگویج ماڈل APIs کے درمیان بیٹھتی ہے اور ہر درخواست کے لیے طے کرتی ہے کہ اسے کون سا ماڈل یا فراہم کنندہ سنبھالے گا، لاگت، latency، دستیابی یا کام کی قسم جیسے اصولوں کی بنیاد پر۔ - [لاگ پروبز (logprobs)](https://broadrouter.com/ur/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): لاگ پروبز وہ لاگ احتمالات ہیں جو لینگویج ماڈل نے اپنے جنریٹ کیے گئے ٹوکنز کو دیے۔ انہیں لوٹانے والا API ہر آؤٹ پٹ ٹوکن کے لیے دکھاتا ہے کہ ماڈل نے اسے کتنا ممکن سمجھا، اور چاہیں تو اس جگہ کے سب سے ممکنہ متبادل ٹوکن بھی۔ - [LoRA (low-rank adaptation)](https://broadrouter.com/ur/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation) ایک پیرامیٹر کفایتی fine-tuning طریقہ ہے جو پہلے سے تربیت یافتہ ماڈل کے weights کو منجمد کر دیتا ہے اور منتخب layers کے لیے چھوٹی low-rank میٹرکس کا ایک جوڑا تربیت دیتا ہے۔ سیکھی گئی تبدیلی چھوٹے adapter کے طور پر رکھی جا سکتی ہے یا weights میں ملائی جا سکتی ہے، انفرنس latency بڑھائے بغیر۔ - [Lost in the middle (درمیان میں کھو جانا)](https://broadrouter.com/ur/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle لینگویج ماڈلز کا مشاہدہ شدہ رجحان ہے کہ وہ متعلقہ معلومات کا سب سے اچھا استعمال تب کرتے ہیں جب وہ لمبے ان پٹ کے شروع یا آخر کے قریب ہو، اور سب سے برا تب جب وہ درمیان میں ہو۔ یہ نام Liu اور ساتھیوں کی 2023 کی تحقیق نے دیا۔ - [max_tokens (آؤٹ پٹ ٹوکن کی حد)](https://broadrouter.com/ur/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens ایک LLM API پیرامیٹر ہے جو طے کرتا ہے کہ ماڈل ایک جواب میں زیادہ سے زیادہ کتنے ٹوکن بنا سکتا ہے۔ حد پر پہنچتے ہی جنریشن رک جاتی ہے اور جواب کو کٹا ہوا قرار دیا جاتا ہے، مثلاً finish_reason کو "length" پر سیٹ کر کے۔ - [MCP (Model Context Protocol)](https://broadrouter.com/ur/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol)، JSON-RPC 2.0 پر بنا ایک اوپن پروٹوکول ہے جو طے کرتا ہے کہ LLM ایپلیکیشنز بیرونی ٹولز اور ڈیٹا سے کیسے جڑیں۔ MCP سرور ٹولز، ریسورسز اور پرامپٹس فراہم کرتا ہے؛ ہوسٹ ایپلیکیشن کے اندر کا MCP کلائنٹ انہیں تلاش کرتا ہے اور ماڈل کی ٹول کالز سرور تک پہنچاتا ہے۔ - [Mixture of experts (MoE)](https://broadrouter.com/ur/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Mixture of experts (MoE) ایک نیورل نیٹ ورک ڈیزائن ہے جس میں ایک layer میں کئی expert ذیلی نیٹ ورکس ہوتے ہیں اور ایک سیکھا ہوا router ہر ٹوکن کو ان میں سے صرف چند کے پاس بھیجتا ہے۔ MoE لینگویج ماڈل کے کل پیرامیٹرز بہت ہوتے ہیں، لیکن ہر ٹوکن پر ایک چھوٹا active حصہ ہی استعمال کرتا ہے۔ - [ماڈل ڈیپریکیشن (model deprecation)](https://broadrouter.com/ur/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): ماڈل ڈیپریکیشن تب ہوتا ہے جب کوئی API فراہم کنندہ اعلان کرتا ہے کہ ایک ماڈل دستیاب نہیں رہے گا، عموماً ریٹائرمنٹ یا شٹ ڈاؤن کی تاریخ اور تجویز کردہ متبادل کے ساتھ۔ اس تاریخ کے بعد اس ماڈل کی درخواستیں ناکام ہوتی ہیں، اس لیے ایپس کو پہلے ہی کسی دوسری ماڈل ID پر منتقل ہونا ہوتا ہے۔ - [ماڈل فال بیک (model fallback)](https://broadrouter.com/ur/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): ماڈل فال بیک قابلِ اعتمادی کا ایک پیٹرن ہے جس میں ناکام LLM درخواست ایک ترتیب وار فہرست کے کسی دوسرے ماڈل یا فراہم کنندہ پر دوبارہ بھیجی جاتی ہے۔ ٹرگر کوئی ایرر، ریٹ لمٹ، context-length ایرر یا content-policy انکار ہو سکتا ہے۔ یہ آپ کے کلائنٹ میں یا APIs کے آگے لگے گیٹ وے میں چلتا ہے۔ - [ماڈل لائسنس (model license)](https://broadrouter.com/ur/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): ماڈل لائسنس وہ قانونی شرائط ہیں جن کے تحت کسی ماڈل کے weights، کوڈ اور آؤٹ پٹ استعمال، تبدیل اور شیئر کیے جا سکتے ہیں۔ اوپن ویٹ ماڈلز Apache 2.0 یا MIT جیسے عام سافٹ ویئر لائسنس، یا کسٹم ماڈل لائسنس استعمال کرتے ہیں، جو بڑی کمرشل سروسز پر حد جیسی شرائط جوڑ سکتے ہیں۔ - [جدید معیاری عربی (Modern Standard Arabic، MSA)](https://broadrouter.com/ur/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): Modern Standard Arabic (MSA) عربی کی معیاری شکل ہے جو پوری عرب دنیا میں تحریر، خبروں، تعلیم اور رسمی گفتگو میں استعمال ہوتی ہے۔ یہ کسی کی روزمرہ مادری زبان نہیں: لوگ مصری یا خلیجی عربی جیسے علاقائی لہجے بولتے ہیں اور رسمی مواقع پر MSA اختیار کرتے ہیں۔ - [ملٹی ایجنٹ سسٹم (multi-agent system)](https://broadrouter.com/ur/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): ملٹی ایجنٹ سسٹم ایک LLM ایپلیکیشن ہے جس میں کئی ایجنٹ، ہر ایک اپنے پرامپٹ، ٹولز اور کانٹیکسٹ کے ساتھ، ایک ہی کام کے مختلف حصوں پر کام کرتے ہیں۔ عام ڈیزائن orchestrator-worker ہے: ایک lead ایجنٹ منصوبہ بناتا ہے، worker ایجنٹس متوازی شروع کرتا ہے اور ان کے نتائج یکجا کرتا ہے۔ - [ملٹی موڈل AI (multimodal AI)](https://broadrouter.com/ur/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): ملٹی موڈل AI ایسا ماڈل یا سسٹم ہے جو ایک سے زیادہ قسم کے ڈیٹا، جیسے متن، تصاویر، ویڈیو اور آڈیو، کے ساتھ کام کرتا ہے۔ کوئی ملٹی موڈل ماڈل تصویر اور سوال ساتھ پڑھ سکتا ہے، یا آواز لے کر متن میں جواب دے سکتا ہے۔ vision-language ماڈل اس کی سب سے عام قسم ہے۔ - [نیمڈ اینٹیٹی ریکگنیشن (NER)](https://broadrouter.com/ur/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): Named entity recognition (NER) متن میں نام تلاش کرنے اور ہر نام کو شخص، ادارہ یا مقام جیسی کسی قسم کا لیبل دینے کا کام ہے۔ روایتی سسٹم ہر ٹوکن کو ٹیگ کرتے ہیں؛ LLMs سے اکثر entities کی فہرست JSON میں مانگی جاتی ہے، جسے پھر ماخذ متن سے ملا کر جانچا جاتا ہے۔ - [OCR (optical character recognition)](https://broadrouter.com/ur/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): OCR (optical character recognition) وہ ٹیکنالوجی ہے جو کسی تصویر، اسکین یا PDF میں چھپے یا ہاتھ سے لکھے متن کو پہچان کر مشین کے پڑھنے کے قابل حروف میں بدلتی ہے۔ روایتی OCR متن اور صفحے پر اس کی جگہ دیتا ہے؛ vision-language ماڈل متن پڑھ کر اس کے بارے میں براہِ راست سوالوں کے جواب دے سکتے ہیں۔ - [Ogg Opus](https://broadrouter.com/ur/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus وہ آڈیو ہے جو Opus codec (RFC 6716) سے انکوڈ ہو کر Ogg کنٹینر (RFC 7845) میں رکھا جاتا ہے۔ یہ کم bitrate پر گفتگو اور انٹرایکٹو آڈیو کے لیے بنا ہے، اور WhatsApp وائس نوٹس کے لیے یہی فارمیٹ استعمال کرتا ہے، عموماً .ogg یا .opus فائلوں کی صورت میں۔ - [اوپن ویٹ ماڈل (open-weight model)](https://broadrouter.com/ur/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): اوپن ویٹ ماڈل ایسا مشین لرننگ ماڈل ہے جس کے ٹرین شدہ weights ہر کسی کے ڈاؤن لوڈ کرنے اور چلانے کے لیے شائع کیے جاتے ہیں، عموماً شرائط والے لائسنس کے ساتھ۔ مکمل اوپن سورس AI اس کے علاوہ ٹریننگ کوڈ اور ٹریننگ ڈیٹا کی تفصیل بھی جاری کرتا ہے، تاکہ دوسرے ماڈل کو سمجھ اور دوبارہ بنا سکیں۔ - [OpenAI-compatible API](https://broadrouter.com/ur/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): OpenAI-compatible API ایسا HTTP API ہے جو OpenAI API والی ہی درخواست اور ریسپانس کی ساخت قبول کرتا ہے، جیسے POST /v1/chat/completions، تاکہ OpenAI کے SDK اور ٹولز صرف base URL، API کلید اور ماڈل کا نام بدل کر اسے استعمال کر سکیں۔ - [متوازی ٹول کالز (parallel tool calls)](https://broadrouter.com/ur/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Parallel tool calls تب ہوتی ہیں جب ماڈل ایک ہی جواب میں ایک سے زیادہ ٹول کالز واپس کرتا ہے، مثلاً ایک ساتھ تین فائلیں پڑھنا۔ OpenAI فارمیٹ میں یہ tool_calls کی کئی اندراجات کے طور پر آتی ہیں، اور parallel_tool_calls کو false کرنے پر ماڈل ہر ٹرن میں صفر یا ایک کال تک محدود رہتا ہے۔ - [استعمال کے مطابق ادائیگی (pay-as-you-go)](https://broadrouter.com/ur/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): استعمال کے مطابق ادائیگی (pay-as-you-go) قیمت طے کرنے کا ایک ماڈل ہے جس میں آپ صرف اتنا دیتے ہیں جتنا استعمال کرتے ہیں، ٹوکن، درخواستوں یا آڈیو منٹ جیسی اکائیوں میں ناپ کر، بغیر سبسکرپشن یا کم از کم وابستگی کے۔ زیادہ تر LLM API اسی طرح بل بناتے ہیں، پری پیڈ بیلنس یا وقتاً فوقتاً بننے والی انوائس سے۔ - [PDPL (سعودی اور UAE کے ڈیٹا تحفظ کے قوانین)](https://broadrouter.com/ur/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) سعودی عرب کے ذاتی ڈیٹا قانون کا عام نام ہے، جو Royal Decree M/19 سے جاری ہوا اور SDAIA نافذ کرتا ہے، اور UAE کے Federal Decree-Law No. 45 of 2021 کا بھی۔ دونوں ذاتی ڈیٹا کی پروسیسنگ اور بیرونِ ملک منتقلی کی شرائط طے کرتے ہیں، غیر ملکی LLM APIs کو بھیجے پرامپٹس سمیت۔ - [PII (personally identifiable information)](https://broadrouter.com/ur/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): PII (personally identifiable information) وہ معلومات ہے جس سے کسی شخص کی شناخت الگ کی یا کھوجی جا سکے، جیسے نام، ID نمبر یا بایومیٹرک ریکارڈ، نیز اس شخص سے جڑی یا جوڑی جا سکنے والی دوسری معلومات۔ GDPR کی وسیع تر اصطلاح personal data ہے۔ LLM میں PII redaction اسے سسٹم سے نکلنے سے پہلے پرامپٹس سے ہٹاتا ہے۔ - [Prefill اور decode](https://broadrouter.com/ur/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill اور decode، LLM انفرنس کے دو مراحل ہیں۔ Prefill میں ماڈل پورا پرامپٹ متوازی طور پر پروسیس کر کے اپنا KV cache بھرتا ہے؛ decode میں وہ اسی cache کو دوبارہ استعمال کرتے ہوئے جواب ایک ایک ٹوکن کر کے بناتا ہے۔ Prefill زیادہ تر پہلے ٹوکن تک کا وقت طے کرتا ہے؛ decode آؤٹ پٹ کی رفتار۔ - [پری پیڈ کریڈٹ (prepaid credit)](https://broadrouter.com/ur/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): پری پیڈ کریڈٹ وہ رقم ہے جو استعمال سے پہلے API اکاؤنٹ میں ڈالی جاتی ہے۔ ہر درخواست کی لاگت بیلنس سے کٹتی ہے، اور بیلنس ختم ہونے پر میٹر والی درخواستیں تب تک رک جاتی ہیں جب تک آپ ٹاپ اپ نہ کریں۔ اس سے خرچ اتنا ہی رہتا ہے جتنا آپ نے جمع کیا ہے۔ - [پرامپٹ انجینئرنگ (prompt engineering)](https://broadrouter.com/ur/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): پرامپٹ انجینئرنگ، لینگویج ماڈل کو دی جانے والی ہدایات کو لکھنے، ترتیب دینے اور ٹیسٹ کرنے کا عمل ہے، تاکہ ماڈل مسلسل آپ کی ضروریات کے مطابق آؤٹ پٹ دے۔ اس میں رول پیغامات، واضح ہدایات، مثالیں، فارمیٹنگ اور ٹیسٹ کیسز پر نتائج کی پیمائش شامل ہے۔ - [پرامپٹ انجیکشن (prompt injection)](https://broadrouter.com/ur/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): Prompt injection ایک حملہ ہے جس میں لینگویج ماڈل کو دیا گیا متن، چاہے صارف نے لکھا ہو یا کسی ویب صفحے، ای میل، فائل یا ٹول کے نتیجے میں چھپا ہو، ماڈل کا رویہ ڈویلپر کے ارادے کے خلاف بدل دیتا ہے۔ OWASP اپنی LLM ایپلیکیشنز کی Top 10 فہرست میں اسے پہلے نمبر (LLM01:2025) پر رکھتا ہے۔ - [کوانٹائزیشن (quantization)](https://broadrouter.com/ur/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): کوانٹائزیشن، نیورل نیٹ ورک کے اعداد، بنیادی طور پر اس کے weights، کو 16-bit کے بجائے 8-bit یا 4-bit جیسے کم پریسیژن فارمیٹ میں رکھنے کا عمل ہے، تاکہ ماڈل کم میموری لے اور اکثر تیز چلے۔ اچھے طریقے درستی میں ہونے والے نقصان کو کم رکھتے ہیں۔ - [RAG (retrieval-augmented generation)](https://broadrouter.com/ur/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG (retrieval-augmented generation) ایک طریقہ ہے جس میں ایپلیکیشن پہلے کسی بیرونی اسٹور، جیسے سرچ یا ویکٹر انڈیکس، سے سوال سے متعلقہ حصے نکالتی ہے، اور پھر انہیں پرامپٹ میں جوڑتی ہے، تاکہ لینگویج ماڈل صرف اپنی یادداشت کے بجائے اس متن سے جواب دے سکے۔ - [ریٹ لمٹ (rate limit)](https://broadrouter.com/ur/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): ریٹ لمٹ وہ حد ہے جو کوئی API طے کرتا ہے کہ ایک کلائنٹ کسی مدت میں اسے کتنا استعمال کر سکتا ہے، جیسے فی منٹ درخواستیں یا فی منٹ ٹوکن۔ حد سے زیادہ درخواستیں رد کر دی جاتی ہیں، عموماً HTTP 429 Too Many Requests کے ساتھ، جب تک استعمال دوبارہ حد سے نیچے نہ آ جائے۔ - [ReAct (reason and act)](https://broadrouter.com/ur/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (reason and act) ایک پرامپٹنگ پیٹرن ہے، جسے Yao اور ساتھیوں نے 2022 میں متعارف کرایا، جس میں لینگویج ماڈل reasoning traces کو سرچ یا ٹول کالز جیسی کارروائیوں کے ساتھ باری باری چلاتا ہے۔ وہ ہر observation پڑھتا ہے، اپنا منصوبہ اپ ڈیٹ کرتا ہے، اور حتمی جواب تک یہی دہراتا ہے۔ - [ریزننگ ٹوکن (reasoning tokens)](https://broadrouter.com/ur/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): ریزننگ ٹوکن وہ ٹوکن ہیں جو ریزننگ ماڈل اپنا نظر آنے والا جواب لکھنے سے پہلے مسئلہ حل کرنے کے لیے بناتا ہے۔ ان کا بل عموماً آؤٹ پٹ ٹوکنز کی طرح بنتا ہے، یہ آؤٹ پٹ کی حد اور کانٹیکسٹ ونڈو میں شمار ہوتے ہیں، اور API انہیں چھپا سکتا ہے یا ان کا خلاصہ دے سکتا ہے۔ - [ریکوئسٹ ٹائم آؤٹ (request timeout)](https://broadrouter.com/ur/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): ریکوئسٹ ٹائم آؤٹ وہ زیادہ سے زیادہ وقت ہے جتنا کوئی HTTP کلائنٹ سرور کا انتظار کرتا ہے، پھر درخواست چھوڑ دیتا ہے۔ کلائنٹ عموماً اسے کنکشن کھولنے کے connect ٹائم آؤٹ اور ڈیٹا کے انتظار والے read ٹائم آؤٹ میں بانٹتے ہیں۔ LLM کالز کو کھلے read ٹائم آؤٹ چاہییں، کیونکہ لمبے جوابات میں وقت لگتا ہے۔ - [ری رینکنگ (reranking)](https://broadrouter.com/ur/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Reranking، retrieval کا دوسرا مرحلہ ہے جس میں ایک زیادہ درست لیکن سست ماڈل، عموماً cross-encoder، پہلی تلاش سے آئے ہر query اور دستاویز کے جوڑے کو اسکور کر کے ان کی ترتیب بدلتا ہے، تاکہ صرف سب سے متعلقہ چند ہی لینگویج ماڈل تک پہنچیں۔ - [دائیں سے بائیں متن (right-to-left، RTL text)](https://broadrouter.com/ur/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): دائیں سے بائیں (RTL) متن وہ تحریر ہے، جیسے عربی یا عبرانی، جو دائیں سے بائیں دکھائی جاتی ہے۔ کمپیوٹر اسے منطقی پڑھنے کی ترتیب میں محفوظ کرتے ہیں، اور جب RTL متن بائیں سے دائیں متن، ہندسوں، ماڈل ناموں یا کوڈ کے ساتھ ملتا ہے تو Unicode Bidirectional Algorithm دکھانے کی ترتیب طے کرتا ہے۔ - [سیمپل ریٹ (sample rate)](https://broadrouter.com/ur/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): سیمپل ریٹ ہر سیکنڈ میں ریکارڈ ہونے والے آڈیو سیمپلز کی تعداد ہے، جو ہرٹز (Hz) میں ناپی جاتی ہے۔ اسپیچ ریکگنیشن ماڈل عموماً 16,000 Hz (16 kHz) مونو آڈیو کی توقع رکھتے ہیں۔ بٹ ڈیپتھ اور چینلز کی تعداد کے ساتھ مل کر سیمپل ریٹ طے کرتا ہے کہ غیر کمپریسڈ آڈیو فائل فی منٹ کتنی بڑی ہوگی۔ - [seed پیرامیٹر (seed parameter)](https://broadrouter.com/ur/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): seed پیرامیٹر ایک عدد صحیح ہے جو آپ درخواست کے ساتھ بھیجتے ہیں تاکہ وہی درخواست دہرانے پر API وہی ٹوکن سیمپل کرنے کی کوشش کرے۔ OpenAI اسے best effort کہتا ہے: determinism کی ضمانت نہیں، اور system_fingerprint فیلڈ دکھاتا ہے کہ بیک اینڈ کنفیگریشن کب بدلی۔ - [سرور لیس انفرنس (serverless inference)](https://broadrouter.com/ur/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): سرور لیس انفرنس مشین لرننگ ماڈل چلانے کا ایسا طریقہ ہے جس میں پلیٹ فارم اینڈ پوائنٹ کے پیچھے کمپیوٹ خود بخود شروع، اسکیل اور بند کرتا ہے، اکثر فارغ ہونے پر صفر تک، اور مخصوص مشینوں کے بجائے استعمال کا بل لیتا ہے۔ اس کی قیمت ممکنہ کولڈ اسٹارٹ تاخیر ہے۔ - [SLA (سروس لیول ایگریمنٹ)](https://broadrouter.com/ur/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): SLA (سروس لیول ایگریمنٹ) فراہم کنندہ کی طرف سے سروس کی قابلِ پیمائش سطح کا معاہداتی وعدہ ہے، عموماً ماہانہ اپ ٹائم فیصد، اور اس کے پورا نہ ہونے پر طے شدہ تلافی، عموماً سروس کریڈٹس۔ SLA بتاتا ہے کہ ڈاؤن ٹائم کسے مانا جائے گا، کیا خارج ہے اور دعویٰ کیسے کریں۔ - [اسپیکیولیٹو ڈیکوڈنگ (speculative decoding)](https://broadrouter.com/ur/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): اسپیکیولیٹو ڈیکوڈنگ ایک انفرنس تکنیک ہے جس میں چھوٹا، تیز ڈرافٹ ماڈل اگلے کئی ٹوکنز تجویز کرتا ہے اور بڑا ٹارگٹ ماڈل انہیں ایک ہی forward pass میں جانچ کر قبول شدہ ٹوکنز رکھتا ہے۔ یہ آؤٹ پٹ کی ڈسٹری بیوشن بدلے بغیر جنریشن تیز کرتی ہے۔ - [اسپیچ ٹو ٹیکسٹ (speech-to-text)](https://broadrouter.com/ur/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): اسپیچ ٹو ٹیکسٹ، جسے automatic speech recognition (ASR) بھی کہتے ہیں، بولے گئے آڈیو کو تحریری متن میں بدلنے کی ٹیکنالوجی ہے۔ speech-to-text API آڈیو فائل یا اسٹریم لے کر ٹرانسکرپٹ لوٹاتا ہے، عموماً فی آڈیو منٹ قیمت پر۔ درستگی ورڈ ایرر ریٹ سے ناپی جاتی ہے اور لہجے، بولی اور ریکارڈنگ کے معیار کے ساتھ بدلتی ہے۔ - [اخراجات کی حد (spend limit)](https://broadrouter.com/ur/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): اخراجات کی حد وہ سقف ہے کہ کوئی API اکاؤنٹ، پروجیکٹ یا کلید ایک مدت میں، عموماً ایک مہینے میں، کتنا خرچ کر سکتی ہے۔ استعمال وہاں پہنچنے پر فراہم کنندہ مزید درخواستیں رد کر دیتا ہے، اکثر HTTP 429 یا 400 ایرر کے ساتھ، جب تک مدت ری سیٹ نہ ہو یا حد بڑھائی نہ جائے۔ - [SRT سب ٹائٹلز (SRT subtitles)](https://broadrouter.com/ur/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) ایک سادہ متن والا سب ٹائٹل فارمیٹ ہے۔ ہر کیپشن میں ایک ترتیبی نمبر، hh:mm:ss,mmm --> hh:mm:ss,mmm کی صورت میں لکھا شروع اور اختتام کا وقت، متن کی ایک یا زیادہ لائنیں اور ایک خالی لائن ہوتی ہے۔ WebVTT اس سے ملتا جلتا ویب فارمیٹ ہے جو HTML ویڈیو استعمال کرتا ہے۔ - [Server-sent events (SSE)](https://broadrouter.com/ur/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Server-sent events (SSE) ایک ویب معیار ہے جس میں سرور text/event-stream کانٹینٹ ٹائپ والے ایک دیر تک کھلے HTTP جواب پر کلائنٹ کو متنی events کی اسٹریم بھیجتا ہے۔ ہر event میں data: جیسی ایک یا زیادہ فیلڈ لائنیں ہوتی ہیں، اور وہ ایک خالی لائن پر ختم ہوتا ہے۔ - [اسٹاپ سیکوئنس (stop sequences)](https://broadrouter.com/ur/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): اسٹاپ سیکوئنس وہ اسٹرنگز ہیں جو آپ درخواست کے ساتھ بھیجتے ہیں تاکہ ماڈل ان میں سے کوئی بھی لکھتے ہی جنریشن روک دے۔ OpenAI اس فیلڈ کو stop کہتا ہے اور 4 سیکوئنس تک قبول کرتا ہے؛ Anthropic اسے stop_sequences کہتا ہے۔ ملا ہوا متن عموماً لوٹائے گئے آؤٹ پٹ میں شامل نہیں ہوتا۔ - [اسٹریمنگ (streaming)](https://broadrouter.com/ur/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): اسٹریمنگ، لینگویج ماڈل کے جواب کو آخر میں ایک ساتھ لوٹانے کے بجائے، بنتے وقت ہی چھوٹے چھوٹے ٹکڑوں (chunks) میں لوٹانے کا طریقہ ہے۔ OpenAI-compatible APIs، درخواست میں stream کو true سیٹ کرنے پر server-sent events کے ذریعے اسٹریم کرتے ہیں۔ - [اسٹرکچرڈ آؤٹ پٹ (structured output)](https://broadrouter.com/ur/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): اسٹرکچرڈ آؤٹ پٹ LLM API کا ایک فیچر ہے جو ماڈل کے جواب کو ڈویلپر کے طے کردہ فارمیٹ، عموماً JSON Schema، تک محدود کرتا ہے، تاکہ کوڈ اسے قابلِ اعتماد طریقے سے پارس کر سکے۔ یہ JSON mode سے زیادہ سخت ہے، جو درست JSON کی ضمانت دیتا ہے، لیکن کسی خاص ساخت کی نہیں۔ - [سب ایجنٹ (subagent)](https://broadrouter.com/ur/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): سب ایجنٹ ایک الگ ایجنٹ ہے جسے مرکزی ایجنٹ کسی ایک مخصوص کام کے لیے شروع کرتا ہے۔ یہ اپنی کانٹیکسٹ ونڈو، سسٹم پرامپٹ اور ٹول رسائی کے ساتھ چلتا ہے، کام کرتا ہے، اور مرکزی ایجنٹ کو صرف نتیجہ یا خلاصہ لوٹاتا ہے، جس سے مرکزی گفتگو مختصر رہتی ہے۔ - [سب پروسیسر (subprocessor)](https://broadrouter.com/ur/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): سب پروسیسر وہ تیسرا فریق ہے جسے پروسیسر کنٹرولر کی جانب سے ذاتی ڈیٹا پروسیس کرنے کے لیے رکھتا ہے۔ LLM API کے معاملے میں سب پروسیسرز میں GPU ہوسٹس، پارٹنر انفرنس فراہم کنندگان اور بیک اپ گنجائش شامل ہو سکتے ہیں۔ GDPR آرٹیکل 28 کے تحت انہیں کنٹرولر کی اجازت اور وہی ڈیٹا تحفظ کی ذمہ داریاں درکار ہیں۔ - [سسٹم پرامپٹ (system prompt)](https://broadrouter.com/ur/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): سسٹم پرامپٹ وہ ہدایتی متن ہے جو کوئی ایپلیکیشن گفتگو سے پہلے لینگویج ماڈل کو بھیجتی ہے، عموماً system یا developer رول والے پیغام کے طور پر۔ یہ ماڈل کا کردار، قواعد، لہجہ اور آؤٹ پٹ فارمیٹ طے کرتا ہے، اور ٹکراؤ کی صورت میں ماڈل کو صارف کے پیغامات کے بجائے اسی پر عمل کرنا ہوتا ہے۔ - [تائے مربوطہ (taa marbuta، ة)](https://broadrouter.com/ur/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): تائے مربوطہ (ة، U+0629) ایک عربی حرف ہے جو صرف لفظ کے آخر میں آتا ہے، اکثر مؤنث اسم اور صفت کی علامت کے طور پر۔ یہ دو نقطوں والی ہ (ه، U+0647) جیسا دکھتا ہے، اس لیے ٹائپ کیے گئے متن میں دونوں گڈمڈ ہو جاتے ہیں، اور کئی سرچ نارملائزرز ة کو ه میں بدل دیتے ہیں۔ - [ٹمپریچر (temperature)](https://broadrouter.com/ur/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): ٹمپریچر ایک سیمپلنگ پیرامیٹر ہے جو طے کرتا ہے کہ لینگویج ماڈل کا آؤٹ پٹ کتنا رینڈم ہوگا۔ ماڈل اپنے ٹوکن اسکورز کو احتمالات میں بدلنے سے پہلے ٹمپریچر سے تقسیم کرتا ہے: کم قدر پر سب سے ممکنہ ٹوکن غالب رہتا ہے، زیادہ قدر پر تقسیم ہموار ہو جاتی ہے اور کم ممکنہ ٹوکن زیادہ بار چنے جاتے ہیں۔ - [ٹیکسٹ ٹو اسپیچ (text-to-speech، TTS)](https://broadrouter.com/ur/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): ٹیکسٹ ٹو اسپیچ (TTS)، یا اسپیچ سنتھیسس، لکھے متن کو بولی گئی آڈیو میں بدلتا ہے۔ TTS API متن، ایک آواز اور MP3 یا WAV جیسا آؤٹ پٹ فارمیٹ لیتا ہے اور تیار کردہ آواز لوٹاتا ہے۔ یہ اسپیچ ٹو ٹیکسٹ کا الٹ ہے، جو آڈیو کو متن میں بدلتا ہے۔ - [تھروپٹ (throughput)](https://broadrouter.com/ur/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): LLM سرونگ میں تھروپٹ وہ کل کام ہے جو کوئی سسٹم تمام بیک وقت درخواستوں میں فی اکائی وقت مکمل کرتا ہے، عموماً آؤٹ پٹ ٹوکن فی سیکنڈ یا درخواستیں فی سیکنڈ میں ناپا جاتا ہے۔ یہ فی اسٹریم رفتار سے مختلف ہے، جو بتاتی ہے کہ ایک درخواست کو اس کے ٹوکن کتنی تیزی سے ملتے ہیں۔ - [پہلے ٹوکن تک کا وقت (TTFT، time to first token)](https://broadrouter.com/ur/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): Time to first token (TTFT) وہ وقت ہے جو لینگویج ماڈل کو درخواست بھیجنے سے پہلا آؤٹ پٹ ٹوکن آنے تک لگتا ہے۔ اس میں نیٹ ورک کا وقت، سرور پر قطار اور prefill شامل ہیں، یعنی وہ مرحلہ جس میں ماڈل پورا پرامپٹ پڑھتا ہے، اس لیے یہ پرامپٹ کی لمبائی اور لوڈ کے ساتھ بڑھتا ہے۔ - [ٹائم اسٹیمپس (اسپیچ ٹو ٹیکسٹ، timestamps)](https://broadrouter.com/ur/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): اسپیچ ٹو ٹیکسٹ میں ٹائم اسٹیمپس ٹرانسکرپٹ کے حصوں سے جڑے شروع اور اختتام کے اوقات ہیں، سیکنڈز میں۔ سیگمنٹ ٹائم اسٹیمپس کسی فقرے یا جملے کا احاطہ کرتے ہیں؛ ورڈ ٹائم اسٹیمپس ہر لفظ کا۔ یہ سب ٹائٹلز، ریکارڈنگز کے اندر سرچ اور متن کو بولنے والوں سے ملانے کے لیے ضروری ہیں۔ - [ٹوکن (token)](https://broadrouter.com/ur/glossary/token/) (translation of https://broadrouter.com/glossary/token/): ٹوکن متن کی وہ اکائی ہے جسے لینگویج ماڈل پڑھتا اور لکھتا ہے: کوئی لفظ، لفظ کا حصہ، رموزِ اوقاف کی کوئی علامت یا بائٹس کا سلسلہ، جو ماڈل کا ٹوکنائزر بناتا ہے۔ LLM API استعمال کو ٹوکنز میں گنتے ہیں، انہی میں حدیں لگاتے ہیں اور عموماً بل بھی بناتے ہیں۔ - [ٹوکنائزر (tokenizer)](https://broadrouter.com/ur/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): ٹوکنائزر وہ جزو ہے جو متن کو انٹیجر ٹوکن IDs کی اس ترتیب میں بدلتا ہے جسے لینگویج ماڈل پڑھتا ہے، اور تیار شدہ IDs کو واپس متن میں بدلتا ہے۔ ہر ماڈل اپنے ٹوکنائزر کے ساتھ آتا ہے، اس لیے ایک ہی متن مختلف ماڈلز پر مختلف تعداد میں ٹوکنز بنا سکتا ہے۔ - [ٹوکن فی سیکنڈ (tokens per second، TPS)](https://broadrouter.com/ur/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): ٹوکن فی سیکنڈ (TPS) وہ شرح ہے جس پر لینگویج ماڈل آؤٹ پٹ ٹوکن بناتا ہے، عموماً ایک درخواست کے پہلے ٹوکن کے بعد ناپی جاتی ہے۔ یہ طے کرتی ہے کہ جواب کتنی تیزی سے اسٹریم ہوگا۔ تمام بیک وقت درخواستوں کا مجموعہ سسٹم تھروپٹ ہے؛ پہلے ٹوکن سے پہلے کا انتظار الگ ناپا جاتا ہے۔ - [ٹول کالنگ (tool calling)](https://broadrouter.com/ur/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): ٹول کالنگ، جسے function calling بھی کہتے ہیں، LLM API کی وہ سہولت ہے جس میں ماڈل ڈویلپر کے طے کردہ کسی فنکشن کو چلانے کی ایک منظم درخواست، JSON آرگیومنٹس کے ساتھ، لوٹاتا ہے۔ ایپلیکیشن فنکشن چلا کر نتیجہ واپس بھیجتی ہے؛ ماڈل خود کبھی کوڈ نہیں چلاتا۔ - [tool_choice](https://broadrouter.com/ur/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice ایک درخواست پیرامیٹر ہے جو ٹول کے استعمال کو کنٹرول کرتا ہے۔ OpenAI کے Chat Completions میں یہ auto (ماڈل طے کرے)، required (ایک یا زیادہ ٹولز کال کرنا لازم)، none (کوئی ٹول نہیں) یا لازمی کال کے لیے کسی فنکشن کا نام ہو سکتا ہے۔ Anthropic کے مساوی آپشنز auto، any، tool اور none ہیں۔ - [ٹاپ پی (top-p، nucleus sampling)](https://broadrouter.com/ur/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): ٹاپ پی، جسے nucleus sampling بھی کہتے ہیں، ڈیکوڈنگ کا ایک طریقہ ہے جس میں ماڈل اگلا ٹوکن صرف سب سے ممکنہ ٹوکنز کے اس سب سے چھوٹے سیٹ سے چنتا ہے جن کا مجموعی احتمال 0.9 جیسی حد p تک پہنچ جائے، اور باقی کم ممکنہ ٹوکنز چھوڑ دیتا ہے۔ - [Unicode نارملائزیشن (Unicode normalization)](https://broadrouter.com/ur/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Unicode نارملائزیشن متن کو ایک معیاری شکل میں بدلتی ہے تاکہ ایک جیسی دکھنے والی اسٹرنگز ایک جیسے کوڈ پوائنٹس استعمال کریں۔ چار شکلیں NFD، NFC، NFKD اور NFKC ہیں؛ K والی شکلیں compatibility حروف، جیسے عربی presentation forms اور ligatures، کو بھی عام حروف میں بدل دیتی ہیں۔ - [وائس ایکٹیویٹی ڈیٹیکشن (VAD)](https://broadrouter.com/ur/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): وائس ایکٹیویٹی ڈیٹیکشن (VAD) سگنل پروسیسنگ کا ایک مرحلہ ہے جو آڈیو کے چھوٹے فریمز کو اسپیچ یا نان اسپیچ میں درجہ بند کرتا ہے۔ اسپیچ سسٹم اس سے خاموشی چھوڑتے ہیں، لمبی ریکارڈنگ کو وقفوں پر کاٹتے ہیں، پہچانتے ہیں کہ بولنے والا کب رکا، اور پس منظر کے شور کو ٹرانسکرائب ہونے سے بچاتے ہیں۔ - [ویکٹر ڈیٹابیس (vector database)](https://broadrouter.com/ur/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): ویکٹر ڈیٹابیس ایک ڈیٹا اسٹور ہے جو embedding ویکٹرز، اکثر ان کے سورس متن اور میٹا ڈیٹا کے ساتھ، رکھتا ہے اور nearest-neighbour کوئریز کا جواب دیتا ہے: کوئری ویکٹر ملنے پر یہ سب سے ملتے جلتے محفوظ ویکٹرز لوٹاتا ہے، عموماً cosine فاصلے، Euclidean فاصلے یا inner product سے۔ - [ویژن لینگویج ماڈل (VLM)](https://broadrouter.com/ur/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): ویژن لینگویج ماڈل (VLM) ایسا ماڈل ہے جو تصویر اور متن کو ایک ساتھ ان پٹ کے طور پر لے کر متن بناتا ہے، تاکہ وہ تصویروں کی وضاحت کر سکے، ان کے بارے میں سوالوں کے جواب دے سکے اور دستاویزات، چارٹس یا اسکرین شاٹس پڑھ سکے۔ VLM کو image-text-to-text یا multimodal ماڈل بھی کہتے ہیں۔ - [vLLM](https://broadrouter.com/ur/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM، LLM inference اور سرونگ کے لیے ایک اوپن سورس لائبریری ہے، جو پہلے UC Berkeley کی Sky Computing Lab میں بنائی گئی۔ یہ PagedAttention سے KV cache میموری سنبھالتی ہے، آنے والی درخواستوں کو مسلسل بیچ کرتی ہے، کئی quantization فارمیٹس سپورٹ کرتی ہے اور OpenAI-compatible HTTP سرور دیتی ہے۔ - [ووکیبلری سائز (vocabulary size)](https://broadrouter.com/ur/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): ووکیبلری سائز کسی لینگویج ماڈل کے ٹوکنائزر میں الگ الگ ٹوکنز کی تعداد ہے، ہر ایک کی اپنی ID اور embedding کے ساتھ۔ بڑی ووکیبلری عام متن کو کم ٹوکنز میں لکھ سکتی ہے، مگر اس سے embedding اور آؤٹ پٹ لیئرز بڑی ہوتی ہیں، اور یہ صرف اسی متن پر مدد کرتی ہے جس پر یہ تربیت یافتہ ہے۔ - [WAV فائل (WAV)](https://broadrouter.com/ur/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format، یا WAVE) مائیکروسافٹ اور IBM کے RIFF فارمیٹ پر بنا ایک آڈیو کنٹینر ہے۔ اس میں عموماً غیر کمپریسڈ linear PCM سیمپلز ہوتے ہیں، اس لیے اس کا سائز پہلے سے معلوم ہوتا ہے: فی سیکنڈ سیمپل ریٹ ضرب فی سیمپل بائٹس ضرب چینلز۔ - [ورڈ ایرر ریٹ (WER)](https://broadrouter.com/ur/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): ورڈ ایرر ریٹ (WER) اسپیچ ریکگنیشن کی درستی کا معیاری پیمانہ ہے: ٹرانسکرپٹ کو درست ریفرنس میں بدلنے کے لیے درکار الفاظ کے substitutions، deletions اور insertions کی تعداد، جسے ریفرنس کے کل الفاظ سے تقسیم کیا جاتا ہے۔ جتنا کم، اتنا بہتر۔ - [زیرو ڈیٹا ریٹینشن (zero data retention، ZDR)](https://broadrouter.com/ur/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): زیرو ڈیٹا ریٹینشن (ZDR) ڈیٹا سنبھالنے کی ایک پالیسی ہے، جس میں AI فراہم کنندہ پرامپٹس اور آؤٹ پٹ صرف درخواست پوری کرنے کے لیے پروسیس کرتا ہے اور بعد میں وہ مواد لاگز، کیش یا تربیتی ڈیٹا میں محفوظ نہیں کرتا۔ ٹوکن گنتی اور اوقات جیسے بغیر مواد کے ریکارڈ عموماً پھر بھی رکھے جاتے ہیں۔ ### Hindi (hi, हिन्दी): 294 translated pages - [Articles in Hindi](https://broadrouter.com/hi/blog/) - [अक्टूबर 2026 में LLM API कीमतों में बदलाव: 3 बढ़ोतरी, कटौती नहीं](https://broadrouter.com/hi/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): अक्टूबर 2026 में 7 में से 5 LLM API प्रोवाइडर के प्राइसिंग पेज पर 23 बदलाव: 3 बढ़ोतरी, कोई कटौती नहीं, 9 नए मॉडल की कीमत आई और 11 हटे। हर एक का आर्काइव लिंक। - [AI कोडिंग एजेंट के एक रन की लागत: स्टेप, कॉन्टेक्स्ट और कैशिंग](https://broadrouter.com/hi/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): कोडिंग एजेंट का रन जितना लंबा चलता है, हर स्टेप उतना महँगा क्यों पड़ता है, कैशिंग कुल लागत कैसे बदलती है, और फ़्लैट सेशन कब जीतता है। टोकन का पूरा हिसाब। - [LLM से अरबी टेक्स्ट में हरकतें (तश्कील) लगाना: प्रॉम्प्ट, जाँच, DER](https://broadrouter.com/hi/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): LLM API से अरबी हरकतें (tashkeel) वापस लगाएँ: हर अक्षर बचाने वाला प्रॉम्प्ट, बदलाव पकड़ने वाली जाँच, DER और WER स्कोरिंग, और CATT कब चुनें। - [LLM से अरबी व्याकरण और वर्तनी सुधार: एक how-to](https://broadrouter.com/hi/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): LLM API से अरबी वर्तनी और व्याकरण सुधारें: कम से कम बदलाव वाला प्रॉम्प्ट, दोबारा लिखने से बचाव, ग़लती की क़िस्म के हिसाब से edits और M2 जैसी स्कोरिंग। - [अरबी रिकॉर्डिंग से मीटिंग मिनट्स: एक API पाइपलाइन](https://broadrouter.com/hi/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): अरबी मीटिंग रिकॉर्डिंग को फ़ैसलों और एक्शन आइटम वाले मिनट्स में बदलें: ट्रांसक्राइब करें, साफ़ करें, LLM से JSON में सारांश लें, और नाम व तारीख़ें जाँचें। - [LLM से अरबी named entity extraction: JSON, तारीख़ें और जाँच](https://broadrouter.com/hi/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): LLM से अरबी नाम, संगठन, जगहें और हिजरी तारीख़ें JSON में निकालें: एक प्रॉम्प्ट, हरकतें अनदेखी करने वाली span जाँच, तारीख़ बदलना और स्कोरिंग स्क्रिप्ट। - [अरबी OCR API की तुलना: Google, Azure, Textract, Mistral, LLM](https://broadrouter.com/hi/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): कौन-से OCR API आधिकारिक रूप से छपी और हाथ से लिखी अरबी सपोर्ट करते हैं, हर एक क्या लौटाता है और प्रति 1,000 पेज कितना लेता है, vision LLM API के मुक़ाबले। - [speech-to-text API से अरबी सबटाइटल (SRT): आपको क्या चाहिए](https://broadrouter.com/hi/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): SRT फ़ाइल को क्या चाहिए, कौन-से speech-to-text API अरबी के लिए टाइमस्टैम्प या SRT देते हैं, और सिर्फ़-टेक्स्ट API से सबटाइटल बनाने वाली टेस्ट की गई पाइपलाइन। - [LLM से अरबी टेक्स्ट classification: sentiment, intent, triage](https://broadrouter.com/hi/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): LLM API से अरबी रिव्यू, संदेश और सपोर्ट टिकट classify करें: लेबल डिज़ाइन, बोली-सजग प्रॉम्प्ट, JSON आउटपुट, macro F1 टेस्टिंग और बड़े वॉल्यूम पर लागत। - [सर्च और RAG के लिए अरबी टेक्स्ट normalization: क्या बदलें, क्या नहीं](https://broadrouter.com/hi/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): तश्कील, alef के रूप, ta marbuta, hamza, फ़ारसी अक्षर, अंक और Unicode: अरबी normalization के कौन-से नियम सर्च में मदद करते हैं, कौन नुक़सान, और टेस्टेड Python। - [LLM से अरबी टेक्स्ट का सारांश: प्रॉम्प्ट, लंबाई और स्कोरिंग](https://broadrouter.com/hi/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): LLM API से अरबी लेख, रिपोर्ट और बोली वाले मैसेज का सारांश: प्रॉम्प्ट पैटर्न, शब्द-गिनती, संख्याओं की जाँच, और डिफ़ॉल्ट ROUGE अरबी को 0 क्यों देता है। - [Artificial Analysis के विकल्प: 2026 में LLM की तुलना कहाँ करें](https://broadrouter.com/hi/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Artificial Analysis के विकल्पों की तुलना: Arena, Epoch AI, Vals AI, OpenRouter रैंकिंग, Hugging Face और प्राइस डेटा, हर एक क्या मापता है और कब काम आता है। - [BroadNet का GPU हार्डवेयर और सर्विंग का तरीक़ा, BroadNet के शब्दों में](https://broadrouter.com/hi/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): BroadNet Inference के पीछे के हार्डवेयर के बारे में BroadNet क्या कहता है: B300 से क्लाइंट appliance तक तीन GPU स्तर, प्रति GPU throughput, और टेस्ट से मिली सीख। - [LLM API रिस्पॉन्स कैश करके लागत घटाएँ: Python में एक्ज़ैक्ट-मैच कैश](https://broadrouter.com/hi/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Python में LLM API रिस्पॉन्स के लिए टेस्ट किया हुआ एक्ज़ैक्ट-मैच कैश: key में क्या जाता है, क्या कभी कैश न करें, TTL, प्राइवेसी और बनाने से पहले हिट रेट कैसे मापें। - [ओपन मॉडल के साथ Claude Code: प्रॉक्सी, नेटिव एंडपॉइंट और सीमाएँ](https://broadrouter.com/hi/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code, Anthropic Messages API बोलता है। LiteLLM प्रॉक्सी या नेटिव एंडपॉइंट से इसे ओपन मॉडल पर कैसे चलाएँ, और रास्ते में क्या-क्या टूटता है। - [डॉक्यूमेंट AI वर्कफ़्लो में ह्यूमन रिव्यू को शामिल करें](https://broadrouter.com/hi/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): डॉक्यूमेंट-AI पाइपलाइन का ह्यूमन रिव्यू चरण कैसे बनाएँ: वैलिडेशन नियम, किसे इंसान तक भेजें, रैंडम ऑडिट सैंपल, ऑडिट ट्रेल और रिव्यू दर। - [मिस्र का डेटा सुरक्षा क़ानून (Law 151 of 2020) और LLM API](https://broadrouter.com/hi/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): मिस्र के लोगों का निजी डेटा विदेश के LLM या speech-to-text API को भेजना: Law 151 of 2020 और Decision 816 of 2025 ट्रांसफ़र, लाइसेंस और ब्रीच पर क्या कहते हैं। - [फ़ाइन-ट्यूनिंग से पहले: हल करने वाली समस्या तय करें](https://broadrouter.com/hi/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): कब प्रॉम्प्टिंग, RAG या evaluation समस्या हल कर देते हैं और फ़ाइन-ट्यूनिंग ज़रूरी नहीं, कैसा ट्रेनिंग डेटा चाहिए, और फ़ाइन-ट्यून माँगने से पहले उसे कैसे जाँचें। - [LLM का अमान्य JSON कैसे ठीक करें: कारण, रिपेयर, रीट्राई](https://broadrouter.com/hi/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): LLM टूटा JSON क्यों लौटाते हैं (कोड फ़ेंस, गद्य, max_tokens पर कटाव, ट्रेलिंग कॉमा), और टेस्ट किया Python फ़ंक्शन जो सुरक्षित चीज़ें ठीक करे, बाक़ी ठुकराए। - [LLM API की कीमतें कितनी बार बदलती हैं? आर्काइव पेजों के 3 साल](https://broadrouter.com/hi/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): हमने 2023 से 2026 तक 11 LLM API प्रोवाइडर के 308 आर्काइव किए प्राइसिंग पेज पढ़े। 80% मॉडल की कीमत कभी नहीं बदली; 75% बदलाव अकेले DeepInfra ने किए। - [LLM API की लागत कैसे कम करें: 10 कदमों की चेकलिस्ट, सही क्रम में](https://broadrouter.com/hi/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): LLM API का बिल सही क्रम में घटाएँ: पहले मापें, प्रीफ़िक्स कैश करें, प्रॉम्प्ट छोटा करें, आउटपुट और रीज़निंग सीमित करें, सही मॉडल चुनें और फ़्लैट रेट का हिसाब जानें। - [LLM API फ़ेलओवर: अपने कोड में दूसरे प्रोवाइडर पर फ़ॉलबैक](https://broadrouter.com/hi/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): OpenAI-कम्पैटिबल LLM प्रोवाइडरों के बीच फ़ेलओवर कैसे करें: किन एरर पर स्विच करें, सर्किट ब्रेकर, मॉडल ID मैपिंग, स्ट्रीमिंग, लागत और प्राइवेसी का मेल। - [LLM API कीमतों में कटौती 2023–2026: सबसे बड़ी गिरावटें, स्रोतों के साथ](https://broadrouter.com/hi/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): 2023–2026 की सबसे बड़ी LLM API कीमत कटौतियाँ, पहले और बाद के Internet Archive लिंक के साथ: o3, Codestral, Gemini 1.5, और कम चर्चित बढ़ोतरियाँ भी। - [क्या LLM API की कीमतें सिर्फ़ घटती हैं? 52 बढ़ोतरियाँ और बजट का तरीक़ा](https://broadrouter.com/hi/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): LLM API की कीमतें बढ़ती भी हैं। 2024 से 2026 के आर्काइव किए प्राइसिंग पेजों में एक ही मॉडल पर 52 बढ़ोतरियाँ मिलीं, मीडियन 40%। सारे स्रोत और बजट का नियम। - [प्रोवाइडर के हिसाब से LLM API रेट लिमिट: RPM, TPM और कॉन्करेंसी](https://broadrouter.com/hi/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): 13 प्रोवाइडर की एंट्री-टियर LLM API रेट लिमिट, उनके डॉक्स से: OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks वगैरह, और RPM व TPM का मतलब। - [LLM API SLA तुलना: Azure, Bedrock, Google और OpenAI क्या वादा करते हैं](https://broadrouter.com/hi/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): हर LLM API प्रोवाइडर लिखित में क्या वादा करता है: अपटाइम प्रतिशत, डाउनटाइम किसे माना जाता है, सर्विस क्रेडिट और लेटेंसी टारगेट, SLA पेजों से पढ़कर। - [Batch API के बिना LLM बैच प्रोसेसिंग: दोबारा शुरू होने वाला runner](https://broadrouter.com/hi/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Batch API के बिना chat-completions API से 10,000 से 1,000,000 प्रॉम्प्ट चलाएँ: concurrency सीमा, रेट पेसिंग, retry, दोबारा शुरू होने वाले checkpoint और समय का हिसाब। - [Safety Map: 39 LLM संवेदनशील डेटा को कैसे संभालते हैं](https://broadrouter.com/hi/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet ने 14 लैब के 39 LLM को credentials, PII, हेल्थ और वित्तीय डेटा पर परखा: क्या transfer हुआ, क्या नहीं, और tool call व रीज़निंग से हुए लीक। - [संवेदनशील डेटा की सुरक्षा के लिए LLM को कैसे परखें: BroadNet का तरीक़ा](https://broadrouter.com/hi/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): डिप्लॉय करने से पहले LLM credentials, PII, हेल्थ या वित्तीय डेटा लीक करता है या नहीं, यह परखने का BroadNet का प्रकाशित तरीक़ा: कैटेगरी, leak बनाम miss, 3+ रन। - [OpenAI API से माइग्रेशन: Python, Node और cURL में कोड वॉकथ्रू](https://broadrouter.com/hi/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): OpenAI से OpenAI-कम्पैटिबल API पर जाने के लिए कोड के बदलाव: क्लाइंट सेटअप, Responses से Chat Completions, टूल, स्ट्रीमिंग, embeddings और एक स्मोक टेस्ट। - [BroadNet ने किन मॉडलों को परखकर ख़ारिज किया, और क्यों](https://broadrouter.com/hi/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet अपने टेस्ट किए और ख़ारिज मॉडल कारण सहित प्रकाशित करता है: क्रेडेंशियल-सुरक्षा जाँच में फ़ेल चैट मॉडल, और हारे हुए स्पीच, OCR, एम्बेडिंग मॉडल। - [कई भाषाओं के लिए AI प्रोडक्ट कैसे डिज़ाइन करें](https://broadrouter.com/hi/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): अरबी, अंग्रेज़ी और स्पेनिश यूज़र्स के लिए AI प्रोडक्ट डिज़ाइन: UI लोकेल बनाम मॉडल की भाषा, RTL और मिला-जुला टेक्स्ट, संख्याएँ, तारीख़ें, भाषा पहचान और लागत। - [ओपन-वेट बनाम क्लोज़्ड मॉडल API की लागत: कीमतों का डेटा क्या कहता है](https://broadrouter.com/hi/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): क्या API के ज़रिए ओपन-वेट मॉडल GPT, Claude या Gemini से सस्ते हैं? प्रोवाइडरों के अपने पेज से प्रति टोकन कीमतें, होस्ट का फ़ासला और प्रति यूज़र लागत। - [OpenAI बनाम Anthropic बनाम Google: API कीमतों का इतिहास, 2023–2026](https://broadrouter.com/hi/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): आर्काइव किए प्राइसिंग पेजों से GPT, Claude और Gemini API की कीमतें साथ-साथ, 2023 से 2026: कौन-सी लाइन सस्ती हुई, कौन-सी महँगी, हर एक आर्काइव लिंक के साथ। - [OpenAI-compatible API के base URL: 14 प्रोवाइडर, डॉक्स से जाँचे हुए](https://broadrouter.com/hi/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): 14 OpenAI-compatible LLM API का base_url, हर एक प्रोवाइडर के अपने डॉक्स से लिया गया, उनके बताए एंडपॉइंट और वे बारीकियाँ जिनसे सीधा स्विच टूट जाता है। - [सस्ते मॉडल पर OpenClaw और Hermes Agent: सेटअप और टोकन का हिसाब](https://broadrouter.com/hi/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): OpenClaw या Hermes Agent को OpenAI-compatible endpoint से जोड़ें, समझें कि शेड्यूल्ड चेक टोकन क्यों खाते हैं, और हल्के कॉन्टेक्स्ट व सस्ते मॉडल से बिल घटाएँ। - [प्राइवेट एंडपॉइंट की शुरुआत साफ़ वर्कलोड से होती है](https://broadrouter.com/hi/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): प्राइवेट LLM एंडपॉइंट माँगने से पहले क्या तय करें: वर्कलोड, प्रति सेकंड टोकन, समानांतर अनुरोध, कॉन्टेक्स्ट, रीजन, लाइसेंस और बजट, फिर उसे टेस्ट कैसे करें। - [कोडिंग एजेंट में prompt injection: बचाव जो सच में टिकते हैं](https://broadrouter.com/hi/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): prompt injection कोडिंग एजेंट तक कहाँ से पहुँचता है, और वेंडर और रिसर्चर कौन-से परतदार बचाव बताते हैं: सैंडबॉक्स, नेटवर्क पर रोक, allow-list और मंज़ूरी। - [RAG चंक साइज़ और ओवरलैप: रिट्रीवल और लागत पर इनका असर](https://broadrouter.com/hi/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): चंक साइज़ और ओवरलैप RAG रिकॉल, इंडेक्स साइज़, एम्बेडिंग लागत और हर क्वेरी में LLM को जाने वाले टोकन कैसे बदलते हैं, प्रकाशित नतीजों और हिसाब के साथ। - [रीज़निंग मॉडल के लिए max_tokens कैसे तय करें: बजट का तरीक़ा](https://broadrouter.com/hi/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): कम max_tokens पर रीज़निंग मॉडल ख़ाली जवाब क्यों देते हैं, और नापे गए percentile, कॉन्टेक्स्ट विंडो और बढ़ते retry से बजट कैसे तय करें। - [ब्राउज़र में ऑडियो रिकॉर्ड करें और speech-to-text API को भेजें](https://broadrouter.com/hi/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): MediaRecorder से माइक्रोफ़ोन रिकॉर्ड करें, WebM या Safari का MP4 16 kHz WAV में बदलें, और API key रखने वाले छोटे सर्वर से ट्रांसक्रिप्शन API को भेजें। - [प्रॉम्प्ट टोकन कैसे घटाएँ: सिस्टम प्रॉम्प्ट, टूल, डेटा की मापी कटौती](https://broadrouter.com/hi/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): सिस्टम प्रॉम्प्ट, few-shot उदाहरण, टूल स्कीमा, डेटा फ़ॉर्मैट, चैट इतिहास और अरबी हरकात छोटे करने से टोकन की मापी गई बचत, असली टोकनाइज़र से गिनी गई। - [एक मॉडल, अलग होस्ट: LLM API की कीमतें कितनी अलग होती हैं](https://broadrouter.com/hi/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): कई होस्ट पर बिकने वाले 119 मॉडल में सबसे महँगी लिस्ट प्राइस सबसे सस्ती की मीडियन 1.76 गुना है। 41 प्रोवाइडरों के अपने प्राइसिंग पेज से डेटा, नवंबर 2026। - [सऊदी और UAE PDPL और LLM API: cross-border डेटा ट्रांसफ़र चेकलिस्ट](https://broadrouter.com/hi/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): सऊदी अरब या UAE से LLM API को निजी डेटा भेजना: PDPL के ट्रांसफ़र नियम क्या कहते हैं, प्रोवाइडर से क्या पूछें, और देश के भीतर inference कब ज़रूरी है। - [टेक्स्ट, स्पीच या vision? पहले टास्क देखें](https://broadrouter.com/hi/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): टास्क के हिसाब से टेक्स्ट चैट, इमेज इनपुट या ऑडियो ट्रांसक्रिप्शन चुनें, BroadNet के डॉक्यूमेंटेड मॉडल और सीमाओं के साथ, और वह भी जो वह नहीं देता, जैसे TTS। - [25 MB से बड़ा लंबा ऑडियो ट्रांसक्राइब करें: बाँटें, भेजें, जोड़ें](https://broadrouter.com/hi/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): speech-to-text API की फ़ाइल सीमा से बड़ी रिकॉर्डिंग कैसे ट्रांसक्राइब करें: फ़ाइल छोटी करें, ffmpeg से ख़ामोशी पर बाँटें, टुकड़े समानांतर भेजें और टेक्स्ट जोड़ें। - [LLM से ऐप strings का अरबी अनुवाद: placeholders और plurals](https://broadrouter.com/hi/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): LLM API से i18n JSON फ़ाइलों का अरबी अनुवाद करें: placeholders फ़्रीज़ करें, अरबी के छहों plural रूप बनवाएँ, CI में ICU syntax जाँचें और डाले गए नाम isolate करें। - [API से अरबी ऑडियो का अंग्रेज़ी में अनुवाद: टेस्ट की हुई पाइपलाइन](https://broadrouter.com/hi/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): दो API कॉल से अरबी रिकॉर्डिंग को अंग्रेज़ी टेक्स्ट बनाएँ: language=ar से ट्रांसक्राइब करें, फिर LLM से अनुवाद, संख्या जाँच, chunking, लागत और स्कोरिंग के साथ। - [Live, on request, research: मॉडल कैटलॉग को साफ़-साफ़ कैसे पढ़ें](https://broadrouter.com/hi/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): कैसे पहचानें कि कैटलॉग के कौन-से मॉडल आज कॉल हो सकते हैं, किनके लिए पूछताछ चाहिए और कौन सिर्फ़ टेस्ट हुए, BroadNet के कैटलॉग और /v1/models के उदाहरण से। - [vLLM में Uno: BroadNet की सर्विंग स्टडी ने क्या मापा](https://broadrouter.com/hi/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet ने Uno self-drafting adapter को vLLM में जोड़ा और H100 पर मापा: एक stream पर सादे vLLM से 2.58x थ्रूपुट, 32 पर 1.74x। विधि, DOI और सीमाएँ। - [Gemma 4 26B पर Uno: BroadNet की mixture-of-experts स्टडी](https://broadrouter.com/hi/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet ने mixture-of-experts मॉडल Gemma 4 26B A4B के लिए पहला Uno self-drafting adapter ट्रेन किया। क्या मापा गया, self-drafting कब फ़ायदेमंद है, और DOI। - [एजेंट टोकन बजट: एजेंट लूप के टोकन कहाँ जाते हैं](https://broadrouter.com/hi/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): सिस्टम प्रॉम्प्ट, टूल स्कीमा, फ़ाइल पढ़ना, हिस्ट्री और रीज़निंग: एजेंट लूप टोकन कैसे ख़र्च करता है, usage डेटा से हर हिस्सा कैसे मापें, और पहले क्या घटाएँ। - [अरबी ASR में word error rate (WER) क्या है, आसान भाषा में](https://broadrouter.com/hi/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): word error rate का मतलब क्या है, इसे कैसे निकालें, और अरबी WER हरकतों, वर्तनी normalisation और बोलियों पर इतना निर्भर क्यों है। Python और असली नंबरों के साथ। - [अरबी बोलियों की speech recognition: खाड़ी, मिस्री, लेवेंटाइन, मग़रिबी](https://broadrouter.com/hi/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): खाड़ी, मिस्री, लेवेंटाइन और मग़रिबी अरबी पर speech-to-text क्यों लड़खड़ाता है, और ऐसा बोली टेस्ट सेट कैसे बनाएँ जो दिखाए कि प्रोवाइडर आपके ऑडियो पर कैसा करेगा। - [LLM से अरबी–अंग्रेज़ी अनुवाद: प्रॉम्प्ट, शब्दावली, मूल्यांकन](https://broadrouter.com/hi/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): LLM API से अरबी से अंग्रेज़ी और वापस अनुवाद करें: लहजा बचाने वाले प्रॉम्प्ट, तय शब्दों की शब्दावली, संख्याएँ और नाम, और क्वालिटी का मूल्यांकन कैसे करें। - [खाड़ी, मिस्री और लेवेंटाइन अरबी: LLM बोलियों को कैसे संभालते हैं](https://broadrouter.com/hi/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): LLM खाड़ी, मिस्री और लेवेंटाइन अरबी कैसे संभालते हैं: कहाँ चूकते हैं, बोली के लिए प्रॉम्प्ट और टेस्ट कैसे करें, और Qwen3.8 27B पर BroadNet की रिपोर्ट। - [अरबी OCR बनाम vision LLM: अरबी दस्तावेज़ और इनवॉइस पढ़ना](https://broadrouter.com/hi/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): अरबी इनवॉइस, रसीदों और फ़ॉर्म के लिए कब पारंपरिक OCR और कब vision LLM: सटीकता के जाल, JSON आउटपुट, प्रति पेज लागत और एक validation पाइपलाइन। - [अरबी प्रॉम्प्ट इंजीनियरिंग: register, RTL, अंक और मिला-जुला टेक्स्ट](https://broadrouter.com/hi/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): ऐसे प्रॉम्प्ट लिखें जो LLM से अच्छी अरबी निकलवाएँ: register और बोली चुनें, अंक और तारीख़ें नियंत्रित करें, RTL फ़ॉर्मैटिंग साफ़ रखें, अरबी–अंग्रेज़ी मिश्रण संभालें। - [अरबी RAG: normalisation, chunking और retrieval जो सच में काम करे](https://broadrouter.com/hi/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): अरबी दस्तावेज़ों पर RAG: अलिफ़, या और ता मरबूता normalise करें, समझदारी से chunk करें, बहुभाषी embeddings चुनें और keyword search जोड़ें। - [अरबी speech-to-text API: 2026 में सही API कैसे चुनें](https://broadrouter.com/hi/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): अरबी ट्रांसक्रिप्शन API कैसे चुनें: word error rate, बोलियाँ, फ़ाइल सीमाएँ, प्रति ऑडियो घंटा कीमत और प्राइवेसी, साथ में आपकी रिकॉर्डिंग के लिए टेस्ट प्लान। - [अरबी WhatsApp chatbot का आर्किटेक्चर: वॉइस नोट से जवाब तक](https://broadrouter.com/hi/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): अरबी WhatsApp असिस्टेंट का काम करता आर्किटेक्चर: webhook, queue, वॉइस नोट ट्रांसक्रिप्शन, LLM जवाब, पॉलिसी नियम और लागत, Python कोड के साथ। - [Arabizi और LLM: 3 और 7 जैसे अंकों वाली Franco-Arabic को संभालना](https://broadrouter.com/hi/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): Arabizi अरबी को लैटिन अक्षरों में लिखती है, ع के लिए 3 और ح के लिए 7 जैसे अंकों के साथ। इसे कैसे पहचानें, LLM से अरबी लिपि में बदलें, और सही तरह जवाब दें। - [Aider के लिए सबसे अच्छा LLM: मॉडल कैसे चुनें और जोड़ें](https://broadrouter.com/hi/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Aider के लिए सबसे अच्छा LLM कैसे चुनें: एडिट फ़ॉर्मैट का पालन, कॉन्टेक्स्ट साइज़ और प्रति घंटा लागत, अपनी रिपो पर एक घंटे का टेस्ट, और OPENAI_API_BASE का सटीक सेटअप। - [कोडिंग एजेंट के लिए बेस्ट ओपन मॉडल: टूल, कॉन्टेक्स्ट, लागत से चुनें](https://broadrouter.com/hi/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): आपके कोडिंग एजेंट के लिए GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B या Gemma 4? टूल कॉलिंग, कॉन्टेक्स्ट, रीज़निंग और प्रति टास्क लागत से चुनें, पूरे हिसाब के साथ। - [कैश्ड इनपुट टोकन क्या हैं: LLM API का बिल कैसे घटाएँ](https://broadrouter.com/hi/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): कैश्ड इनपुट सामान्य इनपुट की कीमत के छोटे हिस्से पर बिल होता है। प्रॉम्प्ट कैशिंग का बिल, usage में cached_tokens पढ़ना, और प्रॉम्प्ट को कैश के लायक बनाना। - [सबसे सस्ता विज़न LLM API: प्रति इमेज और प्रति पेज लागत का हिसाब](https://broadrouter.com/hi/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): विज़न LLM कॉल की प्रति इमेज और प्रति डॉक्यूमेंट पेज लागत: Gemma 4, Qwen3.8 27B और DeepSeek V4.1 Flash के टोकन का हिसाब, और DeepInfra कहाँ सस्ता है। - [LLM API प्रोवाइडर कैसे चुनें: 20 सवालों की चेकलिस्ट](https://broadrouter.com/hi/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): LLM API चुनने से पहले पूछने लायक़ 20 सवाल: कीमत, सीमाएँ, OpenAI कम्पैटिबिलिटी, प्राइवेसी, मॉडल, सपोर्ट और अपटाइम, साथ में स्कोरिंग शीट और टेस्ट प्लान। - [अरबी के लिए LLM API कैसे चुनें: क्या-क्या टेस्ट करें](https://broadrouter.com/hi/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): अरबी ऐप के लिए LLM कैसे चुनें: बोली की समझ, लेखन की गुणवत्ता, अनुवाद, अरबी दस्तावेज़ और इमेज, लागत, और एक दिन में बनने वाला टेस्ट सेट। - [Cline बनाम Roo Code बनाम Aider: ओपन मॉडल वाली टीमों के लिए (2026)](https://broadrouter.com/hi/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (अब आर्काइव, Zoo Code के रूप में जारी) और Aider की तुलना, ओपन मॉडल के लिए अहम बातों पर: टूल कॉल, कॉन्टेक्स्ट, MCP, git और अप्रूवल। - [LLM अनुरोध भेजने से पहले टोकन कैसे गिनें](https://broadrouter.com/hi/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): LLM API कॉल से पहले भाषा और मॉडल के हिसाब से प्रॉम्प्ट टोकन का अनुमान: टोकनाइज़र क्यों अलग हैं, पाँच भाषाओं के मापे गए उदाहरण और सुरक्षित max_tokens मार्जिन। - [कोडिंग एजेंट की LLM लागत घटाने के 10 तरीक़े, अनुमानित बचत के साथ](https://broadrouter.com/hi/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Cline, Aider, opencode या आपका अपना कोडिंग एजेंट टोकन पर जो ख़र्च करता है, उसे घटाने के दस ठोस तरीक़े। हर एक के साथ बचत का हिसाब और उसे मापने का तरीक़ा। - [DeepSeek API के विकल्प: DeepSeek मॉडल कॉल करने के 5 तरीक़े](https://broadrouter.com/hi/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): DeepSeek मॉडल को आधिकारिक API, एग्रीगेटर, क्लाउड कैटलॉग, अपने GPU या BroadNet Inference से कॉल करना: कीमत, प्राइवेसी और सीमाओं के फ़ायदे-नुक़सान। - [DeepSeek V4.1 Flash API की कीमत: प्रति टोकन, कैश्ड और प्लान](https://broadrouter.com/hi/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): DeepSeek V4.1 Flash API की कीमत प्रति 1M इनपुट, कैश्ड और आउटपुट टोकन, मासिक, साप्ताहिक और घंटे वाले फ़्लैट प्लान, और DeepSeek, DeepInfra व Fireworks से तुलना। - [फ़ीचर लॉन्च से पहले LLM API की लागत का अनुमान कैसे लगाएँ](https://broadrouter.com/hi/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): नए AI फ़ीचर के लिए क़दम-दर-क़दम लागत मॉडल: यूज़र × एक्शन × कॉल × टोकन, कैश, रीज़निंग, रीट्राई, रेट लिमिट और एक हल की हुई मासिक टेबल के साथ। - [प्रोवाइडर बदलने से पहले LLM का मूल्यांकन कैसे करें](https://broadrouter.com/hi/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): असली ट्रैफ़िक से एक छोटा eval सेट बनाएँ, कोड चेक और पक्षपात-रहित LLM जज से स्कोर करें, और माइग्रेट करने से पहले क्वालिटी, लागत और लेटेंसी की तुलना करें। - [वीज़न LLM से इनवॉइस, रसीद और फ़ॉर्म का डेटा JSON में निकालें](https://broadrouter.com/hi/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): इनवॉइस, रसीद और फ़ॉर्म के लिए प्रोडक्शन पाइपलाइन: इमेज की तैयारी, JSON स्कीमा वाला प्रॉम्प्ट, Pydantic validation, जोड़ की जाँच, retry और ह्यूमन रिव्यू क्यू। - [फ़्लैट रेट बनाम प्रति टोकन LLM API: ब्रेक-ईवन का हिसाब](https://broadrouter.com/hi/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): फ़्लैट रेट LLM प्लान प्रति टोकन भुगतान से कब सस्ता पड़ता है? DeepSeek V4.1 Flash के चार हल किए हुए मासिक बिल, ब्रेक-ईवन वॉल्यूम, और अपना उपयोग मापने का तरीक़ा। - [मुफ़्त LLM API क्रेडिट: शुरुआती क्रेडिट असल में कितना चलता है](https://broadrouter.com/hi/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): मुफ़्त LLM API क्रेडिट से असल में क्या मिलता है: BroadNet के शुरुआती क्रेडिट पर टोकन, ऑडियो मिनट और घंटे वाली key, Gemini और OpenRouter के फ़्री टियर से तुलना। - [GDPR और LLM API: अपने प्रोवाइडर से पूछने लायक़ 15 सवाल](https://broadrouter.com/hi/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): LLM API के लिए GDPR चेकलिस्ट: DPA, सबप्रोसेसर, डेटा रिटेंशन, ट्रेनिंग, अंतरराष्ट्रीय ट्रांसफ़र और ब्रीच नोटिस, हर सवाल के पीछे के GDPR अनुच्छेद के साथ। - [GLM 5.3 बनाम DeepSeek V4.1 Flash: कौन-सा चुनें और कब](https://broadrouter.com/hi/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 बनाम DeepSeek V4.1 Flash: कीमत, कॉन्टेक्स्ट, टूल, स्ट्रक्चर्ड आउटपुट, इमेज और रीज़निंग की तुलना, असली कामों की लागत और किसे कब चुनें, इसका सीधा नियम। - [LLM API में 429 एरर कैसे सँभालें: बैकऑफ़, जिटर और लिमिटर](https://broadrouter.com/hi/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 बनाम 5xx, फ़ुल जिटर के साथ exponential backoff, LLM API के लिए कॉन्करेंसी लिमिटर और क्यू, साथ में प्रोडक्शन में कॉपी करने लायक Python और Node कोड। - [घंटे वाली LLM API key: प्रति घंटा तय कीमत, आसान भाषा में](https://broadrouter.com/hi/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): BroadNet की घंटे वाली key तय कीमत पर एक stream पर Gemma 4 VL 26B का एक घंटा देती है। इसमें क्या शामिल है, सीमाएँ, यह किसके लिए है और कब फ़ायदेमंद है। - [LLM API डेटा रिटेंशन की तुलना: OpenAI, Anthropic, Gemini, Mistral](https://broadrouter.com/hi/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): OpenAI, Anthropic, Gemini API, Mistral और BroadNet, API प्रॉम्प्ट रखने और उन पर ट्रेनिंग के बारे में आधिकारिक रूप से क्या कहते हैं, नवंबर 2026 के पेजों से उद्धृत। - [LLM API key को सुरक्षित कैसे रखें: स्टोरेज, रोटेशन और IP allow-list](https://broadrouter.com/hi/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): LLM API key को कोड और लॉग से दूर रखें, हर वर्कलोड को अलग key दें, तय समय पर रोटेट करें, key को सर्वर IP से बाँधें और लीक हुई key सेकंडों में रद्द करें। - [LLM API की कीमत कैसे तय होती है: इनपुट, कैश्ड, आउटपुट और रीज़निंग](https://broadrouter.com/hi/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): प्रति टोकन LLM API की कीमत कैसे बनती है: इनपुट, कैश्ड इनपुट, आउटपुट, रीज़निंग टोकन, इमेज और प्रति मिनट ऑडियो, एक प्रकाशित रेट कार्ड के उदाहरणों के साथ। - [LLM API को संवेदनशील डेटा भेजना: redact करें, कम भेजें, कम लॉग करें](https://broadrouter.com/hi/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): ग्राहक, स्वास्थ्य या वित्तीय डेटा LLM API को सुरक्षित रूप से कैसे भेजें: फ़ील्ड वर्गीकृत करें, वापस बदले जा सकने वाले placeholder लगाएँ, ज़ीरो रिटेंशन, साफ़ लॉग। - [LLM API ख़र्च पर नियंत्रण: बजट, key सीमाएँ और लूप से बचाव](https://broadrouter.com/hi/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): LLM API का बिल बेक़ाबू होने से रोकें: प्रीपेड बैलेंस, हर ऐप की अलग key, IP allow-list, usage लॉगिंग, कोड में बजट, लूप की सीमा और फ़्लैट-कीमत वाली key। - [LLM API usage लॉगिंग: हर यूज़र और फ़ीचर की लागत, PII हटाकर](https://broadrouter.com/hi/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): हर LLM API कॉल को टोकन, लागत, लेटेंसी और अनुरोध ID के साथ लॉग करें, ख़र्च को यूज़र और फ़ीचर से जोड़ें, और लिखने से पहले PII हटाएँ। Python कोड के साथ। - [प्रति यूज़र LLM लागत: SaaS के AI फ़ीचर की यूनिट इकोनॉमिक्स](https://broadrouter.com/hi/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): हिसाब लगाएँ कि एक AI फ़ीचर प्रति सक्रिय यूज़र प्रति माह कितने का पड़ता है, भारी यूज़र औसत से ज़्यादा क्यों मायने रखते हैं, और SaaS टियर की कीमत कैसे तय करें। - [इन्फ़रेंस प्रोवाइडर ओपन मॉडल की कीमत कैसे तय करते हैं: 6 की तुलना](https://broadrouter.com/hi/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): OpenRouter, Together AI, Fireworks, Groq, DeepInfra और BroadNet ओपन मॉडल की कीमत कैसे लेते हैं: प्रति टोकन, कैश्ड, बैच, GPU घंटे और फ़्लैट प्लान। - [LLM लेटेंसी: TTFT, टोकन प्रति सेकंड और एंड-टू-एंड समय](https://broadrouter.com/hi/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): टाइम टू फ़र्स्ट टोकन, टोकन प्रति सेकंड और एंड-टू-एंड लेटेंसी क्या मापते हैं, क्लाइंट से हर एक को कोड से कैसे मापें, और कौन-सा सुधार कौन-सी संख्या बदलता है। - [LLM स्ट्रीमिंग समझें: SSE चंक, डेल्टा, usage और कैंसल](https://broadrouter.com/hi/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Chat-completion स्ट्रीमिंग वायर पर कैसे चलती है, डेल्टा से जवाब कैसे जोड़ें, आख़िरी चंक से usage पढ़ें और Python, Node व curl में साफ़ तरीक़े से कैंसल करें। - [लॉन्ग कॉन्टेक्स्ट बनाम RAG: कब 1M टोकन वेक्टर DB की जगह लेते हैं](https://broadrouter.com/hi/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): 1M-टोकन कॉन्टेक्स्ट विंडो कब वेक्टर डेटाबेस की जगह ले सकती है और कब RAG ही बेहतर है: प्रति क्वेरी लागत, लेटेंसी, रिकॉल और रखरखाव पर तुलना। - [लंबे चलने वाले एजेंट में 1M विंडो के साथ कॉन्टेक्स्ट मैनेजमेंट](https://broadrouter.com/hi/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): 1M-टोकन विंडो के साथ लंबे एजेंट सेशन को सटीक और सस्ता कैसे रखें: append-only हिस्ट्री, कैश के अनुकूल प्रॉम्प्ट, कब compact करें और पहले क्या हटाएँ। - [लंबे चलने वाले एजेंट: टिकाऊ timeout, retry और checkpoint](https://broadrouter.com/hi/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): LLM एजेंट को घंटों चलाते रहें: हर कॉल पर timeout, अटकने का पता लगाने के लिए स्ट्रीमिंग, सुरक्षित retry, idempotent tool side effect और हर चरण के बाद checkpoint। - [OpenAI-कम्पैटिबल मॉडल के साथ MCP सर्वर (किसी भी प्रोवाइडर पर)](https://broadrouter.com/hi/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP किसी भी LLM के साथ चलता है: क्लाइंट MCP टूल को फ़ंक्शन कॉल में बदलता है। Cline, Continue, opencode, n8n या अपने Python कोड में ओपन मॉडल के साथ MCP कैसे चलाएँ। - [1M-टोकन कॉन्टेक्स्ट विंडो कब फ़ायदेमंद है (और कब नहीं)](https://broadrouter.com/hi/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): 1,048,576 टोकन में क्या समाता है, DeepSeek V4.1 Flash पर इसकी लागत क्या है, रीज़निंग और कैशिंग हिसाब कैसे बदलते हैं, और कब RAG अब भी बेहतर है। - [मल्टी-एजेंट सेटअप की असली लागत: प्रति टोकन बनाम फ़्लैट सेशन](https://broadrouter.com/hi/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): मल्टी-एजेंट सिस्टम एक एजेंट से कई गुना टोकन ख़र्च करते हैं। फ़्लैट प्लान बनाम प्रति-टोकन बिलिंग का सेशन वाला हिसाब, छिपे गुणक, और कब कौन जीतता है। - [मल्टीलिंगुअल speech-to-text API: 14 भाषाएँ, एक एंडपॉइंट](https://broadrouter.com/hi/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): अंग्रेज़ी, स्पेनिश, फ़्रेंच, चीनी, अरबी और 9 और भाषाओं को प्रति मिनट कीमत वाले एक OpenAI-compatible एंडपॉइंट से ट्रांसक्राइब करें, और जानें कि अरबी मॉडल कब चुनें। - [सस्ते LLM पर n8n AI Agent: सेटअप, मॉडल का चुनाव और प्रति रन लागत](https://broadrouter.com/hi/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): कम लागत वाले OpenAI-कम्पैटिबल मॉडल पर n8n AI Agent वर्कफ़्लो बनाएँ: टूल के लिए कौन-सा मॉडल, कौन-सी सेटिंग टोकन घटाती हैं, और हर 10,000 रन का हिसाब। - [बिज़नेस के लिए open-weight बनाम closed AI मॉडल: लागत, नियंत्रण, जोखिम](https://broadrouter.com/hi/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): आपके बिज़नेस के लिए open-weight या closed LLM? लागत, नियंत्रण, प्राइवेसी, क्वालिटी का अंतर और lock-in की तुलना, open मॉडल चलाने के तीन तरीक़ों के साथ। - [OpenAI-कम्पैटिबल API पर माइग्रेट करना: पूरी चेकलिस्ट](https://broadrouter.com/hi/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): base_url बदलना आसान हिस्सा है। टूटते हैं मॉडल ID, रीज़निंग टोकन, स्ट्रीमिंग usage, एरर, रेट लिमिट और key। कोड के साथ पूरी चेकलिस्ट। - [AI एजेंट के लिए OpenRouter बनाम फ़्लैट रेट प्लान: कौन सस्ता है?](https://broadrouter.com/hi/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter प्रति टोकन और क्रेडिट पर फ़ीस लेता है; फ़्लैट प्लान प्रति सेशन। हमेशा चालू एजेंट की लागत का हिसाब, OpenRouter की फ़ीस और सीमाएँ, और कौन किसके लिए ठीक है। - [कोडिंग एजेंट को रात भर बिना निगरानी के सुरक्षित कैसे चलाएँ](https://broadrouter.com/hi/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): बिना निगरानी वाले कोडिंग एजेंट रन का व्यावहारिक सेटअप: सैंडबॉक्स, git चेकपॉइंट, टास्क फ़ाइल, स्टेप की सीमा, सुबह की समीक्षा, और रात भर के लिए फ़्लैट प्लान सेशन। - [प्रॉम्प्ट को कैसे बनाएँ ताकि प्रीफ़िक्स कैश हो](https://broadrouter.com/hi/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): LLM प्रॉम्प्ट को ऐसे सजाएँ कि कैश्ड प्रीफ़िक्स एक जैसा रहे: सिस्टम प्रॉम्प्ट, टूल स्कीमा, डॉक्यूमेंट, सिर्फ़ आगे जुड़ता इतिहास, और इसे साबित करने वाला टेस्ट। - [रीज़निंग टोकन: दिखने से ज़्यादा आउटपुट का भुगतान क्यों होता है](https://broadrouter.com/hi/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): रीज़निंग मॉडल छिपी सोच का बिल आउटपुट टोकन के रूप में बनाते हैं। इसे कैसे मापें, हर मॉडल पर इसकी लागत, और reasoning_effort, max_tokens व रूटिंग से इसे कैसे घटाएँ। - [फ़्लैट-रेट LLM API पर पैरेलल कोडिंग एजेंट कैसे चलाएँ](https://broadrouter.com/hi/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): अनलिमिटेड DeepSeek V4.1 Flash पर सब-एजेंट, ऑर्केस्ट्रेटर और बैच वर्कर। कितने प्लान सेशन चाहिए, उन्हें key में कैसे बाँटें, और ओवरलैप पर सीमा कैसे लगाएँ। - [LLM सेल्फ़-होस्ट करें या API लें: असली लागत और ब्रेक-ईवन](https://broadrouter.com/hi/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): आधिकारिक पेजों से GPU किराए की कीमतें, अपने GPU पर प्रति 1M टोकन लागत का फ़ॉर्मूला, API को मात देने के लिए ज़रूरी उपयोग दर, और ऑपरेशन की भूली हुई लागतें। - [स्पीच-टू-टेक्स्ट API की कीमतों की तुलना: 2026 में प्रति मिनट कीमत](https://broadrouter.com/hi/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq और BroadNet Transcribe की प्रति मिनट और प्रति 1,000 घंटे कीमतें, आधिकारिक पेजों से, ऐड-ऑन समेत। - [JSON mode बनाम json_schema: भरोसेमंद स्ट्रक्चर्ड आउटपुट](https://broadrouter.com/hi/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON mode वैध JSON देता है; json_schema आपका ढाँचा। OpenAI-कम्पैटिबल मॉडल पर दोनों कैसे माँगें, Pydantic से वैलिडेट करें, और ग़लत आउटपुट लूप में ठीक करें। - [अरबी बनाम अंग्रेज़ी में प्रति शब्द टोकन: अरबी महँगी क्यों पड़ती है](https://broadrouter.com/hi/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): एक ही अर्थ के लिए अरबी में अक्सर अंग्रेज़ी से ज़्यादा LLM टोकन क्यों लगते हैं, प्रति शब्द टोकन क्यों भ्रामक है, और usage फ़ील्ड से अपना अनुपात मापने की स्क्रिप्ट। - [ओपन मॉडल के साथ टूल कॉलिंग: स्कीमा, पैरेलल कॉल, रीट्राई](https://broadrouter.com/hi/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): OpenAI-कम्पैटिबल ओपन मॉडल पर फ़ंक्शन कॉलिंग: स्कीमा के टिप्स, पैरेलल टूल कॉल, टूल कॉल फ़ेल होने के पाँच तरीक़े, और वैलिडेट व रीट्राई करने वाला Python लूप। - [कॉल सेंटर रिकॉर्डिंग कैसे ट्रांसक्राइब करें: अरबी-रेडी पाइपलाइन](https://broadrouter.com/hi/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): स्टीरियो कॉल को चैनल से अलग करें, बोलने के turns काटें, अरबी और अंग्रेज़ी ट्रांसक्राइब करें, PII redact करें और LLM से सारांश बनाएँ। लागत के साथ Python पाइपलाइन। - [कोडिंग एजेंट में ओपन मॉडल: Aider, Cline, opencode, Codex](https://broadrouter.com/hi/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Aider, Cline, Roo Code, opencode, Continue और Codex CLI को किसी भी OpenAI-कम्पैटिबल एंडपॉइंट से जोड़ें। हर एक के फ़ील्ड के सटीक नाम, और कौन-सी सेटिंग टूटती हैं। - [बिज़नेस के लिए वॉइस नोट से टेक्स्ट: WhatsApp और Telegram से टिकट](https://broadrouter.com/hi/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): WhatsApp Business और Telegram बॉट पर आए ग्राहकों के वॉइस नोट को ट्रांसक्रिप्ट और सपोर्ट टिकट में बदलें: डाउनलोड, ट्रांसक्राइब, LLM से फ़ील्ड निकालें, रूट करें। - [छोटे प्रोजेक्ट के लिए साप्ताहिक LLM प्लान: Orbit बनाम घंटा बनाम महीना](https://broadrouter.com/hi/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): एक हफ़्ते का अनलिमिटेड DeepSeek V4.1 Flash प्लान कब घंटे वाले सेशन, मासिक प्लान या प्रति-टोकन भुगतान से बेहतर है: हैकथॉन, स्प्रिंट और माइग्रेशन का हिसाब। - [LLM API प्लान में “अनलिमिटेड” का असली मतलब क्या है?](https://broadrouter.com/hi/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): टोकन मीटर, 5 घंटे की रोलिंग विंडो, साप्ताहिक सीमा, रेट लिमिट, कॉन्करेंसी: ये सात सवाल बताते हैं कि “अनलिमिटेड” AI प्लान आपके एजेंट के लिए सचमुच अनलिमिटेड है या नहीं। - [Whisper API का विकल्प: कब बदलें, और उसे कैसे टेस्ट करें](https://broadrouter.com/hi/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): OpenAI Whisper API का होस्टेड विकल्प कब सही है: आपके ऑडियो पर word error rate, भाषाएँ, प्रति घंटा कीमत, सीमाएँ और प्राइवेसी, साथ में टेस्ट कोड। - [ज़ीरो डेटा रिटेंशन वाला LLM API: 10 सवालों की चेकलिस्ट](https://broadrouter.com/hi/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): LLM API के लिए “ज़ीरो डेटा रिटेंशन” का क्या मतलब होना चाहिए: प्रॉम्प्ट, आउटपुट, लॉग, ट्रेनिंग, सबप्रोसेसर और ऑडिट। डेटा भेजने से पहले पूछने लायक दस सवाल। - [Guides in Hindi](https://broadrouter.com/hi/guides/) - [BroadNet Inference के साथ OpenAI Python SDK इस्तेमाल करें](https://broadrouter.com/hi/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): सिर्फ़ base_url बदलकर आधिकारिक OpenAI Python SDK को BroadNet Inference से जोड़ें। इंस्टॉल करें, key सेट करें, पहला अनुरोध भेजें और जवाब स्ट्रीम करें। - [BroadNet Inference के साथ OpenAI Node.js SDK इस्तेमाल करें](https://broadrouter.com/hi/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): baseURL सेट करके आधिकारिक OpenAI Node.js और TypeScript SDK को BroadNet Inference से जोड़ें: इंस्टॉल, पहला अनुरोध, टोकन स्ट्रीमिंग और रिक्वेस्ट ID लॉग करना। - [BroadNet Inference को ब्राउज़र में आज़माएँ, बिना अकाउंट](https://broadrouter.com/hi/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): BroadNet के ब्राउज़र प्लेग्राउंड में बिना साइन-अप या कार्ड के अरबी स्पीच-टू-टेक्स्ट, डॉक्यूमेंट रीडिंग और Gemma 4 चैट टेस्ट करें, फिर वही अनुरोध API पर ले जाएँ। - [curl से BroadNet Inference को पहला अनुरोध](https://broadrouter.com/hi/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): curl से BroadNet Inference की पहली कॉल करें: एक chat completion, एक स्ट्रीम किया गया जवाब और एक ऑडियो ट्रांसक्रिप्शन, फिर रिस्पॉन्स, usage और एरर पढ़ना सीखें। - [chat completions स्ट्रीम करें और टूटे कनेक्शन सँभालें](https://broadrouter.com/hi/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): BroadNet Inference पर स्ट्रीमिंग कैसे काम करती है: Server-Sent Events, [DONE] मार्कर, आख़िरी चंक में टोकन usage, और स्ट्रीम बीच में ख़त्म हो जाए तो क्या करें। - [रीज़निंग मॉडल ख़ाली जवाब क्यों लौटाता है](https://broadrouter.com/hi/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3 और Qwen3.8 27B जवाब से पहले सोचते हैं, और यह रीज़निंग max_tokens में गिनी जाती है। ख़ाली जवाब आना ऐसे रोकें। - [BroadNet Inference के एरर कोड और रीट्राई कब करें](https://broadrouter.com/hi/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): BroadNet Inference API के हर HTTP स्टेटस का मतलब, किसे बैकऑफ़ के साथ रीट्राई करें, किसे पहले ठीक करें, और सपोर्ट को रिक्वेस्ट ID के साथ क्या भेजें। - [IP allow-list से API key को अपने सर्वर तक सीमित करें](https://broadrouter.com/hi/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): IP allow-list से BroadNet Inference की key सुरक्षित करें, इन्फ़रेंस key को पोर्टल क्रेडेंशियल से अलग रखें और लीक हुई key सेकंडों में रद्द करें। curl उदाहरणों के साथ। - [एजेंट के लिए अनलिमिटेड LLM API प्लान कैसे काम करते हैं](https://broadrouter.com/hi/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): अनलिमिटेड DeepSeek V4.1 Flash प्लान में न टोकन मीटर है, न साप्ताहिक सीमा। जानें कि सेशन, समानांतर एजेंट, key और बिलिंग असल में कैसे काम करते हैं। - [Aider को अनलिमिटेड DeepSeek V4.1 Flash पर चलाएँ](https://broadrouter.com/hi/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Aider कोडिंग एजेंट को उसकी OpenAI-कम्पैटिबल सेटिंग से BroadNet Inference से जोड़ें और फ़्लैट कीमत वाले अनलिमिटेड DeepSeek V4.1 Flash प्लान पर चलाएँ। - [BroadNet Inference के साथ LangChain इस्तेमाल करें](https://broadrouter.com/hi/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): base_url और api_key से LangChain के ChatOpenAI को BroadNet Inference से जोड़ें, और जानें कि LangChain रीज़निंग मॉडल के जवाब का क्या रखता है और क्या नहीं। - [Cline में BroadNet Inference (OpenAI Compatible प्रोवाइडर)](https://broadrouter.com/hi/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Cline VS Code एक्सटेंशन को उसके OpenAI Compatible प्रोवाइडर से BroadNet Inference पर सेट करें: सटीक फ़ील्ड, मॉडल ID, कॉन्टेक्स्ट साइज़ और आम समस्याओं के हल। - [Roo Code में BroadNet Inference (OpenAI Compatible)](https://broadrouter.com/hi/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code मई 2026 में बंद हो गया। मौजूदा इंस्टॉल और फ़ोर्क के लिए सेटिंग, और BroadNet Inference के साथ आगे कहाँ जाएँ (Cline या Zoo Code फ़ोर्क)। - [Continue में BroadNet Inference (config.yaml और apiBase)](https://broadrouter.com/hi/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): config.yaml में provider openai और apiBase से Continue IDE एक्सटेंशन में BroadNet Inference के मॉडल जोड़ें, tool_use क्षमता और कॉन्टेक्स्ट लंबाई समेत। - [opencode में BroadNet Inference को कस्टम प्रोवाइडर के रूप में जोड़ें](https://broadrouter.com/hi/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): opencode.jsonc में BroadNet Inference को opencode कोडिंग एजेंट के कस्टम OpenAI-कम्पैटिबल प्रोवाइडर के रूप में जोड़ें, मॉडल alias, capabilities और limit के साथ। - [Open WebUI को BroadNet Inference से जोड़ें](https://broadrouter.com/hi/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Open WebUI में BroadNet Inference को OpenAI-compatible कनेक्शन की तरह जोड़ें, मॉडल ID allowlist में डालें, और verify, ख़ाली जवाब व रेट लिमिट की समस्याएँ ठीक करें। - [LibreChat में BroadNet Inference को कस्टम एंडपॉइंट के रूप में जोड़ें](https://broadrouter.com/hi/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): librechat.yaml में BroadNet Inference को OpenAI-compatible कस्टम एंडपॉइंट बनाएँ: .env में key, मॉडल लिस्ट, टाइटल मॉडल और आम समस्याओं के समाधान। - [n8n में BroadNet Inference इस्तेमाल करें (OpenAI credential, Base URL)](https://broadrouter.com/hi/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): n8n के OpenAI credential में Base URL फ़ील्ड से BroadNet Inference सेट करें, फिर OpenAI Chat Model नोड को ID से चलाएँ और Responses API टॉगल बंद रखें। - [BroadNet Inference के साथ Vercel AI SDK इस्तेमाल करें](https://broadrouter.com/hi/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): @ai-sdk/openai-compatible के createOpenAICompatible से Vercel AI SDK में BroadNet Inference कॉल करें: generateText, streamText, टूल और reasoning effort। - [अरबी speech-to-text API, स्टेप बाय स्टेप](https://broadrouter.com/hi/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): OpenAI-compatible एंडपॉइंट के ज़रिए BroadNet Transcribe Arabic से अरबी ऑडियो ट्रांसक्राइब करें। फ़ॉर्मैट, language फ़ील्ड, Python और curl उदाहरण, और कीमत। - [LiteLLM को BroadNet Inference पर रूट करें (SDK और प्रॉक्सी)](https://broadrouter.com/hi/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): openai/ मॉडल प्रीफ़िक्स और api_base के साथ LiteLLM से BroadNet Inference कॉल करें, Python SDK में या प्रॉक्सी के config.yaml में, मेल खाती रेट लिमिट के साथ। - [BroadNet Inference पर OpenAI Agents SDK चलाएँ](https://broadrouter.com/hi/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): AsyncOpenAI और OpenAIChatCompletionsModel के ज़रिए Python वाले OpenAI Agents SDK को BroadNet Inference से जोड़ें, और OpenAI key माँगने वाली ट्रेसिंग बंद करें। - [API से WhatsApp वॉइस नोट को टेक्स्ट में बदलें](https://broadrouter.com/hi/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): WhatsApp के .ogg वॉइस नोट या फ़ोन की .m4a रिकॉर्डिंग को ffmpeg से WAV में बदलें और एक API कॉल से अरबी या 13 दूसरी भाषाओं में ट्रांसक्राइब करें। - [n8n में WhatsApp वॉइस नोट को टेक्स्ट में बदलें](https://broadrouter.com/hi/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): एक n8n वर्कफ़्लो जो WhatsApp वॉइस नोट लेता है, Meta से डाउनलोड करता है, ffmpeg से बदलता है और अरबी या 13 दूसरी भाषाओं में ट्रांसक्राइब करता है। - [विज़न मॉडल से अरबी इनवॉइस और रसीदें पढ़ें](https://broadrouter.com/hi/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): इनवॉइस या रसीद की फ़ोटो BroadNet Inference के विज़न मॉडल को भेजें और कुल रक़म, तारीख़ और विक्रेता JSON में पाएँ, सवाल अरबी या अंग्रेज़ी में पूछें। - [Zed में BroadNet Inference को OpenAI-कम्पैटिबल प्रोवाइडर बनाएँ](https://broadrouter.com/hi/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Zed एडिटर की settings.json में BroadNet Inference को OpenAI-कम्पैटिबल प्रोवाइडर के रूप में जोड़ें: मॉडल ID, टोकन सीमाएँ, capabilities और key वेरिएबल। - [Kilo Code में BroadNet Inference (OpenAI Compatible प्रोवाइडर)](https://broadrouter.com/hi/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Kilo Code को BroadNet Inference से कस्टम OpenAI Compatible प्रोवाइडर के रूप में जोड़ें, VS Code डायलॉग या kilo.json में: मॉडल ID, सीमाएँ और टूल फ़्लैग के साथ। - [goose में BroadNet Inference को कस्टम प्रोवाइडर के रूप में जोड़ें](https://broadrouter.com/hi/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): goose AI एजेंट को BroadNet Inference से OpenAI Compatible कस्टम प्रोवाइडर के रूप में जोड़ें: goose Desktop, goose configure या एक JSON फ़ाइल से। - [BroadNet Inference के साथ Pydantic AI (OpenAICompatibleProvider)](https://broadrouter.com/hi/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): OpenAIChatModel और OpenAICompatibleProvider से BroadNet Inference पर Pydantic AI एजेंट चलाएँ, साथ में टूल कॉल, स्ट्रक्चर्ड आउटपुट और मॉडल प्रोफ़ाइल के समाधान। - [CrewAI में BroadNet Inference को कस्टम OpenAI एंडपॉइंट बनाएँ](https://broadrouter.com/hi/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): LLM(custom_openai=True, base_url=...) से CrewAI एजेंट को BroadNet Inference पर लाएँ, फिर टूल, लंबे कॉन्टेक्स्ट और फ़्लैट-रेट प्लान के लिए मॉडल चुनें। - [OpenAILike के ज़रिए BroadNet Inference के साथ LlamaIndex](https://broadrouter.com/hi/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): OpenAILike क्लास से LlamaIndex को BroadNet Inference से जोड़ें, और api_base, context_window, is_chat_model और is_function_calling_model सही सेट करें। - [OpenAI-API-compatible प्लगइन से Dify में BroadNet Inference जोड़ें](https://broadrouter.com/hi/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): OpenAI-API-compatible मॉडल प्रोवाइडर प्लगइन से Dify को BroadNet Inference से जोड़ें: सटीक फ़ील्ड वैल्यू, कॉन्टेक्स्ट साइज़ और function calling सेटिंग के साथ। - [Jan में BroadNet Inference को कस्टम OpenAI-compatible प्रोवाइडर बनाएँ](https://broadrouter.com/hi/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Jan डेस्कटॉप ऐप में BroadNet Inference को कस्टम OpenAI-compatible प्रोवाइडर के रूप में जोड़ें, मॉडल ID हाथ से डालें, और हर मॉडल के लिए टूल और विज़न चालू करें। - [SillyTavern को BroadNet Inference से जोड़ें (Custom OpenAI-compatible)](https://broadrouter.com/hi/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): SillyTavern को Custom (OpenAI-compatible) Chat Completion सोर्स से BroadNet Inference से जोड़ें: सही मॉडल ID, प्रॉम्प्ट पोस्ट-प्रोसेसिंग और टोकन सीमाएँ। - [Spring AI को BroadNet Inference के साथ चलाएँ (OpenAI base-url)](https://broadrouter.com/hi/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Spring AI के OpenAI चैट स्टार्टर को spring.ai.openai.base-url, BroadNet मॉडल ID और max-tokens से BroadNet Inference पर लगाएँ, फिर ChatClient से कॉल करें। - [openai-go SDK से Go में BroadNet Inference कॉल करें](https://broadrouter.com/hi/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): option.WithBaseURL से आधिकारिक openai-go v3 SDK को BroadNet Inference से जोड़ें: Go में chat completions, usage के साथ स्ट्रीमिंग और एरर हैंडलिंग। - [BroadNet Inference पर LangGraph एजेंट बनाएँ](https://broadrouter.com/hi/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): ChatOpenAI, base_url और BroadNet मॉडल ID से LangGraph का StateGraph या create_agent BroadNet Inference पर चलाएँ, साथ में टूल, स्ट्रीमिंग और मेमोरी। - [BroadNet Inference के साथ Haystack पाइपलाइन और एजेंट चलाएँ](https://broadrouter.com/hi/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Haystack के OpenAIChatGenerator को api_base_url, BroadNet मॉडल ID और Secret key से BroadNet Inference पर सेट करें: पाइपलाइन, टूल, स्ट्रीमिंग और एजेंट के लिए। - [Semantic Kernel को BroadNet Inference के साथ इस्तेमाल करें](https://broadrouter.com/hi/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Microsoft Semantic Kernel को BroadNet Inference से जोड़ें: Python में AsyncOpenAI क्लाइंट से, या C# में कस्टम endpoint से, चैट, प्लगइन और एजेंट के लिए। - [BroadNet Inference के साथ Mastra एजेंट इस्तेमाल करें](https://broadrouter.com/hi/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): id, url और apiKey वाले model ऑब्जेक्ट से Mastra एजेंट BroadNet Inference पर चलाएँ, साथ में वह दोहरा broadnet/ प्रीफ़िक्स जो मॉडल ID को चाहिए। - [BroadNet Inference पर Instructor से स्ट्रक्चर्ड आउटपुट पाएँ](https://broadrouter.com/hi/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Instructor से BroadNet Inference पर Pydantic ऑब्जेक्ट निकालें: कस्टम base_url के साथ from_openai या from_provider इस्तेमाल करें, और सही Mode चुनें। - [DSPy में BroadNet Inference (model_type="chat" सेट करें)](https://broadrouter.com/hi/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): openai/ प्रीफ़िक्स, api_base और model_type="chat" से BroadNet Inference के लिए dspy.LM सेट करें, क्योंकि DSPy 3.5 OpenAI मॉडल को Responses API पर भेजता है। - [promptfoo से BroadNet मॉडल का इवैल्युएशन करें](https://broadrouter.com/hi/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): openai:chat प्रोवाइडर, apiBaseUrl और apiKeyEnvar से BroadNet Inference पर promptfoo evals और मॉडलों के आमने-सामने टेस्ट चलाएँ, ग्रेडर के साथ। - [OpenAIModel के साथ smolagents को BroadNet Inference पर चलाएँ](https://broadrouter.com/hi/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Hugging Face smolagents को OpenAIModel और api_base से BroadNet Inference से जोड़ें: किसी भी मॉडल पर CodeAgent, या GLM 5.3 और Qwen3.8 27B पर ToolCallingAgent। - [Flowise को BroadNet Inference से जोड़ें (OpenAI Custom Model)](https://broadrouter.com/hi/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Flowise चैटफ़्लो में OpenAI Custom Model नोड से BroadNet Inference इस्तेमाल करें: Base Path, OpenAI API क्रेडेंशियल और हाथ से टाइप किया BroadNet मॉडल नाम। - [Langflow को BroadNet Inference से जोड़ें](https://broadrouter.com/hi/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Langflow के OpenAI कंपोनेंट में OpenAI API Base फ़ील्ड और टाइप किए मॉडल नाम से BroadNet Inference इस्तेमाल करें, और जानें OpenAI Compatible प्रोवाइडर क्यों नहीं। - [Glossary entries in Hindi](https://broadrouter.com/hi/glossary/) - [AI एजेंट (AI agent)](https://broadrouter.com/hi/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): AI एजेंट ऐसा सिस्टम है जिसमें लैंग्वेज मॉडल किसी लक्ष्य की ओर अपने क़दम ख़ुद तय करता है: कौन-से टूल कॉल करने हैं, चुनता है, नतीजे पढ़ता है और अंतिम जवाब या किसी सीमा तक दोहराता रहता है। तय बिंदुओं पर मॉडल को कॉल करने वाली स्क्रिप्ट को आम तौर पर workflow कहते हैं। - [API key](https://broadrouter.com/hi/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): API key एक गुप्त स्ट्रिंग है जो API कॉल करने वाले क्लाइंट को पहचानती और अधिकृत करती है। LLM API आम तौर पर इसे Authorization: Bearer हेडर में चाहते हैं, और जिसके पास भी यह key हो, वह ऐसे अनुरोध कर सकता है जिनका बिल key वाले अकाउंट पर बनता है। - [अरबी डायक्रिटिक्स (tashkeel)](https://broadrouter.com/hi/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): अरबी डायक्रिटिक्स, या tashkeel, अक्षरों के ऊपर या नीचे लगने वाले छोटे चिह्न हैं जो छोटे स्वर, दोहरे व्यंजन और स्वर के न होने को दिखाते हैं। ये वैकल्पिक हैं: ख़बरों समेत ज़्यादातर अरबी टेक्स्ट इन्हें छोड़ देता है, इसलिए पाठक और सॉफ़्टवेयर संदर्भ से स्वर का अंदाज़ा लगाते हैं। - [अरबी बोली की पहचान (Arabic dialect identification)](https://broadrouter.com/hi/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): अरबी बोली की पहचान वह टास्क है जिसमें तय किया जाता है कि कोई टेक्स्ट या रिकॉर्डिंग किस अरबी रूप में है, जैसे मिस्री, खाड़ी, लेवेंटाइन, मग़रिबी या मॉडर्न स्टैंडर्ड अरबी, और यह क्षेत्र, देश या शहर के स्तर पर होता है। इसका इस्तेमाल मैसेज रूट करने, प्रॉम्प्ट चुनने और मूल्यांकन सेट बाँटने में होता है। - [अरबी टेक्स्ट नॉर्मलाइज़ेशन (Arabic text normalization)](https://broadrouter.com/hi/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): अरबी टेक्स्ट नॉर्मलाइज़ेशन वह चरण है जो मिलान से पहले एक ही शब्द की अलग-अलग वर्तनियों को एक रूप में बदलता है: डायक्रिटिक्स और tatweel हटाना, alef के रूप एक करना, और अक्सर ى को ي और ة को ه करना। यह सर्च key और मूल्यांकन पर लगता है, दिखाए जाने वाले टेक्स्ट पर नहीं। - [Arabizi (अरबीज़ी)](https://broadrouter.com/hi/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): Arabizi लैटिन अक्षरों में अनौपचारिक रूप से लिखी गई अरबी है, जिसमें उन अरबी ध्वनियों के लिए अंक लगते हैं जिनका लैटिन में कोई बराबर अक्षर नहीं, जैसे ع के लिए 3। यह चैट और सोशल मीडिया में आम है, ज़्यादातर बोलचाल की होती है, इसकी तय वर्तनी नहीं, और यह अक्सर अंग्रेज़ी या फ़्रेंच से मिली होती है। - [AWQ (activation-aware weight quantization)](https://broadrouter.com/hi/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (activation-aware weight quantization) एक post-training तरीक़ा है जो LLM के weights को कम प्रिसिज़न, आम तौर पर 4-bit integers, में दबाता है। यह छोटे calibration सेट के activation आँकड़ों से सबसे अहम weight चैनल पहचानता है और उन्हें scale करता है, ताकि उनकी सटीकता कम घटे। इसे backpropagation की ज़रूरत नहीं। - [base URL](https://broadrouter.com/hi/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): base URL किसी API का तय मूल पता है, जैसे https://api.example.com/v1, जिसे क्लाइंट /chat/completions जैसे हर एंडपॉइंट पाथ के आगे जोड़ता है। OpenAI-compatible SDK में base URL को API key और मॉडल ID के साथ बदलने से वही कोड किसी दूसरे प्रोवाइडर से जुड़ जाता है। - [Batch API (बैच API)](https://broadrouter.com/hi/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Batch API एक asynchronous LLM एंडपॉइंट है जो एक साथ कई अनुरोध लेता है, आम तौर पर एक फ़ाइल के रूप में, उन्हें तुरंत नहीं बल्कि एक तय समय-सीमा में प्रोसेस करता है, और नतीजे एक साथ लौटाता है। OpenAI और Anthropic जैसे प्रोवाइडर batch अनुरोधों की कीमत अपने सामान्य API से कम रखते हैं। - [BM25](https://broadrouter.com/hi/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) कीवर्ड सर्च का एक रैंकिंग फ़ंक्शन है जो दस्तावेज़ को तीन बातों से स्कोर करता है: क्वेरी के शब्द उसमें कितनी बार आते हैं, वे पूरे कलेक्शन में कितने दुर्लभ हैं, और दस्तावेज़ कितना लंबा है। यह Lucene और Elasticsearch की डिफ़ॉल्ट similarity है। - [बाइट पेयर एन्कोडिंग (byte pair encoding, BPE)](https://broadrouter.com/hi/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Byte pair encoding (BPE) एक subword tokenization एल्गोरिदम है जो अकेले कैरेक्टर या bytes से शुरू होता है और सबसे आम पास-पास के जोड़े को बार-बार मिलाकर नई शब्दावली एंट्री बनाता है, जब तक लक्ष्य शब्दावली आकार न आ जाए। आज के ज़्यादातर LLM tokenizer BPE या उसका byte-level रूप इस्तेमाल करते हैं। - [कैश्ड इनपुट टोकन (cached input tokens)](https://broadrouter.com/hi/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): कैश्ड इनपुट टोकन वे प्रॉम्प्ट टोकन हैं जिन्हें LLM प्रोवाइडर ने हाल ही में प्रोसेस किया है और दोबारा गणना करने के बजाय कैश से फिर इस्तेमाल कर सकता है, आम तौर पर सिस्टम प्रॉम्प्ट या डॉक्यूमेंट जैसा दोहराया गया प्रीफ़िक्स। प्रोवाइडर इनका बिल नए इनपुट टोकन से कम दर पर बनाते हैं। - [चेन-ऑफ़-थॉट (chain-of-thought, CoT)](https://broadrouter.com/hi/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Chain-of-thought (CoT) एक प्रॉम्प्टिंग तकनीक है जिसमें लैंग्वेज मॉडल को अंतिम जवाब से पहले बीच के रीज़निंग चरण लिखकर दिखाए जाते हैं या लिखने को कहा जाता है। इससे गणित और तर्क जैसी कई-चरण वाली समस्याओं पर सटीकता बढ़ती है, और रीज़निंग मॉडल अब इसे अपने-आप बनाते हैं। - [कैरेक्टर एरर रेट (character error rate, CER)](https://broadrouter.com/hi/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Character error rate (CER) ट्रांसक्रिप्शन या OCR की सटीकता कैरेक्टर स्तर पर मापता है: आउटपुट को रेफ़रंस टेक्स्ट बनाने के लिए ज़रूरी substitutions, deletions और insertions, भाग रेफ़रंस के कैरेक्टर की संख्या। कम बेहतर है, और 0 पूरा मेल है। - [Chat Completions API](https://broadrouter.com/hi/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Chat Completions API, OpenAI का HTTP एंडपॉइंट POST /v1/chat/completions है, जो एक मॉडल का नाम और रोल वाले मैसेज की सूची लेता है और मॉडल का अगला मैसेज लौटाता है। इसके अनुरोध और जवाब का ढाँचा वह आम फ़ॉर्मैट बन गया जिसे OpenAI-कम्पैटिबल प्रोवाइडर और SDK अपनाते हैं। - [चैट टेम्पलेट (chat template)](https://broadrouter.com/hi/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): चैट टेम्पलेट वह फ़ॉर्मैटिंग नियम है, आम तौर पर Jinja में लिखा और मॉडल के tokenizer के साथ आने वाला, जो roles वाले चैट मैसेज की सूची को उस एक टोकन क्रम में बदलता है जिसकी मॉडल उम्मीद करता है, उसके special control टोकन समेत। ग़लत टेम्पलेट से मॉडल के जवाब बहुत बिगड़ सकते हैं। - [चंकिंग (chunking, RAG)](https://broadrouter.com/hi/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Chunking retrieval पाइपलाइन का वह चरण है जो दस्तावेज़ों को embed और इंडेक्स करने से पहले छोटे टुकड़ों, यानी chunks, में बाँटता है। Chunk size तय करता है कि हर टुकड़े में कितना टेक्स्ट हो, और chunk overlap पड़ोसी chunks के बीच कुछ टेक्स्ट दोहराता है ताकि सीमा पर कटे विचार न खोएँ। - [कोड-स्विचिंग (code-switching)](https://broadrouter.com/hi/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Code-switching एक ही बातचीत या वाक्य में, बोलकर या लिखकर, दो या ज़्यादा भाषाओं, या एक भाषा के अलग रूपों, का इस्तेमाल है। उदाहरण: अरबी बोलने वाले अपनी बोली में अंग्रेज़ी शब्द मिलाते हैं, या मॉडर्न स्टैंडर्ड अरबी और अपनी बोली के बीच आते-जाते हैं। - [AI कोडिंग एजेंट (AI coding agent)](https://broadrouter.com/hi/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): AI कोडिंग एजेंट एक सॉफ़्टवेयर टूल है जिसमें लैंग्वेज मॉडल कोडबेस पर लूप में काम करता है: फ़ाइलें पढ़ता है, उन्हें बदलता है, टेस्ट जैसी कमांड चलाता है, आउटपुट पढ़ता है और अगला क़दम तय करता है, जब तक टास्क पूरा न हो या कोई सीमा न आ जाए। - [कॉन्करेंसी (concurrency)](https://broadrouter.com/hi/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): कॉन्करेंसी, किसी क्लाइंट के एक ही समय पर चल रहे (in-flight) अनुरोधों की संख्या है। LLM API अक्सर इसकी सीमा हर key या प्लान पर, प्रति मिनट अनुरोधों से अलग रखते हैं, क्योंकि हर खुला अनुरोध जवाब पूरा होने तक GPU क्षमता घेरे रहता है। - [कॉन्टेक्स्ट कॉम्पैक्शन (context compaction)](https://broadrouter.com/hi/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): Context compaction वह तरीक़ा है जिसमें कॉन्टेक्स्ट विंडो की सीमा के पास पहुँची बातचीत का सारांश बनाया जाता है और सेशन उसी सारांश से आगे चलता है, ताकि एजेंट मुख्य फ़ैसले और स्थिति रखे और पुराने टूल आउटपुट जैसी कच्ची हिस्ट्री छोड़ दे। - [कॉन्टेक्स्ट इंजीनियरिंग (context engineering)](https://broadrouter.com/hi/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): कॉन्टेक्स्ट इंजीनियरिंग वह काम है जिसमें तय किया और सँभाला जाता है कि लैंग्वेज मॉडल हर अनुरोध में कौन-से टोकन देखे: सिस्टम प्रॉम्प्ट, टूल डेफ़िनिशन, इतिहास, रिट्रीव किए डॉक्यूमेंट और नोट्स, ताकि मॉडल को काम के लिए ज़रूरी, सबसे छोटी और सबसे काम की जानकारी मिले। - [कॉन्टेक्स्ट रॉट (context rot)](https://broadrouter.com/hi/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): कॉन्टेक्स्ट रॉट सटीकता और भरोसेमंदी का वह धीरे-धीरे होने वाला नुक़सान है जो लैंग्वेज मॉडल में इनपुट लंबा होने पर दिखता है, तब भी जब काम वही रहे और इनपुट कॉन्टेक्स्ट विंडो के अंदर ही हो। यह शब्द 2025 की Chroma तकनीकी रिपोर्ट से मशहूर हुआ, जिसने 18 मॉडल टेस्ट किए। - [कॉन्टेक्स्ट विंडो (context window)](https://broadrouter.com/hi/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): कॉन्टेक्स्ट विंडो टोकन की वह अधिकतम संख्या है जिसके साथ लैंग्वेज मॉडल एक अनुरोध में काम कर सकता है। इसमें प्रॉम्प्ट, बातचीत का इतिहास, टूल डेफ़िनिशन, इमेज और जनरेट हुआ आउटपुट, रीज़निंग टोकन समेत, सब आता है। जो इसमें न समाए, उसे काटना, सारांश बनाना या कई अनुरोधों में बाँटना पड़ता है। - [डेटा रेज़िडेंसी (data residency)](https://broadrouter.com/hi/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): डेटा रेज़िडेंसी वह शर्त या वादा है कि डेटा किसी तय देश या क्षेत्र में ही स्टोर और प्रोसेस होगा। LLM API के लिए यह बताती है कि इन्फ़रेंस के दौरान प्रॉम्प्ट कहाँ प्रोसेस होते हैं और कोई स्टोर किया कंटेंट कहाँ रहता है; अक्सर अकाउंट और बिलिंग डेटा इसमें शामिल नहीं होता। - [अरबी बोलियाँ (dialectal Arabic)](https://broadrouter.com/hi/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): अरबी बोलियाँ किसी क्षेत्र या शहर में रोज़ बोली जाने वाली अरबी हैं, जैसे मिस्री, खाड़ी, लेवेंटाइन या मग़रिबी अरबी। ये आधुनिक मानक अरबी से आवाज़ों, शब्दों और व्याकरण में अलग हैं, इनकी कोई मानक वर्तनी नहीं, और ऑनलाइन इन्हें तेज़ी से लिखा जा रहा है, जिससे लैंग्वेज मॉडल के लिए ये मुश्किल हैं। - [स्पीकर डायराइज़ेशन (speaker diarization)](https://broadrouter.com/hi/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): स्पीकर डायराइज़ेशन ऑडियो रिकॉर्डिंग को बोलने वालों के हिसाब से हिस्सों में बाँटने का काम है, जो बताता है कि "कौन कब बोला"। यह बोली के हर हिस्से पर SPEAKER_00 जैसा गुमनाम स्पीकर ID लगाता है, और आम तौर पर ट्रांसक्रिप्ट के साथ मिलाकर दिखाता है कि हर लाइन किसने कही। - [डिस्टिलेशन (knowledge distillation)](https://broadrouter.com/hi/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): डिस्टिलेशन, या नॉलेज डिस्टिलेशन, ट्रेनिंग का वह तरीक़ा है जिसमें एक छोटा स्टूडेंट मॉडल बड़े टीचर मॉडल के आउटपुट, जैसे उसके प्रायिकता वितरण या जनरेट किए जवाब, दोहराना सीखता है, ताकि स्टूडेंट कम लागत पर चलते हुए टीचर की काफ़ी क्वालिटी बनाए रखे। - [डेटा प्रोसेसिंग एग्रीमेंट (DPA)](https://broadrouter.com/hi/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): डेटा प्रोसेसिंग एग्रीमेंट (DPA) कंट्रोलर और प्रोसेसर के बीच का अनुबंध है, जो तय करता है कि प्रोसेसर पर्सनल डेटा कैसे संभालेगा। GDPR अनुच्छेद 28(3) इसे ज़रूरी बनाता है और इसकी सामग्री गिनाता है: लिखित निर्देश, गोपनीयता, सुरक्षा, सबप्रोसेसर के नियम, डेटा मिटाना या लौटाना, और ऑडिट। - [एम्बेडिंग (embeddings)](https://broadrouter.com/hi/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): एम्बेडिंग, फ़्लोटिंग-पॉइंट संख्याओं की सूचियाँ (वेक्टर) हैं जो कोई मॉडल टेक्स्ट, इमेज या दूसरे इनपुट को दर्शाने के लिए बनाता है। मिलते-जुलते अर्थ वाले इनपुट के वेक्टर पास-पास होते हैं, इसलिए वेक्टर के बीच की दूरी से सर्च नतीजों की रैंकिंग, दस्तावेज़ों का समूह बनाना या जुड़ी चीज़ें ढूँढना संभव होता है। - [Eval (LLM इवैल्यूएशन)](https://broadrouter.com/hi/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Eval (evaluation का छोटा रूप) किसी लैंग्वेज मॉडल या प्रॉम्प्ट का दोहराने लायक़ टेस्ट है: इनपुट का एक तय सेट, अच्छे आउटपुट का वर्णन, और एक ग्रेडर (कोड, इंसान या दूसरा मॉडल) जो हर आउटपुट को स्कोर करता है, ताकि अलग-अलग वर्ज़न की तुलना हो सके। - [एक्सपोनेंशियल बैकऑफ़ (exponential backoff)](https://broadrouter.com/hi/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): एक्सपोनेंशियल बैकऑफ़ री-ट्राई की एक रणनीति है जिसमें क्लाइंट हर नाकाम कोशिश के बाद, दोबारा कोशिश करने से पहले, ज़्यादा देर इंतज़ार करता है, आम तौर पर देरी को एक सीमा तक दोगुना करते हुए। हर इंतज़ार में रैंडम jitter जोड़ने से कई क्लाइंट के री-ट्राई बिखर जाते हैं, ताकि वे एक ही पल सर्वर पर न पहुँचें। - [फ़्यू-शॉट प्रॉम्प्टिंग (few-shot prompting)](https://broadrouter.com/hi/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): फ़्यू-शॉट प्रॉम्प्टिंग वह तकनीक है जिसमें प्रॉम्प्ट असली इनपुट से पहले इनपुट-आउटपुट के कुछ हल किए उदाहरण देता है, ताकि लैंग्वेज मॉडल उनसे काम और फ़ॉर्मैट समझ ले। मॉडल के वेट नहीं बदलते; ज़ीरो-शॉट में कोई उदाहरण नहीं होता और वन-शॉट में एक। - [फ़ाइन-ट्यूनिंग (fine-tuning)](https://broadrouter.com/hi/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): फ़ाइन-ट्यूनिंग पहले से ट्रेन हुए लैंग्वेज मॉडल को किसी छोटे, काम-विशेष डेटासेट, जैसे उदाहरण वाली बातचीत, पर आगे ट्रेन करने की प्रक्रिया है, ताकि उसके वेट बदलें और वह हर प्रॉम्प्ट में लंबे निर्देश या उदाहरण दिए बिना वह काम या स्टाइल ज़्यादा भरोसे से करे। - [finish_reason (जवाब रुकने की वजह)](https://broadrouter.com/hi/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason, chat completion जवाब की हर choice में एक फ़ील्ड है जो बताता है कि मॉडल ने जनरेट करना क्यों रोका। OpenAI फ़ॉर्मैट में इसकी वैल्यू stop, length, tool_calls, content_filter और पुरानी (deprecated) function_call हैं। Anthropic का Messages API यही बात stop_reason में बताता है। - [फ़्लैट-रेट LLM प्लान (flat-rate LLM plan)](https://broadrouter.com/hi/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): फ़्लैट-रेट LLM प्लान किसी लैंग्वेज मॉडल API तक पहुँच के लिए प्रति टोकन बिल बनाने के बजाय एक अवधि, जैसे एक महीने, की तय कीमत लेता है। ऐसे प्लान आम तौर पर कुछ और सीमित करते हैं, जैसे concurrency, स्पीड, मॉडल या इस्तेमाल की समय-सीमा, इसलिए सीमाएँ कीमत जितनी ही अहम हैं। - [FP8 (8-बिट फ़्लोटिंग पॉइंट)](https://broadrouter.com/hi/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 डीप लर्निंग में इस्तेमाल होने वाला 8-बिट फ़्लोटिंग-पॉइंट नंबर फ़ॉर्मैट है, जिसकी दो एन्कोडिंग हैं: E4M3, जिसमें 4 एक्सपोनेंट और 3 मैंटिसा बिट हैं, और E5M2, जिसमें 5 एक्सपोनेंट और 2 मैंटिसा बिट हैं। यह 16-बिट फ़ॉर्मैट के मुक़ाबले मेमोरी आधी करता है और नए GPU पर नेटिव चलता है। - [फ़्रीक्वेंसी और प्रेज़ेंस पेनल्टी](https://broadrouter.com/hi/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): फ़्रीक्वेंसी और प्रेज़ेंस पेनल्टी सैंपलिंग सेटिंग हैं जो आउटपुट में पहले आ चुके टोकन का स्कोर घटाती हैं। प्रेज़ेंस पेनल्टी टोकन के एक बार भी आने पर लगती है; फ़्रीक्वेंसी पेनल्टी हर दोहराव के साथ बढ़ती है। OpenAI के API में दोनों की रेंज -2.0 से 2.0 है, डिफ़ॉल्ट 0। - [GDPR (जनरल डेटा प्रोटेक्शन रेगुलेशन)](https://broadrouter.com/hi/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): GDPR (General Data Protection Regulation, Regulation (EU) 2016/679) पर्सनल डेटा की प्रोसेसिंग पर EU का क़ानून है। यह 25 मई 2018 से लागू है और EU में स्थापित संगठनों के साथ EU के बाहर के कुछ संगठनों को भी कवर करता है। LLM API के लिए यह प्रोसेसर अनुबंध, अंतरराष्ट्रीय ट्रांसफ़र और डेटा मिनिमाइज़ेशन तय करता है। - [GGUF](https://broadrouter.com/hi/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF लैंग्वेज मॉडल स्टोर करने का एक बाइनरी फ़ाइल फ़ॉर्मैट है, जो llama.cpp जैसे GGML-आधारित रनटाइम पर इन्फ़रेंस के लिए बना है। एक GGUF फ़ाइल में वेट, अक्सर क्वांटाइज़्ड, और मॉडल लोड करने का ज़रूरी मेटाडेटा होता है, जैसे टोकनाइज़र और आर्किटेक्चर की जानकारी, और इसे तेज़ लोडिंग के लिए मेमोरी-मैप किया जा सकता है। - [GPU-घंटा (GPU-hour)](https://broadrouter.com/hi/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): GPU-घंटा एक ग्राफ़िक्स प्रोसेसर का एक घंटे तक इस्तेमाल है। क्लाउड प्रोवाइडर किराए के GPU इंस्टेंस का बिल घंटे या सेकंड के हिसाब से बनाते हैं, और रिसर्चर किसी जॉब का कंप्यूट GPU-घंटों में बताते हैं। आप मशीन के समय का भुगतान करते हैं, चाहे वह व्यस्त हो या नहीं। - [ग्राउंडिंग (grounding)](https://broadrouter.com/hi/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): ग्राउंडिंग लैंग्वेज मॉडल के आउटपुट को जानकारी के ठोस, जाँचे जा सकने वाले स्रोतों, जैसे दिए गए डॉक्यूमेंट, सर्च नतीजे या डेटाबेस रिकॉर्ड, से जोड़ने और मॉडल को उन्हीं से जवाब देने का निर्देश देने का तरीक़ा है। ग्राउंडेड जवाब का स्रोत खोजा जा सकता है और उसमें गढ़े हुए तथ्य कम होते हैं। - [गार्डरेल (LLM guardrails)](https://broadrouter.com/hi/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): गार्डरेल वे जाँचें हैं जो कोई ऐप्लिकेशन लैंग्वेज मॉडल कॉल के इर्द-गिर्द चलाता है, ताकि वह तय सीमाओं में रहे। इनपुट गार्डरेल मॉडल में जाने वाली चीज़ परखते हैं, आउटपुट गार्डरेल बाहर आने वाली, और टूल गार्डरेल कार्रवाई चलने से पहले उसे जाँचते हैं। जाँच फ़ेल होने पर अनुरोध रुकता, बदलता या फ़्लैग होता है। - [हैलुसिनेशन (hallucination)](https://broadrouter.com/hi/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): हैलुसिनेशन, लैंग्वेज मॉडल का ऐसा आउटपुट है जो सुनने में सही लगता है लेकिन ग़लत, गढ़ा हुआ या दिए गए इनपुट से असमर्थित होता है, जैसे नक़ली हवाला, ग़लत संख्या या मनगढ़ंत API। यह इसलिए होता है क्योंकि मॉडल तथ्य जाँचने के बजाय संभावित टेक्स्ट का अनुमान लगाता है। - [हम्ज़ा (hamza)](https://broadrouter.com/hi/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): हम्ज़ा (ء) ग्लॉटल स्टॉप के लिए अरबी अक्षर है। यह अकेले या किसी कुर्सी पर लिखा जाता है: अलिफ़ (أ إ), वाव (ؤ) या ये (ئ), और آ अलिफ़ को मद्दा के साथ जोड़ता है। टाइप किए टेक्स्ट में यह एक-सा नहीं लिखा जाता, सो सर्च इंजन और स्पीच स्कोरिंग अक्सर अलिफ़ वाले रूपों को ا में बदल देते हैं। - [हाइब्रिड सर्च (hybrid search)](https://broadrouter.com/hi/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): हाइब्रिड सर्च एक रिट्रीवल तरीक़ा है जो एक ही क्वेरी पर BM25 जैसी कीवर्ड सर्च और embeddings पर वेक्टर सर्च दोनों चलाता है, फिर दोनों रैंक की गई सूचियों को एक में मिलाता है, अक्सर reciprocal rank fusion (RRF) या normalized स्कोर के weighted sum से। - [Idempotency key (आइडेम्पोटेंसी key)](https://broadrouter.com/hi/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Idempotency key एक यूनिक वैल्यू है, आम तौर पर UUID जैसी रैंडम ID, जिसे क्लाइंट स्टेट बदलने वाले अनुरोध, जैसे पेमेंट या ख़रीद, के साथ जोड़ता है। अगर वही अनुरोध उसी key के साथ दोबारा आए, तो सर्वर काम दो बार करने के बजाय पहला नतीजा लौटाता है। - [इमेज टोकन (image tokens)](https://broadrouter.com/hi/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): इमेज टोकन वे इनपुट टोकन हैं जिनसे कोई vision language मॉडल प्रॉम्प्ट में एक इमेज को दर्शाता है। प्रोवाइडर इन्हें इमेज के साइज़ से निकालते हैं, आम तौर पर रीसाइज़ के बाद उसे तय साइज़ के patches या tiles में बाँटकर, और टेक्स्ट इनपुट टोकन की तरह बिल करते हैं, इसलिए बड़ी इमेज ज़्यादा महँगी पड़ती है। - [इन्फ़रेंस प्रोवाइडर (inference provider)](https://broadrouter.com/hi/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): इन्फ़रेंस प्रोवाइडर वह कंपनी है जो मशीन लर्निंग मॉडल, अक्सर ओपन-वेट LLM, अपने, लीज़ पर लिए या किराए के GPU पर चलाती है और API के ज़रिए उन तक पहुँच बेचती है, आम तौर पर प्रति टोकन, प्रति अनुरोध या प्रति समय-इकाई की कीमत पर। सर्विंग, स्केलिंग और हार्डवेयर वही सँभालती है। - [इनपुट और आउटपुट टोकन (input and output tokens)](https://broadrouter.com/hi/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): इनपुट टोकन वे टोकन हैं जिन्हें लैंग्वेज मॉडल किसी अनुरोध में पढ़ता है, जिनमें सिस्टम प्रॉम्प्ट, बातचीत का इतिहास, टूल डेफ़िनिशन और इमेज शामिल हैं। आउटपुट टोकन वे हैं जो वह जवाब में बनाता है, रीज़निंग समेत। LLM API आम तौर पर दोनों की कीमत अलग रखते हैं, और आउटपुट आम तौर पर महँगा होता है। - [INT4 (4-bit क्वांटाइज़ेशन)](https://broadrouter.com/hi/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 एक 4-bit integer नंबर फ़ॉर्मैट है जो 16 स्तर दर्शा सकता है। LLM इन्फ़रेंस में इसका मतलब आम तौर पर 4-bit weight-only क्वांटाइज़ेशन (W4A16) होता है: weights scale factors के साथ INT4 में रखे जाते हैं और गणना के दौरान 16-bit में फैलाए जाते हैं, जिससे 16-bit weights की लगभग एक चौथाई मेमोरी लगती है। - [IP allow list (IP अनुमति सूची)](https://broadrouter.com/hi/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): IP allow list एक सुरक्षा नियंत्रण है जो सिर्फ़ सूची में दिए IP पतों या CIDR रेंज से आए अनुरोध स्वीकार करता है और बाक़ी सबको ठुकरा देता है। किसी API पर यह लीक हुई key से होने वाला नुक़सान सीमित करता है, क्योंकि सूची से बाहर के नेटवर्क से key काम नहीं करती। - [जेलब्रेक (LLM jailbreak)](https://broadrouter.com/hi/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): जेलब्रेक एक ऐसा इनपुट है जो किसी लैंग्वेज मॉडल से उसकी सेफ़्टी ट्रेनिंग या उसे चलाने वाले ऐप के नियमों की अनदेखी करवाने के लिए बनाया जाता है, ताकि वह ऐसा कंटेंट दे या ऐसे काम करे जिनसे वह आम तौर पर इनकार करता। OWASP जेलब्रेकिंग को prompt injection का एक रूप मानता है। - [JSON मोड (JSON mode)](https://broadrouter.com/hi/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): JSON मोड, LLM API की एक सेटिंग है, OpenAI जैसे API में response_format {"type": "json_object"}, जो मॉडल के जवाब को वैलिड और पार्स होने लायक़ JSON बनाती है। JSON Schema वाले स्ट्रक्चर्ड आउटपुट के उलट, यह गारंटी नहीं देती कि ऑब्जेक्ट में कौन-सी keys, टाइप या वैल्यू होंगी। - [JSON Schema (JSON स्कीमा)](https://broadrouter.com/hi/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema, JSON में लिखी एक शब्दावली है, जो JSON डेटा की बनावट बताती और जाँचती है: टाइप, ज़रूरी फ़ील्ड, स्वीकार्य वैल्यू और नेस्टिंग। मौजूदा वर्ज़न 2020-12 है। LLM API इससे टूल पैरामीटर तय करते हैं, और वह सटीक आकार भी जिसका स्ट्रक्चर्ड आउटपुट को पालन करना है। - [KV cache (की-वैल्यू कैश)](https://broadrouter.com/hi/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): KV cache वह मेमोरी है जहाँ ट्रांसफ़ॉर्मर लैंग्वेज मॉडल पहले से प्रोसेस हुए हर टोकन के लिए निकाले गए attention keys और values रखता है, ताकि हर नया टोकन पूरी sequence दोबारा गिनने के बजाय उन्हें इस्तेमाल करे। यह context की लंबाई के साथ बढ़ता है और अक्सर तय करता है कि GPU कितने अनुरोध संभाले। - [लेटेंसी (latency)](https://broadrouter.com/hi/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): लैंग्वेज मॉडल API के लिए लेटेंसी वह समय है जो किसी अनुरोध के भेजे जाने से अंतिम आउटपुट टोकन मिलने तक लगता है। यह पहले टोकन तक के समय और जनरेशन टाइम का जोड़ है, और आम तौर पर औसत के बजाय p50 और p95 जैसे परसेंटाइल में बताई जाती है। - [LLM-as-a-judge (LLM जज)](https://broadrouter.com/hi/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge एक मूल्यांकन तरीक़ा है जिसमें एक लैंग्वेज मॉडल दूसरे मॉडल के आउटपुट को ग्रेड करता है, या तो एक जवाब को rubric पर स्कोर देकर या दो जवाबों में से बेहतर चुनकर। यह इंसानी फ़ैसले के क़रीब पहुँचता है और बहुत सस्ता है, लेकिन इसके जाने-माने bias हैं, जैसे पहले या लंबे जवाब को तरजीह देना। - [LLM बेंचमार्क (LLM benchmark)](https://broadrouter.com/hi/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): LLM बेंचमार्क टास्क का एक तय, आम तौर पर पब्लिक सेट है जिसके साथ स्कोर करने का तरीक़ा भी तय होता है, और जिससे लैंग्वेज मॉडल को किसी एक हुनर, जैसे ज्ञान के सवाल, कोडिंग या स्पीच रिकग्निशन, पर मापा और तुलना की जाती है। बेंचमार्क स्कोर उस टेस्ट पर तुलनात्मक क्षमता दिखाता है, आपके अपने वर्कलोड पर प्रदर्शन नहीं। - [LLM राउटर (LLM router)](https://broadrouter.com/hi/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): LLM राउटर एक सॉफ़्टवेयर या सर्विस है जो ऐप और एक या ज़्यादा लैंग्वेज मॉडल API के बीच बैठती है और हर अनुरोध के लिए तय करती है कि उसे कौन-सा मॉडल या प्रोवाइडर संभालेगा, लागत, latency, उपलब्धता या टास्क के प्रकार जैसे नियमों के आधार पर। - [लॉगप्रॉब्स (logprobs)](https://broadrouter.com/hi/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): लॉगप्रॉब्स वे लॉग प्रायिकताएँ हैं जो लैंग्वेज मॉडल ने अपने जनरेट किए टोकन को दीं। इन्हें लौटाने वाला API हर आउटपुट टोकन के लिए दिखाता है कि मॉडल ने उसे कितना संभावित माना, और चाहें तो उस जगह के सबसे संभावित वैकल्पिक टोकन भी। - [LoRA (low-rank adaptation)](https://broadrouter.com/hi/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation) एक पैरामीटर-कुशल fine-tuning तरीक़ा है जो पहले से ट्रेन हुए मॉडल के weights को जमा (freeze) देता है और चुनी हुई layers के लिए छोटी low-rank मैट्रिक्स की एक जोड़ी ट्रेन करता है। सीखा गया बदलाव छोटे adapter के रूप में रखा जा सकता है या weights में मिलाया जा सकता है, बिना इन्फ़रेंस latency बढ़ाए। - [Lost in the middle (बीच में खो जाना)](https://broadrouter.com/hi/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle लैंग्वेज मॉडल का देखा गया रुझान है कि वे काम की जानकारी का सबसे अच्छा इस्तेमाल तब करते हैं जब वह लंबे इनपुट की शुरुआत या अंत के पास हो, और सबसे ख़राब तब जब वह बीच में हो। यह नाम Liu और साथियों की 2023 की स्टडी ने दिया। - [max_tokens (आउटपुट टोकन सीमा)](https://broadrouter.com/hi/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens एक LLM API पैरामीटर है जो तय करता है कि मॉडल एक जवाब में ज़्यादा से ज़्यादा कितने टोकन जनरेट कर सकता है। सीमा पर पहुँचते ही जनरेशन रुक जाता है और जवाब को कटा हुआ चिह्नित किया जाता है, जैसे finish_reason को "length" पर सेट करके। - [MCP (Model Context Protocol)](https://broadrouter.com/hi/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol) JSON-RPC 2.0 पर बना एक ओपन प्रोटोकॉल है, जो तय करता है कि LLM ऐप्लिकेशन बाहरी टूल और डेटा से कैसे जुड़ें। MCP सर्वर टूल, रिसोर्स और प्रॉम्प्ट उपलब्ध कराता है; होस्ट ऐप्लिकेशन के अंदर का MCP क्लाइंट उन्हें ढूँढता है और मॉडल की टूल कॉल सर्वर तक पहुँचाता है। - [Mixture of experts (MoE)](https://broadrouter.com/hi/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Mixture of experts (MoE) एक न्यूरल नेटवर्क डिज़ाइन है जिसमें एक layer में कई expert सब-नेटवर्क होते हैं और एक सीखा हुआ router हर टोकन को उनमें से सिर्फ़ कुछ के पास भेजता है। MoE लैंग्वेज मॉडल के कुल पैरामीटर बहुत होते हैं, लेकिन हर टोकन पर वह एक छोटा active हिस्सा ही इस्तेमाल करता है। - [मॉडल डिप्रिकेशन (model deprecation)](https://broadrouter.com/hi/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): मॉडल डिप्रिकेशन तब होता है जब कोई API प्रोवाइडर घोषणा करता है कि एक मॉडल मिलना बंद हो जाएगा, आम तौर पर रिटायरमेंट या शटडाउन की तारीख़ और सुझाए गए विकल्प के साथ। उस तारीख़ के बाद उस मॉडल के अनुरोध फ़ेल होते हैं, इसलिए ऐप को पहले ही किसी दूसरी मॉडल ID पर जाना होता है। - [मॉडल फ़ॉलबैक (model fallback)](https://broadrouter.com/hi/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): मॉडल फ़ॉलबैक भरोसेमंदी का एक पैटर्न है जिसमें फ़ेल हुआ LLM अनुरोध एक क्रमबद्ध सूची के किसी दूसरे मॉडल या प्रोवाइडर पर दोबारा भेजा जाता है। ट्रिगर कोई एरर, रेट लिमिट, context-length एरर या content-policy इनकार हो सकता है। यह आपके क्लाइंट में या API के आगे लगे गेटवे में चलता है। - [मॉडल लाइसेंस (model license)](https://broadrouter.com/hi/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): मॉडल लाइसेंस वे क़ानूनी शर्तें हैं जिनके तहत किसी मॉडल के weights, कोड और आउटपुट इस्तेमाल, बदले और साझा किए जा सकते हैं। ओपन-वेट मॉडल Apache 2.0 या MIT जैसे आम सॉफ़्टवेयर लाइसेंस, या कस्टम मॉडल लाइसेंस इस्तेमाल करते हैं, जो बड़ी कमर्शियल सेवाओं पर सीमा जैसी शर्तें जोड़ सकते हैं। - [मानक आधुनिक अरबी (Modern Standard Arabic, MSA)](https://broadrouter.com/hi/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): Modern Standard Arabic (MSA) अरबी का मानक रूप है, जो पूरी अरब दुनिया में लेखन, समाचार, शिक्षा और औपचारिक भाषण में इस्तेमाल होता है। यह किसी की रोज़मर्रा की मातृभाषा नहीं है: लोग मिस्री या खाड़ी अरबी जैसी क्षेत्रीय बोलियाँ बोलते हैं और औपचारिक मौक़ों पर MSA अपनाते हैं। - [मल्टी-एजेंट सिस्टम (multi-agent system)](https://broadrouter.com/hi/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): मल्टी-एजेंट सिस्टम एक LLM ऐप्लिकेशन है जिसमें कई एजेंट, हर एक अपने प्रॉम्प्ट, टूल और कॉन्टेक्स्ट के साथ, एक ही काम के अलग-अलग हिस्सों पर काम करते हैं। आम डिज़ाइन orchestrator-worker है: एक lead एजेंट योजना बनाता है, worker एजेंट समानांतर शुरू करता है और उनके नतीजे जोड़ता है। - [मल्टीमोडल AI (multimodal AI)](https://broadrouter.com/hi/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): मल्टीमोडल AI ऐसा मॉडल या सिस्टम है जो एक से ज़्यादा तरह के डेटा, जैसे टेक्स्ट, इमेज, वीडियो और ऑडियो, के साथ काम करता है। कोई मल्टीमोडल मॉडल फ़ोटो और सवाल साथ पढ़ सकता है, या आवाज़ लेकर टेक्स्ट में जवाब दे सकता है। vision-language मॉडल इसका सबसे आम रूप है। - [नेम्ड एंटिटी रिकग्निशन (NER)](https://broadrouter.com/hi/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): Named entity recognition (NER) टेक्स्ट में नाम ढूँढने और हर नाम को व्यक्ति, संगठन या जगह जैसे किसी प्रकार का लेबल देने का काम है। पुराने सिस्टम हर टोकन को टैग करते हैं; LLM से अक्सर entities की सूची JSON में माँगी जाती है, जिसे फिर स्रोत टेक्स्ट से मिलाकर जाँचा जाता है। - [OCR (optical character recognition)](https://broadrouter.com/hi/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): OCR (optical character recognition) वह तकनीक है जो किसी इमेज, स्कैन या PDF में छपे या हाथ से लिखे टेक्स्ट को पहचानकर मशीन से पढ़े जा सकने वाले अक्षरों में बदलती है। क्लासिक OCR टेक्स्ट और पेज पर उसकी जगह देता है; vision-language मॉडल टेक्स्ट पढ़कर उसके बारे में सीधे सवालों के जवाब दे सकते हैं। - [Ogg Opus](https://broadrouter.com/hi/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus वह ऑडियो है जो Opus codec (RFC 6716) से एन्कोड होकर Ogg कंटेनर (RFC 7845) में रखा जाता है। यह कम bitrate पर बोली और इंटरैक्टिव ऑडियो के लिए बना है, और WhatsApp वॉइस नोट के लिए यही फ़ॉर्मैट इस्तेमाल करता है, आम तौर पर .ogg या .opus फ़ाइलों के रूप में। - [ओपन-वेट मॉडल (open-weight model)](https://broadrouter.com/hi/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): ओपन-वेट मॉडल ऐसा मशीन लर्निंग मॉडल है जिसके ट्रेन किए weights सबके डाउनलोड करने और चलाने के लिए प्रकाशित होते हैं, आम तौर पर शर्तों वाले लाइसेंस के साथ। यह पूरी तरह ओपन-सोर्स AI से अलग है, जो ट्रेनिंग कोड और ट्रेनिंग डेटा की विस्तृत जानकारी भी जारी करता है, ताकि दूसरे मॉडल को समझ और दोबारा बना सकें। - [OpenAI-compatible API](https://broadrouter.com/hi/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): OpenAI-compatible API ऐसा HTTP API है जो OpenAI API वाले ही अनुरोध और रिस्पॉन्स फ़ॉर्मैट स्वीकार करता है, जैसे POST /v1/chat/completions, ताकि OpenAI के SDK और टूल सिर्फ़ base URL, API key और मॉडल का नाम बदलकर इसे इस्तेमाल कर सकें। - [पैरेलल टूल कॉल (parallel tool calls)](https://broadrouter.com/hi/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Parallel tool calls तब होते हैं जब मॉडल एक ही जवाब में एक से ज़्यादा टूल कॉल लौटाता है, जैसे एक साथ तीन फ़ाइलें पढ़ना। OpenAI फ़ॉर्मैट में ये tool_calls की कई एंट्री के रूप में आते हैं, और parallel_tool_calls को false करने पर मॉडल हर टर्न में शून्य या एक कॉल तक सीमित रहता है। - [उपयोग के अनुसार भुगतान (pay-as-you-go)](https://broadrouter.com/hi/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): उपयोग के अनुसार भुगतान (pay-as-you-go) एक प्राइसिंग मॉडल है जिसमें आप सिर्फ़ उतना देते हैं जितना इस्तेमाल करते हैं, जिसे टोकन, अनुरोध या ऑडियो मिनट जैसी इकाइयों में मापा जाता है, बिना सब्सक्रिप्शन या न्यूनतम प्रतिबद्धता के। ज़्यादातर LLM API इसी तरह बिल बनाते हैं, प्रीपेड बैलेंस या समय-समय पर बनने वाले इनवॉइस से। - [PDPL (सऊदी और UAE के डेटा संरक्षण क़ानून)](https://broadrouter.com/hi/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) सऊदी अरब के पर्सनल डेटा क़ानून का आम नाम है, जो Royal Decree M/19 से जारी हुआ और SDAIA लागू करता है, और UAE के Federal Decree-Law No. 45 of 2021 का भी। दोनों पर्सनल डेटा की प्रोसेसिंग नियंत्रित करते और विदेश ट्रांसफ़र की शर्तें तय करते हैं, विदेशी LLM API को भेजे प्रॉम्प्ट समेत। - [PII (personally identifiable information)](https://broadrouter.com/hi/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): PII (personally identifiable information) वह जानकारी है जिससे किसी व्यक्ति की पहचान अलग की या खोजी जा सके, जैसे नाम, ID नंबर या बायोमेट्रिक रिकॉर्ड, साथ ही उस व्यक्ति से जुड़ी या जोड़ी जा सकने वाली दूसरी जानकारी। GDPR का व्यापक शब्द personal data है। LLM में PII redaction इसे सिस्टम से निकलने से पहले प्रॉम्प्ट से हटाता है। - [Prefill और decode](https://broadrouter.com/hi/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill और decode LLM इन्फ़रेंस के दो चरण हैं। Prefill में मॉडल पूरा प्रॉम्प्ट समानांतर रूप से प्रोसेस करके अपना KV cache भरता है; decode में वह उसी cache को दोबारा इस्तेमाल करते हुए जवाब एक-एक टोकन करके बनाता है। Prefill मुख्य रूप से पहले टोकन तक का समय तय करता है; decode आउटपुट की रफ़्तार। - [प्रीपेड क्रेडिट (prepaid credit)](https://broadrouter.com/hi/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): प्रीपेड क्रेडिट वह पैसा है जो इस्तेमाल से पहले API अकाउंट में डाला जाता है। हर अनुरोध की लागत बैलेंस से कटती है, और बैलेंस ख़त्म होने पर मीटर वाले अनुरोध तब तक रुक जाते हैं जब तक आप टॉप-अप न करें। इससे खर्च उतने तक सीमित रहता है जितना आपने जमा किया है। - [प्रॉम्प्ट इंजीनियरिंग (prompt engineering)](https://broadrouter.com/hi/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): प्रॉम्प्ट इंजीनियरिंग, लैंग्वेज मॉडल को दिए जाने वाले निर्देशों को लिखने, ढाँचा देने और टेस्ट करने की प्रक्रिया है, ताकि मॉडल लगातार आपकी ज़रूरतों के मुताबिक़ आउटपुट दे। इसमें रोल मैसेज, साफ़ निर्देश, उदाहरण, फ़ॉर्मैटिंग और टेस्ट केस पर नतीजों को मापना शामिल है। - [प्रॉम्प्ट इंजेक्शन (prompt injection)](https://broadrouter.com/hi/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): Prompt injection एक हमला है जिसमें लैंग्वेज मॉडल को दिया गया टेक्स्ट, चाहे यूज़र ने लिखा हो या किसी वेब पेज, ईमेल, फ़ाइल या टूल नतीजे में छिपा हो, मॉडल का बर्ताव डेवलपर की मंशा के ख़िलाफ़ बदल देता है। OWASP अपनी LLM ऐप्लिकेशन Top 10 सूची में इसे पहले स्थान (LLM01:2025) पर रखता है। - [क्वांटाइज़ेशन (quantization)](https://broadrouter.com/hi/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): क्वांटाइज़ेशन, न्यूरल नेटवर्क की संख्याओं, मुख्य रूप से उसके weights, को 16-bit के बजाय 8-bit या 4-bit जैसे कम प्रिसिज़न वाले फ़ॉर्मैट में रखने की प्रक्रिया है, ताकि मॉडल कम मेमोरी ले और अक्सर तेज़ चले। अच्छे तरीक़े सटीकता में होने वाले नुक़सान को कम रखते हैं। - [RAG (retrieval-augmented generation)](https://broadrouter.com/hi/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG (retrieval-augmented generation) एक पैटर्न है जिसमें ऐप्लिकेशन पहले किसी बाहरी स्टोर, जैसे सर्च या वेक्टर इंडेक्स, से सवाल से जुड़े हिस्से निकालता है, और फिर उन्हें प्रॉम्प्ट में जोड़ता है, ताकि लैंग्वेज मॉडल सिर्फ़ अपनी याददाश्त के बजाय उस टेक्स्ट से जवाब दे सके। - [रेट लिमिट (rate limit)](https://broadrouter.com/hi/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): रेट लिमिट वह सीमा है जो कोई API तय करता है कि एक क्लाइंट किसी समय-अवधि में उसका कितना इस्तेमाल कर सकता है, जैसे प्रति मिनट अनुरोध या प्रति मिनट टोकन। सीमा से ऊपर के अनुरोध ठुकरा दिए जाते हैं, आम तौर पर HTTP 429 Too Many Requests के साथ, जब तक इस्तेमाल फिर सीमा के नीचे न आ जाए। - [ReAct (reason and act)](https://broadrouter.com/hi/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (reason and act) एक प्रॉम्प्टिंग पैटर्न है, जिसे Yao और साथियों ने 2022 में पेश किया, जिसमें लैंग्वेज मॉडल reasoning traces को सर्च या टूल कॉल जैसी कार्रवाइयों के साथ बारी-बारी चलाता है। वह हर observation पढ़ता है, अपनी योजना अपडेट करता है, और अंतिम जवाब तक यही दोहराता है। - [रीज़निंग टोकन (reasoning tokens)](https://broadrouter.com/hi/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): रीज़निंग टोकन वे टोकन हैं जो रीज़निंग मॉडल अपना दिखने वाला जवाब लिखने से पहले किसी समस्या को हल करने के लिए बनाता है। इनका बिल आम तौर पर आउटपुट टोकन की तरह बनता है, ये आउटपुट लिमिट और कॉन्टेक्स्ट विंडो में गिने जाते हैं, और API इन्हें छिपा सकता है या इनका सारांश दे सकता है। - [रिक्वेस्ट टाइमआउट (request timeout)](https://broadrouter.com/hi/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): रिक्वेस्ट टाइमआउट वह अधिकतम समय है जितना कोई HTTP क्लाइंट सर्वर का इंतज़ार करता है, उसके बाद अनुरोध छोड़ देता है। क्लाइंट आम तौर पर इसे कनेक्शन खोलने के connect टाइमआउट और डेटा के इंतज़ार वाले read टाइमआउट में बाँटते हैं। LLM कॉल को खुले read टाइमआउट चाहिए, क्योंकि लंबे जवाबों में समय लगता है। - [रीरैंकिंग (reranking)](https://broadrouter.com/hi/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Reranking retrieval का दूसरा चरण है जिसमें एक ज़्यादा सटीक पर धीमा मॉडल, आम तौर पर cross-encoder, पहली सर्च से आए हर query और दस्तावेज़ जोड़े को स्कोर करके उनका क्रम बदलता है, ताकि सिर्फ़ सबसे प्रासंगिक कुछ ही लैंग्वेज मॉडल तक पहुँचें। - [राइट-टू-लेफ़्ट टेक्स्ट (right-to-left, RTL text)](https://broadrouter.com/hi/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): राइट-टू-लेफ़्ट (RTL) टेक्स्ट वह लिखावट है, जैसे अरबी या हिब्रू, जो दाएँ से बाएँ दिखाई जाती है। कंप्यूटर इसे लॉजिकल पढ़ने के क्रम में सेव करते हैं, और जब RTL टेक्स्ट लेफ़्ट-टू-राइट टेक्स्ट, अंकों, मॉडल नामों या कोड के साथ मिलता है, तब Unicode Bidirectional Algorithm दिखने का क्रम तय करता है। - [सैंपल रेट (sample rate)](https://broadrouter.com/hi/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): सैंपल रेट हर सेकंड रिकॉर्ड होने वाले ऑडियो सैंपल की संख्या है, जिसे हर्ट्ज़ (Hz) में नापा जाता है। स्पीच रिकग्निशन मॉडल आम तौर पर 16,000 Hz (16 kHz) मोनो ऑडियो की उम्मीद करते हैं। बिट डेप्थ और चैनल संख्या के साथ मिलकर सैंपल रेट तय करता है कि बिना कंप्रेशन वाली ऑडियो फ़ाइल प्रति मिनट कितनी बड़ी होगी। - [seed पैरामीटर (seed parameter)](https://broadrouter.com/hi/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): seed पैरामीटर एक पूर्णांक है जो आप अनुरोध के साथ भेजते हैं, ताकि वही अनुरोध दोहराने पर API वही टोकन सैंपल करने की कोशिश करे। OpenAI इसे best effort बताता है: determinism की गारंटी नहीं है, और system_fingerprint फ़ील्ड दिखाता है कि बैकएंड कॉन्फ़िगरेशन कब बदला। - [सर्वरलेस इन्फ़रेंस (serverless inference)](https://broadrouter.com/hi/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): सर्वरलेस इन्फ़रेंस मशीन लर्निंग मॉडल चलाने का ऐसा तरीका है जिसमें प्लेटफ़ॉर्म एंडपॉइंट के पीछे का कंप्यूट अपने आप शुरू, स्केल और बंद करता है, अक्सर खाली रहने पर ज़ीरो तक, और रिज़र्व मशीनों के बजाय इस्तेमाल का बिल लेता है। इसकी क़ीमत संभावित कोल्ड-स्टार्ट देरी है। - [SLA (सर्विस लेवल एग्रीमेंट)](https://broadrouter.com/hi/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): SLA (सर्विस लेवल एग्रीमेंट) प्रोवाइडर की ओर से सर्विस के नापे जा सकने वाले स्तर का अनुबंधित वादा है, आम तौर पर मासिक अपटाइम प्रतिशत, और उसके पूरा न होने पर तय उपाय, आम तौर पर सर्विस क्रेडिट। SLA बताता है कि डाउनटाइम किसे माना जाएगा, क्या बाहर है और दावा कैसे करें। - [स्पेक्युलेटिव डिकोडिंग (speculative decoding)](https://broadrouter.com/hi/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): स्पेक्युलेटिव डिकोडिंग एक इन्फ़रेंस तकनीक है जिसमें छोटा, तेज़ ड्राफ़्ट मॉडल अगले कई टोकन सुझाता है और बड़ा टारगेट मॉडल उन्हें एक ही forward pass में जाँचकर स्वीकार किए गए टोकन रखता है। यह आउटपुट के डिस्ट्रिब्यूशन को बदले बिना जनरेशन तेज़ करती है। - [स्पीच-टू-टेक्स्ट (speech-to-text)](https://broadrouter.com/hi/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): स्पीच-टू-टेक्स्ट, जिसे automatic speech recognition (ASR) भी कहते हैं, बोले गए ऑडियो को लिखित टेक्स्ट में बदलने की तकनीक है। speech-to-text API ऑडियो फ़ाइल या स्ट्रीम लेकर ट्रांसक्रिप्ट लौटाता है, आम तौर पर प्रति ऑडियो मिनट की कीमत पर। सटीकता वर्ड एरर रेट से मापी जाती है और लहजे, बोली और रिकॉर्डिंग क्वालिटी के साथ बदलती है। - [स्पेंड लिमिट (spend limit)](https://broadrouter.com/hi/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): स्पेंड लिमिट वह सीमा है कि कोई API अकाउंट, प्रोजेक्ट या key एक अवधि में, आम तौर पर एक महीने में, कितना ख़र्च कर सकती है। इस्तेमाल वहाँ पहुँचने पर प्रोवाइडर आगे के अनुरोध रिजेक्ट करता है, अक्सर HTTP 429 या 400 एरर के साथ, जब तक अवधि रीसेट न हो या लिमिट बढ़ाई न जाए। - [SRT सबटाइटल (SRT subtitles)](https://broadrouter.com/hi/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) एक प्लेन-टेक्स्ट सबटाइटल फ़ॉर्मैट है। हर कैप्शन में एक क्रम संख्या, hh:mm:ss,mmm --> hh:mm:ss,mmm रूप में लिखा शुरू और अंत का समय, टेक्स्ट की एक या ज़्यादा लाइनें और एक ख़ाली लाइन होती है। WebVTT इससे मिलता-जुलता वेब फ़ॉर्मैट है, जिसे HTML वीडियो इस्तेमाल करता है। - [Server-sent events (SSE)](https://broadrouter.com/hi/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Server-sent events (SSE) एक वेब स्टैंडर्ड है, जिसमें सर्वर text/event-stream कंटेंट टाइप वाले एक लंबे खुले HTTP जवाब पर क्लाइंट को टेक्स्ट events की स्ट्रीम भेजता है। हर event में data: जैसी एक या ज़्यादा फ़ील्ड लाइनें होती हैं, और वह एक ख़ाली लाइन पर ख़त्म होता है। - [स्टॉप सीक्वेंस (stop sequences)](https://broadrouter.com/hi/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): स्टॉप सीक्वेंस वे स्ट्रिंग हैं जो आप अनुरोध के साथ भेजते हैं, ताकि मॉडल इनमें से कोई भी लिखते ही जनरेशन रोक दे। OpenAI इस फ़ील्ड को stop कहता है और 4 सीक्वेंस तक लेता है; Anthropic इसे stop_sequences कहता है। मिला हुआ टेक्स्ट आम तौर पर लौटाए गए आउटपुट में नहीं होता। - [स्ट्रीमिंग (streaming)](https://broadrouter.com/hi/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): स्ट्रीमिंग, लैंग्वेज मॉडल के जवाब को अंत में एक साथ लौटाने के बजाय, बनते समय ही छोटे-छोटे टुकड़ों (chunks) में लौटाने का तरीक़ा है। OpenAI-कम्पैटिबल API, अनुरोध में stream को true सेट करने पर server-sent events से स्ट्रीम करते हैं। - [स्ट्रक्चर्ड आउटपुट (structured output)](https://broadrouter.com/hi/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): स्ट्रक्चर्ड आउटपुट LLM API का एक फ़ीचर है जो मॉडल के जवाब को डेवलपर के तय फ़ॉर्मैट, आम तौर पर JSON Schema, तक सीमित करता है, ताकि कोड उसे भरोसे से पार्स कर सके। यह JSON mode से सख़्त है, जो मान्य JSON की गारंटी देता है, लेकिन किसी ख़ास ढाँचे की नहीं। - [सबएजेंट (subagent)](https://broadrouter.com/hi/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): सबएजेंट एक अलग एजेंट है जिसे मुख्य एजेंट किसी एक तय काम के लिए शुरू करता है। यह अपनी कॉन्टेक्स्ट विंडो, सिस्टम प्रॉम्प्ट और टूल एक्सेस के साथ चलता है, काम करता है, और मुख्य एजेंट को सिर्फ़ नतीजा या सारांश लौटाता है, जिससे मुख्य बातचीत छोटी रहती है। - [सबप्रोसेसर (subprocessor)](https://broadrouter.com/hi/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): सबप्रोसेसर वह तीसरा पक्ष है जिसे प्रोसेसर कंट्रोलर की ओर से पर्सनल डेटा प्रोसेस करने के लिए लगाता है। LLM API के मामले में सबप्रोसेसर में GPU होस्ट, पार्टनर इन्फ़रेंस प्रोवाइडर और बैकअप क्षमता शामिल हो सकते हैं। GDPR अनुच्छेद 28 के तहत इन्हें कंट्रोलर की अनुमति और वही डेटा सुरक्षा दायित्व चाहिए। - [सिस्टम प्रॉम्प्ट (system prompt)](https://broadrouter.com/hi/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): सिस्टम प्रॉम्प्ट वह निर्देश टेक्स्ट है जो कोई ऐप्लिकेशन बातचीत से पहले लैंग्वेज मॉडल को भेजता है, आम तौर पर system या developer रोल वाले मैसेज के रूप में। यह मॉडल की भूमिका, नियम, लहजा और आउटपुट फ़ॉर्मैट तय करता है, और टकराव होने पर मॉडल को यूज़र के मैसेज के बजाय इसी का पालन करना होता है। - [ता मरबूता (taa marbuta, ة)](https://broadrouter.com/hi/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): ता मरबूता (ة, U+0629) एक अरबी अक्षर है जो सिर्फ़ शब्द के अंत में आता है, अक्सर स्त्रीलिंग संज्ञा और विशेषण का संकेत देता है। यह दो बिंदुओं वाले हे (ه, U+0647) जैसा दिखता है, इसलिए टाइप किए टेक्स्ट में दोनों मिल जाते हैं, और कई सर्च नॉर्मलाइज़र ة को ه में बदल देते हैं। - [टेम्परेचर (temperature)](https://broadrouter.com/hi/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): टेम्परेचर एक सैंपलिंग पैरामीटर है जो तय करता है कि लैंग्वेज मॉडल का आउटपुट कितना रैंडम होगा। मॉडल अपने टोकन स्कोर को प्रायिकता में बदलने से पहले टेम्परेचर से भाग देता है: कम वैल्यू पर सबसे संभावित टोकन हावी रहता है, ज़्यादा वैल्यू पर वितरण सपाट हो जाता है और कम संभावित टोकन ज़्यादा बार चुने जाते हैं। - [टेक्स्ट-टू-स्पीच (text-to-speech, TTS)](https://broadrouter.com/hi/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): टेक्स्ट-टू-स्पीच (TTS), या स्पीच सिंथेसिस, लिखे टेक्स्ट को बोले गए ऑडियो में बदलता है। TTS API टेक्स्ट, एक आवाज़ और MP3 या WAV जैसा आउटपुट फ़ॉर्मैट लेता है और जनरेट की गई आवाज़ लौटाता है। यह स्पीच-टू-टेक्स्ट का उलटा है, जो ऑडियो को टेक्स्ट में बदलता है। - [थ्रूपुट (throughput)](https://broadrouter.com/hi/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): LLM सर्विंग में थ्रूपुट वह कुल काम है जो कोई सिस्टम सभी कॉन्करेंट अनुरोधों में प्रति इकाई समय पूरा करता है, आम तौर पर आउटपुट टोकन प्रति सेकंड या अनुरोध प्रति सेकंड में नापा जाता है। यह प्रति-स्ट्रीम स्पीड से अलग है, जो बताती है कि एक अनुरोध को उसके टोकन कितनी तेज़ी से मिलते हैं। - [पहले टोकन तक का समय (TTFT, time to first token)](https://broadrouter.com/hi/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): Time to first token (TTFT) वह समय है जो लैंग्वेज मॉडल को अनुरोध भेजने से पहला आउटपुट टोकन आने तक लगता है। इसमें नेटवर्क का समय, सर्वर पर कतार और prefill शामिल हैं, यानी वह चरण जिसमें मॉडल पूरा प्रॉम्प्ट पढ़ता है, इसलिए यह प्रॉम्प्ट की लंबाई और लोड के साथ बढ़ता है। - [टाइमस्टैम्प (स्पीच-टू-टेक्स्ट, timestamps)](https://broadrouter.com/hi/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): स्पीच-टू-टेक्स्ट में टाइमस्टैम्प ट्रांसक्रिप्ट के हिस्सों से जुड़े शुरू और अंत के समय हैं, सेकंड में। सेगमेंट टाइमस्टैम्प किसी वाक्यांश या वाक्य को कवर करते हैं; वर्ड टाइमस्टैम्प हर शब्द को। सबटाइटल, रिकॉर्डिंग के अंदर सर्च और टेक्स्ट को स्पीकर से मिलाने के लिए ये ज़रूरी हैं। - [टोकन (token)](https://broadrouter.com/hi/glossary/token/) (translation of https://broadrouter.com/glossary/token/): टोकन टेक्स्ट की वह इकाई है जिसे लैंग्वेज मॉडल पढ़ता और लिखता है: कोई शब्द, शब्द का हिस्सा, विराम चिह्न या बाइट का क्रम, जो मॉडल का टोकनाइज़र बनाता है। LLM API इस्तेमाल को टोकन में गिनते हैं, उसी में लिमिट लगाते हैं और आम तौर पर बिल भी बनाते हैं। - [टोकनाइज़र (tokenizer)](https://broadrouter.com/hi/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): टोकनाइज़र वह हिस्सा है जो टेक्स्ट को इंटीजर टोकन ID के उस क्रम में बदलता है जिसे लैंग्वेज मॉडल पढ़ता है, और जनरेट हुई ID को वापस टेक्स्ट में बदलता है। हर मॉडल अपने टोकनाइज़र के साथ आता है, इसलिए एक ही टेक्स्ट अलग-अलग मॉडलों पर अलग संख्या में टोकन बना सकता है। - [टोकन प्रति सेकंड (tokens per second, TPS)](https://broadrouter.com/hi/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): टोकन प्रति सेकंड (TPS) वह दर है जिस पर लैंग्वेज मॉडल आउटपुट टोकन बनाता है, आम तौर पर एक अनुरोध का पहला टोकन आने के बाद नापी जाती है। यह तय करती है कि शुरू होकर जवाब कितना तेज़ स्ट्रीम होगा। सभी कॉन्करेंट अनुरोधों का जोड़ सिस्टम थ्रूपुट है; पहले टोकन से पहले का इंतज़ार अलग नापा जाता है। - [टूल कॉलिंग (tool calling)](https://broadrouter.com/hi/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): टूल कॉलिंग, जिसे function calling भी कहते हैं, LLM API की वह सुविधा है जिसमें मॉडल डेवलपर के तय किए किसी फ़ंक्शन को चलाने का एक स्ट्रक्चर्ड अनुरोध, JSON आर्ग्युमेंट के साथ, लौटाता है। ऐप्लिकेशन फ़ंक्शन चलाकर नतीजा वापस भेजता है; मॉडल ख़ुद कभी कोड नहीं चलाता। - [tool_choice](https://broadrouter.com/hi/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice एक अनुरोध पैरामीटर है जो टूल के इस्तेमाल को नियंत्रित करता है। OpenAI के Chat Completions में यह auto (मॉडल तय करे), required (एक या ज़्यादा टूल कॉल करने ही हैं), none (कोई टूल नहीं) या ज़बरदस्ती कॉल के लिए किसी फ़ंक्शन का नाम हो सकता है। Anthropic के बराबर विकल्प auto, any, tool और none हैं। - [टॉप-पी (top-p, nucleus sampling)](https://broadrouter.com/hi/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): टॉप-पी, जिसे nucleus sampling भी कहते हैं, डिकोडिंग का एक तरीक़ा है जिसमें मॉडल अगला टोकन सिर्फ़ सबसे संभावित टोकन के उस सबसे छोटे सेट से चुनता है जिनकी कुल प्रायिकता 0.9 जैसी सीमा p तक पहुँचती है, और बाक़ी कम संभावित टोकन छोड़ देता है। - [Unicode नॉर्मलाइज़ेशन (Unicode normalization)](https://broadrouter.com/hi/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Unicode नॉर्मलाइज़ेशन टेक्स्ट को एक मानक रूप में बदलता है, ताकि एक जैसी दिखने वाली स्ट्रिंग एक जैसे कोड पॉइंट इस्तेमाल करें। चार रूप हैं NFD, NFC, NFKD और NFKC; K वाले रूप compatibility कैरेक्टर, जैसे अरबी presentation forms और ligatures, को भी सामान्य अक्षरों में बदल देते हैं। - [वॉइस एक्टिविटी डिटेक्शन (VAD)](https://broadrouter.com/hi/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): वॉइस एक्टिविटी डिटेक्शन (VAD) सिग्नल-प्रोसेसिंग का एक क़दम है जो ऑडियो के छोटे फ़्रेम को स्पीच या नॉन-स्पीच में बाँटता है। स्पीच सिस्टम इससे ख़ामोशी छोड़ते हैं, लंबी रिकॉर्डिंग को ठहराव पर काटते हैं, पहचानते हैं कि बोलने वाला कब रुका, और बैकग्राउंड शोर को ट्रांसक्राइब होने से बचाते हैं। - [वेक्टर डेटाबेस (vector database)](https://broadrouter.com/hi/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): वेक्टर डेटाबेस एक डेटा स्टोर है जो embedding वेक्टर, अक्सर उनके सोर्स टेक्स्ट और मेटाडेटा के साथ, रखता है और nearest-neighbour क्वेरी का जवाब देता है: क्वेरी वेक्टर मिलने पर यह सबसे मिलते-जुलते सेव वेक्टर लौटाता है, आम तौर पर cosine दूरी, Euclidean दूरी या inner product से। - [विज़न-लैंग्वेज मॉडल (VLM)](https://broadrouter.com/hi/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): विज़न-लैंग्वेज मॉडल (VLM) ऐसा मॉडल है जो इमेज और टेक्स्ट को साथ में इनपुट के रूप में लेकर टेक्स्ट बनाता है, ताकि वह तस्वीरों का वर्णन कर सके, उनके बारे में सवालों के जवाब दे सके और दस्तावेज़, चार्ट या स्क्रीनशॉट पढ़ सके। VLM को image-text-to-text या multimodal मॉडल भी कहते हैं। - [vLLM](https://broadrouter.com/hi/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM, LLM inference और सर्विंग के लिए एक ओपन-सोर्स लाइब्रेरी है, जिसे पहले UC Berkeley की Sky Computing Lab में बनाया गया। यह PagedAttention से KV cache मेमोरी सँभालती है, आने वाले अनुरोधों को लगातार बैच करती है, कई quantization फ़ॉर्मैट सपोर्ट करती है और OpenAI-compatible HTTP सर्वर देती है। - [वोकैबुलरी साइज़ (vocabulary size)](https://broadrouter.com/hi/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): वोकैबुलरी साइज़ किसी लैंग्वेज मॉडल के टोकनाइज़र में अलग-अलग टोकन की संख्या है, जिनमें हर एक का अपना ID और embedding होता है। बड़ी वोकैबुलरी आम टेक्स्ट को कम टोकन में लिख सकती है, पर इससे embedding और आउटपुट लेयर बड़ी होती हैं, और यह सिर्फ़ उसी टेक्स्ट पर मदद करती है जिस पर वोकैबुलरी ट्रेन हुई। - [WAV फ़ाइल (WAV)](https://broadrouter.com/hi/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format, या WAVE) Microsoft और IBM के RIFF फ़ॉर्मैट पर बना एक ऑडियो कंटेनर है। इसमें आम तौर पर बिना कंप्रेशन वाले linear PCM सैंपल होते हैं, इसलिए इसका साइज़ पहले से पता होता है: प्रति सेकंड सैंपल रेट गुणा प्रति सैंपल बाइट गुणा चैनल। - [वर्ड एरर रेट (WER)](https://broadrouter.com/hi/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): वर्ड एरर रेट (WER) स्पीच रिकग्निशन की सटीकता का मानक पैमाना है: ट्रांसक्रिप्ट को सही रेफ़रंस में बदलने के लिए ज़रूरी शब्दों के substitutions, deletions और insertions की संख्या, जिसे रेफ़रंस के कुल शब्दों से भाग दिया जाता है। जितना कम, उतना बेहतर। - [ज़ीरो डेटा रिटेंशन (zero data retention, ZDR)](https://broadrouter.com/hi/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): ज़ीरो डेटा रिटेंशन (ZDR) डेटा सँभालने की एक पॉलिसी है, जिसमें AI प्रोवाइडर प्रॉम्प्ट और आउटपुट सिर्फ़ अनुरोध पूरा करने के लिए प्रोसेस करता है और बाद में वह सामग्री लॉग, कैश या ट्रेनिंग डेटा में स्टोर नहीं करता। टोकन गिनती और समय जैसे बिना सामग्री वाले रिकॉर्ड आम तौर पर फिर भी रखे जाते हैं। ### Indonesian (id, Bahasa Indonesia): 294 translated pages - [Articles in Indonesian](https://broadrouter.com/id/blog/) - [Perubahan harga API LLM Oktober 2026: 3 kenaikan, tanpa penurunan](https://broadrouter.com/id/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 perubahan halaman harga di 5 dari 7 penyedia API LLM pada Oktober 2026: 3 kenaikan, tanpa penurunan, 9 model baru diberi harga, 11 dihapus. Dengan tautan arsip. - [Berapa biaya sekali jalan agen coding AI: langkah, konteks, caching](https://broadrouter.com/id/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): Kenapa biaya per langkah agen coding naik makin lama run-nya, bagaimana caching mengubah total, dan kapan sesi harga tetap mengalahkan per token. Dengan hitungan. - [Menambahkan harakat ke teks Arab dengan LLM: prompt, cek, DER](https://broadrouter.com/id/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): Pulihkan harakat Arab (tasykil) dengan API LLM: prompt yang menjaga setiap huruf, cek yang menolak perubahan, skor DER dan WER, serta kapan memakai CATT. - [Koreksi tata bahasa dan ejaan Arab dengan LLM: panduan praktis](https://broadrouter.com/id/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Koreksi ejaan dan tata bahasa Arab dengan API LLM: prompt edit minimal, pengaman terhadap penulisan ulang, edit per kata menurut kelas kesalahan, dan skor gaya M2. - [Notulen rapat bahasa Arab dari rekaman: pipeline berbasis API](https://broadrouter.com/id/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Ubah rekaman rapat berbahasa Arab menjadi notulen berisi keputusan dan action item: transkripsi, bersihkan, ringkas dengan LLM ke JSON, lalu cek nama dan tanggal. - [Ekstraksi entitas bernama bahasa Arab dengan LLM: JSON, tanggal, cek](https://broadrouter.com/id/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Ekstrak nama, organisasi, tempat, dan tanggal Hijriah dari teks Arab ke JSON dengan LLM: prompt, cek rentang tanpa harakat, konversi tanggal, dan skrip penilaian. - [Perbandingan API OCR Arab: Google, Azure, Textract, Mistral, LLM](https://broadrouter.com/id/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): API OCR mana yang resmi mendukung bahasa Arab cetak dan tulisan tangan, apa yang dikembalikan masing-masing, biayanya per 1.000 halaman, dibanding API LLM vision. - [Subtitle bahasa Arab (SRT) dari API speech-to-text: yang Anda butuhkan](https://broadrouter.com/id/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): Isi wajib file SRT, API speech-to-text mana yang memberi timestamp atau SRT untuk bahasa Arab, dan pipeline teruji untuk subtitle dari API yang hanya memberi teks. - [Klasifikasi teks bahasa Arab dengan LLM: sentimen, intent, triase](https://broadrouter.com/id/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): Klasifikasikan ulasan, pesan, dan tiket support berbahasa Arab dengan API LLM: desain label, prompt sadar dialek, output JSON, uji macro F1, dan biaya skala besar. - [Normalisasi teks Arab untuk pencarian dan RAG: apa yang perlu diubah](https://broadrouter.com/id/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Tasykil, bentuk alif, ta marbutah, dudukan hamzah, huruf Persia, digit, dan Unicode: aturan normalisasi Arab mana yang membantu pencarian, mana yang merugikan. - [Meringkas teks bahasa Arab dengan LLM: prompt, panjang, dan penilaian](https://broadrouter.com/id/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): Ringkas artikel, laporan, dan pesan dialek Arab dengan API LLM: pola prompt, kontrol jumlah kata, cek angka, dan kenapa ROUGE bawaan memberi bahasa Arab nilai 0. - [Alternatif Artificial Analysis: tempat membandingkan LLM di 2026](https://broadrouter.com/id/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Perbandingan alternatif Artificial Analysis: Arena, Epoch AI, Vals AI, peringkat OpenRouter, Hugging Face, dan data harga, menurut apa yang diukur dan kapan dipakai. - [Hardware GPU dan pendekatan serving BroadNet, menurut BroadNet](https://broadrouter.com/id/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): Apa kata BroadNet tentang hardware di balik BroadNet Inference: tiga tier GPU dari B300 hingga appliance klien, throughput per GPU, dan pelajaran dari pengujiannya. - [Cache respons API LLM untuk memangkas biaya: exact cache di Python](https://broadrouter.com/id/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Exact-match cache teruji untuk respons API LLM di Python: isi cache key, apa yang tidak boleh di-cache, TTL, privasi, dan cara mengukur hit rate lebih dulu. - [Claude Code dengan model terbuka: proxy, endpoint native, dan batasnya](https://broadrouter.com/id/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code bicara dengan Anthropic Messages API. Cara menjalankannya dengan model terbuka lewat proxy LiteLLM atau endpoint native, dan apa yang rusak di jalan. - [Jadikan review manusia bagian dari workflow AI dokumen Anda](https://broadrouter.com/id/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): Rancang tahap review manusia di pipeline AI dokumen: aturan validasi, apa yang dikirim ke reviewer, sampel audit acak, jejak audit, dan tingkat review. - [Undang-undang pelindungan data Mesir (UU 151 Tahun 2020) dan API LLM](https://broadrouter.com/id/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Mengirim data pribadi warga Mesir ke API LLM atau speech-to-text di luar negeri: isi UU 151/2020 dan Keputusan 816/2025 soal transfer, lisensi, dan kebocoran. - [Sebelum fine-tuning: definisikan masalah yang ingin diselesaikan](https://broadrouter.com/id/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): Kapan prompting, RAG, atau evaluasi sudah menyelesaikan masalah tanpa fine-tuning, data latih apa yang dibutuhkan, dan cara menguji fine-tune sebelum memintanya. - [Cara memperbaiki JSON tidak valid dari LLM: penyebab, perbaikan, retry](https://broadrouter.com/id/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): Kenapa LLM mengembalikan JSON rusak (code fence, teks ekstra, terpotong max_tokens, koma di akhir), plus fungsi Python teruji: perbaiki yang aman, tolak sisanya. - [Seberapa sering harga API LLM berubah? Arsip halaman harga 3 tahun](https://broadrouter.com/id/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): Kami membaca 308 arsip halaman harga 11 penyedia API LLM, 2023–2026. 80% model tidak pernah berubah harga; DeepInfra membuat 75% dari semua perubahan harga daftar. - [Cara menghemat biaya API LLM: checklist 10 langkah, berurutan](https://broadrouter.com/id/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): Pangkas tagihan API LLM dengan urutan tepat: ukur, cache prefix, ringkas prompt, batasi output dan penalaran, pilih model yang pas, batch, dan kenali titik impas. - [Failover API LLM: pindah ke penyedia kedua lewat kode Anda sendiri](https://broadrouter.com/id/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): Cara failover antarpenyedia LLM yang kompatibel dengan OpenAI: error mana yang memicu failover, circuit breaker, pemetaan ID model, streaming, biaya, dan privasi. - [Penurunan harga API LLM 2023–2026: yang terbesar, beserta sumbernya](https://broadrouter.com/id/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): Penurunan harga API LLM terbesar 2023–2026, tiap baris dengan tautan Internet Archive sebelum dan sesudah: o3, Codestral, Gemini 1.5, plus kenaikan yang terlewat. - [Apakah harga API LLM selalu turun? 52 kenaikan dan cara menganggarkan](https://broadrouter.com/id/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): Harga API LLM bisa naik. Kami temukan 52 kenaikan harga daftar pada model yang sama di arsip halaman harga 2024–2026, median 40%. Semua sumber, plus aturan anggaran. - [Rate limit API LLM per penyedia: perbandingan RPM, TPM, dan konkurensi](https://broadrouter.com/id/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): Rate limit tier awal API LLM dari 13 penyedia, dibaca dari dokumentasinya: OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks, dan lainnya, plus arti RPM dan TPM. - [Perbandingan SLA API LLM: janji Azure, Bedrock, Google, dan OpenAI](https://broadrouter.com/id/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): Apa yang dijanjikan tertulis oleh tiap penyedia API LLM: persentase uptime, apa yang dihitung downtime, kredit layanan, dan target latensi, dibaca dari halaman SLA. - [Batch processing LLM tanpa Batch API: runner yang bisa dilanjutkan](https://broadrouter.com/id/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Jalankan 10.000 hingga 1.000.000 prompt lewat API chat completions tanpa Batch API: batas concurrency, pengaturan laju, retry, checkpoint, dan hitungan waktu. - [Safety Map: bagaimana 39 LLM menangani data sensitif](https://broadrouter.com/id/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet menguji 39 LLM dari 14 lab pada kredensial, PII, data kesehatan dan keuangan: apa yang berpindah, apa yang tidak, serta kebocoran tool call dan penalaran. - [Cara menguji keamanan data sensitif pada LLM: metode BroadNet](https://broadrouter.com/id/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): Metode BroadNet untuk menguji apakah LLM membocorkan kredensial, PII, data kesehatan atau keuangan sebelum deploy: kategori, bocor vs terlewat, 3+ run. - [Migrasi dari API OpenAI: panduan kode Python, Node, dan cURL](https://broadrouter.com/id/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): Perubahan kode untuk pindah dari OpenAI ke API kompatibel OpenAI: setup klien, Responses ke Chat Completions, tools, streaming, embeddings, dan smoke test. - [Model yang dievaluasi lalu dicoret BroadNet, dan alasannya](https://broadrouter.com/id/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet menerbitkan model yang ditolaknya beserta alasan: model chat yang gagal uji keamanan kredensial, serta model speech, OCR, dan embedding yang kalah. - [Desain produk AI multibahasa: Arab, Inggris, dan Spanyol](https://broadrouter.com/id/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): Mendesain produk AI untuk pengguna Arab, Inggris, dan Spanyol: locale UI vs bahasa model, RTL dan teks campuran, angka, tanggal, deteksi bahasa, dan biaya. - [Biaya API model open-weight vs tertutup: apa kata data harga](https://broadrouter.com/id/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): Apakah model open-weight lewat API lebih murah daripada GPT, Claude, atau Gemini? Harga per token dari halaman penyedia, selisih antarhost, dan biaya per user. - [Riwayat harga API OpenAI vs Anthropic vs Google, 2023–2026](https://broadrouter.com/id/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): Harga API GPT, Claude, dan Gemini berdampingan dari arsip halaman harga 2023–2026: lini mana yang makin murah, mana yang makin mahal, tiap harga bertautan arsip. - [Base URL API kompatibel OpenAI: 14 penyedia, sudah dicek](https://broadrouter.com/id/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): base_url untuk 14 API LLM kompatibel OpenAI, masing-masing dibaca dari dokumentasi resmi penyedia, beserta endpoint dan perbedaan yang bisa merusak migrasi. - [OpenClaw dan Hermes Agent dengan model murah: setup dan hitungan token](https://broadrouter.com/id/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): Hubungkan OpenClaw atau Hermes Agent ke endpoint kompatibel OpenAI, pahami kenapa cek terjadwal boros token, dan tekan tagihan dengan konteks ringan dan model murah. - [Endpoint privat dimulai dari beban kerja yang jelas](https://broadrouter.com/id/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): Apa yang perlu ditentukan sebelum meminta endpoint LLM privat: beban kerja, token per detik, konkurensi, konteks, region, lisensi, anggaran, lalu cara mengujinya. - [Prompt injection pada coding agent: pertahanan yang benar-benar ampuh](https://broadrouter.com/id/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): Dari mana prompt injection masuk ke coding agent, dan pertahanan berlapis yang didokumentasikan vendor dan peneliti: sandbox, batas egress, allow-list, persetujuan. - [Chunk size dan overlap RAG: pengaruhnya pada retrieval dan biaya](https://broadrouter.com/id/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): Bagaimana chunk size dan overlap mengubah recall RAG, ukuran indeks, biaya embedding, dan token yang dikirim tiap query ke LLM, dengan hasil publikasi dan hitungan. - [Mengatur max_tokens untuk model penalaran: metode penganggaran](https://broadrouter.com/id/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Kenapa model penalaran memberi balasan kosong saat max_tokens rendah, dan cara menentukan anggaran dari persentil terukur, context window, dan retry yang bertahap. - [Merekam audio di browser dan mengirimnya ke API speech-to-text](https://broadrouter.com/id/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): Rekam mikrofon dengan MediaRecorder, ubah WebM atau MP4 Safari menjadi WAV 16 kHz, lalu kirim ke API transkripsi lewat server kecil yang menyimpan kunci API. - [Mengurangi token prompt: hasil terukur untuk system prompt, tool, data](https://broadrouter.com/id/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): Penghematan token terukur dari memangkas system prompt, contoh few-shot, skema tool, format data, riwayat chat, dan harakat Arab, dihitung dengan tokenizer asli. - [Model sama, host berbeda: seberapa jauh harga API LLM bervariasi](https://broadrouter.com/id/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): Untuk 119 model yang dijual beberapa host, harga daftar termahal secara median 1,76 kali yang termurah. Data dari halaman harga 41 penyedia, November 2026. - [PDPL Saudi dan UEA untuk API LLM: checklist transfer lintas negara](https://broadrouter.com/id/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Mengirim data pribadi ke API LLM dari Arab Saudi atau UEA: isi aturan transfer PDPL, apa yang ditanyakan ke penyedia, dan kapan inferensi di dalam negeri penting. - [Teks, suara, atau vision? Mulailah dari tugasnya](https://broadrouter.com/id/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): Pilih chat teks, input gambar, atau transkripsi audio sesuai tugas, dengan model dan batasan yang didokumentasikan BroadNet, plus yang tidak tersedia seperti TTS. - [Transkripsi audio panjang di atas 25 MB: pecah, kirim, gabungkan](https://broadrouter.com/id/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): Cara mentranskripsi rekaman yang melebihi batas file API speech-to-text: perkecil file, pecah di jeda hening dengan ffmpeg, kirim paralel, lalu gabungkan teksnya. - [String aplikasi ke bahasa Arab dengan LLM: placeholder dan plural](https://broadrouter.com/id/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): Terjemahkan file JSON i18n ke bahasa Arab dengan API LLM: bekukan placeholder, hasilkan enam bentuk plural Arab, validasi sintaks ICU di CI, isolasi nama sisipan. - [Terjemahkan audio bahasa Arab ke teks Inggris via API: pipeline teruji](https://broadrouter.com/id/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): Ubah rekaman bahasa Arab jadi teks Inggris dengan dua panggilan API: transkripsi language=ar, lalu terjemahan LLM, plus cek angka, chunking, biaya, dan penilaian. - [Live, on request, riset: cara membaca katalog model dengan benar](https://broadrouter.com/id/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): Cara membedakan model katalog yang bisa dipanggil hari ini, yang perlu pengajuan, dan yang hanya diuji, dengan katalog BroadNet dan /v1/models sebagai contoh. - [Uno di vLLM: apa yang diukur studi serving BroadNet](https://broadrouter.com/id/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet memasang adapter self-drafting Uno ke vLLM dan mengukurnya di H100: throughput 2,58x vLLM biasa pada satu stream, 1,74x pada 32. Metode, DOI, dan batasan. - [Uno di Gemma 4 26B: studi mixture-of-experts dari BroadNet](https://broadrouter.com/id/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet melatih adapter self-drafting Uno pertama untuk model mixture-of-experts, Gemma 4 26B A4B. Apa yang diukur, kapan self-drafting menguntungkan, dan DOI. - [Anggaran token agen: ke mana token loop agen pergi](https://broadrouter.com/id/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): System prompt, skema tool, pembacaan file, riwayat, dan penalaran: cara loop agen memakai token, cara mengukur tiap bagian, dan apa yang dipangkas lebih dulu. - [Word error rate (WER) ASR bahasa Arab: penjelasan lengkap](https://broadrouter.com/id/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): Apa arti word error rate, cara menghitungnya, dan kenapa WER bahasa Arab sangat peka terhadap harakat, normalisasi ejaan, dan dialek. Dengan Python dan angka nyata. - [Speech recognition dialek Arab: Teluk, Mesir, Levant, Maghribi](https://broadrouter.com/id/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Mengapa speech-to-text kesulitan dengan bahasa Arab Teluk, Mesir, Levant, dan Maghribi, dan cara menyusun test set dialek dari audio Anda sendiri. - [Terjemahan Arab–Inggris dengan LLM: prompt, glosarium, evaluasi](https://broadrouter.com/id/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): Terjemahkan bahasa Arab ke Inggris dan sebaliknya dengan API LLM: prompt yang menjaga nada, glosarium istilah tetap, angka dan nama, serta cara menilai kualitas. - [Bahasa Arab Teluk, Mesir, dan Syam: cara LLM menangani dialek](https://broadrouter.com/id/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): Bagaimana LLM menangani bahasa Arab Teluk, Mesir, dan Syam: di mana gagal, cara menulis prompt untuk dialek, cara mengujinya, dan laporan BroadNet untuk Qwen3.8 27B. - [OCR bahasa Arab vs vision LLM: membaca dokumen dan faktur Arab](https://broadrouter.com/id/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): Kapan memakai OCR klasik dan kapan vision LLM untuk faktur, struk, dan formulir bahasa Arab: jebakan akurasi, output JSON, biaya per halaman, dan pipeline validasi. - [Prompt engineering bahasa Arab: register, RTL, angka, teks campuran](https://broadrouter.com/id/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): Tulis prompt agar LLM menghasilkan bahasa Arab yang baik: pilih register dan dialek, kendalikan angka dan tanggal, rapikan format RTL, tangani campuran Arab–Inggris. - [RAG bahasa Arab: normalisasi, chunking, dan retrieval yang berhasil](https://broadrouter.com/id/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): Retrieval-augmented generation untuk dokumen bahasa Arab: normalisasi alif, ya, dan ta marbuta, chunking yang masuk akal, embedding multibahasa, plus keyword search. - [API speech-to-text bahasa Arab: cara memilihnya di 2026](https://broadrouter.com/id/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): Cara memilih API transkripsi bahasa Arab: word error rate, dialek, batas file, harga per jam audio, dan privasi, plus rencana uji untuk rekaman Anda sendiri. - [Arsitektur chatbot WhatsApp bahasa Arab: dari voice note ke balasan](https://broadrouter.com/id/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): Arsitektur yang berjalan untuk asisten WhatsApp bahasa Arab: webhook, antrean, transkripsi voice note, balasan LLM, aturan kebijakan, dan biaya, dengan kode Python. - [Arabizi dan LLM: menangani Franco-Arab dengan angka seperti 3 dan 7](https://broadrouter.com/id/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): Arabizi menulis bahasa Arab dengan huruf Latin dan angka, misalnya 3 untuk ع dan 7 untuk ح. Cara mendeteksi, mengonversi ke aksara Arab dengan LLM, dan membalasnya. - [LLM terbaik untuk Aider: cara memilih model dan menghubungkannya](https://broadrouter.com/id/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Cara memilih LLM terbaik untuk Aider: disiplin format edit, ukuran konteks, biaya per jam, tes satu jam di repo Anda sendiri, dan setup OPENAI_API_BASE yang persis. - [Model terbuka terbaik untuk agen coding: tools, konteks, dan biaya](https://broadrouter.com/id/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B, atau Gemma 4 untuk agen coding? Pilih dari tool calling, konteks, penalaran, dan biaya per tugas, lengkap dengan hitungan. - [Token input cache: cara memangkas tagihan API LLM](https://broadrouter.com/id/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): Input cache jauh lebih murah daripada input biasa. Cara prompt caching ditagih, cara membaca cached_tokens di usage, dan cara menyusun prompt agar kena cache. - [API LLM vision termurah: biaya per gambar dan per halaman, dihitung](https://broadrouter.com/id/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): Biaya LLM vision per gambar dan per halaman dokumen: hitungan token untuk Gemma 4, Qwen3.8 27B, dan DeepSeek V4.1 Flash, plus di mana DeepInfra lebih murah. - [Cara memilih penyedia API LLM: checklist 20 poin](https://broadrouter.com/id/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): Dua puluh pertanyaan sebelum memilih API LLM: harga, batas, kompatibilitas OpenAI, privasi, model, dukungan, dan uptime, plus lembar penilaian dan rencana uji. - [Memilih API LLM untuk bahasa Arab: apa yang perlu diuji](https://broadrouter.com/id/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): Cara memilih LLM untuk aplikasi bahasa Arab: pemahaman dialek, kualitas tulisan, terjemahan, dokumen dan gambar Arab, biaya, plus test set yang bisa dibuat sehari. - [Cline vs Roo Code vs Aider untuk tim yang memakai model terbuka (2026)](https://broadrouter.com/id/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (kini diarsipkan, lanjut sebagai Zoo Code), dan Aider dibandingkan dari sisi penting bagi model terbuka: tool call, konteks, MCP, git, approval. - [Cara menghitung token sebelum mengirim permintaan LLM](https://broadrouter.com/id/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): Perkirakan token prompt per bahasa dan model sebelum memanggil API LLM: kenapa tokenizer berbeda, contoh terukur dalam lima bahasa, dan margin max_tokens yang aman. - [10 cara menekan biaya LLM agen coding, lengkap dengan estimasi hemat](https://broadrouter.com/id/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Sepuluh cara konkret menekan biaya token Cline, Aider, opencode, atau agen coding buatan Anda sendiri, masing-masing dengan hitungan hematnya dan cara mengukurnya. - [Alternatif API DeepSeek: 5 cara memanggil model DeepSeek](https://broadrouter.com/id/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): Memanggil model DeepSeek lewat API resmi, agregator, katalog cloud, GPU sendiri, atau BroadNet Inference: untung-rugi soal harga, privasi, dan batas pemakaian. - [Harga API DeepSeek V4.1 Flash: per token, input cache, dan paket](https://broadrouter.com/id/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): Harga API DeepSeek V4.1 Flash per 1 juta token input, input cache, dan output, paket bulanan, mingguan, per jam, serta perbandingan DeepSeek, DeepInfra, Fireworks. - [Cara menghitung estimasi biaya API LLM sebelum meluncurkan fitur](https://broadrouter.com/id/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): Model biaya langkah demi langkah untuk fitur AI baru: pengguna × aksi × panggilan × token, dengan cache, penalaran, retry, rate limit, dan contoh tabel bulanan. - [Cara mengevaluasi LLM sebelum berpindah penyedia](https://broadrouter.com/id/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): Susun set eval dari trafik nyata, nilai dengan pengecekan kode dan LLM juri yang dikoreksi biasnya, lalu bandingkan kualitas, biaya, dan latensi sebelum pindah. - [Ekstrak data faktur, struk, dan formulir ke JSON dengan vision LLM](https://broadrouter.com/id/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): Pipeline produksi untuk faktur, struk, dan formulir: persiapan gambar, prompt skema JSON, validasi Pydantic, cek aritmetika, retry, dan antrean review manusia. - [API LLM harga tetap vs bayar per token: hitungan titik impas](https://broadrouter.com/id/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): Kapan paket LLM harga tetap lebih murah daripada bayar per token? Empat contoh tagihan bulanan DeepSeek V4.1 Flash, volume titik impas, dan cara mengukurnya sendiri. - [Kredit gratis API LLM: seberapa jauh kredit awal benar-benar cukup](https://broadrouter.com/id/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): Apa yang didapat dari kredit gratis API LLM? Token, menit audio, dan kunci per jam dengan kredit awal BroadNet, dibanding free tier Gemini dan OpenRouter. - [GDPR dan API LLM: 15 pertanyaan untuk penyedia Anda](https://broadrouter.com/id/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): Checklist GDPR untuk API LLM: DPA, subprosesor, retensi, pelatihan model, transfer internasional, dan notifikasi pelanggaran, lengkap dengan pasal GDPR-nya. - [GLM 5.3 vs DeepSeek V4.1 Flash: mana yang dipakai, dan kapan](https://broadrouter.com/id/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 vs DeepSeek V4.1 Flash dari sisi harga, konteks, tool, output terstruktur, gambar, dan penalaran, dengan contoh biaya beban kerja dan aturan memilih. - [Cara mengatasi error 429 API LLM: backoff, jitter, dan limiter](https://broadrouter.com/id/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 vs 5xx, exponential backoff dengan full jitter, limiter konkurensi, dan antrean untuk API LLM, dengan kode Python dan Node yang siap dipakai di produksi. - [Kunci API LLM per jam: harga tetap per jam, dijelaskan](https://broadrouter.com/id/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): Kunci per jam BroadNet memberi satu jam Gemma 4 VL 26B pada satu stream dengan harga tetap. Apa yang termasuk, batasnya, cocok untuk siapa, dan kapan lebih hemat. - [Perbandingan retensi data API LLM: OpenAI, Anthropic, Gemini, Mistral](https://broadrouter.com/id/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): Apa kata resmi OpenAI, Anthropic, Gemini API, Mistral, dan BroadNet soal menyimpan prompt API dan memakainya untuk pelatihan, dikutip dari halaman mereka (Nov 2026). - [Keamanan kunci API LLM: penyimpanan, rotasi, dan IP allow-list](https://broadrouter.com/id/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): Jauhkan kunci API LLM dari kode dan log, beri tiap workload kuncinya sendiri, rotasi terjadwal, kunci ke IP server, dan cabut kunci yang bocor dalam hitungan detik. - [Harga API LLM: input, input cache, output, dan token penalaran](https://broadrouter.com/id/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): Cara kerja harga API LLM per token: input, input cache, output, token penalaran, gambar, dan audio per menit, dengan contoh hitungan dari daftar tarif resmi. - [Mengirim data sensitif ke API LLM: redaksi, minimalkan, kurangi log](https://broadrouter.com/id/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): Cara aman mengirim data pelanggan, kesehatan, atau keuangan ke API LLM: klasifikasikan field, redaksi dengan placeholder yang bisa dibalik, pakai zero retention. - [Kontrol biaya API LLM: anggaran, batas kunci, dan proteksi loop](https://broadrouter.com/id/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): Cegah tagihan API LLM membengkak: saldo prabayar, satu kunci per aplikasi, IP allow-list, pencatatan pemakaian, anggaran di kode, batas loop, dan kunci harga tetap. - [Logging pemakaian API LLM: biaya per pengguna dan fitur, tanpa PII](https://broadrouter.com/id/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): Catat setiap panggilan API LLM beserta token, biaya, latensi, dan request ID, atribusikan biaya ke pengguna dan fitur, dan redaksi PII sebelum ditulis. Kode Python. - [Biaya LLM per user: unit economics fitur AI di SaaS](https://broadrouter.com/id/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): Hitung biaya fitur AI per user aktif per bulan, kenapa heavy user lebih penting dari rata-rata, dan harga paket SaaS yang menutup tagihan model. - [Cara 6 penyedia inference menetapkan harga API model open-weight](https://broadrouter.com/id/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): Cara OpenRouter, Together AI, Fireworks, Groq, DeepInfra, dan BroadNet menetapkan harga model terbuka: per token, cache, batch, jam GPU, dan paket harga tetap. - [Latensi LLM: TTFT vs token per detik vs waktu end-to-end](https://broadrouter.com/id/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): Apa yang diukur time to first token, token per detik, dan latensi end-to-end, cara mengukurnya dari klien dengan kode, dan perbaikan mana mengubah angka mana. - [Penjelasan streaming LLM: chunk SSE, delta, usage, dan pembatalan](https://broadrouter.com/id/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Cara kerja streaming chat completion, cara menyusun ulang jawaban dari delta, membaca usage dari chunk terakhir, dan membatalkan dengan rapi di Python, Node, curl. - [Long context vs RAG: kapan 1M token menggantikan vector database](https://broadrouter.com/id/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): Kapan context window 1M token bisa menggantikan vector database dan kapan RAG tetap unggul, dibandingkan dari biaya per query, latensi, recall, dan pemeliharaan. - [Manajemen konteks untuk agent yang berjalan lama dengan window 1M](https://broadrouter.com/id/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): Cara menjaga sesi agent panjang tetap akurat dan hemat dengan window 1M token: riwayat append-only, prompt ramah cache, kapan compaction, dan apa yang dibuang dulu. - [Agent yang berjalan lama: timeout, retry, dan checkpoint yang andal](https://broadrouter.com/id/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): Jaga agent LLM berjalan berjam-jam: timeout per panggilan, streaming untuk deteksi macet, retry aman, tool idempoten, dan checkpoint di setiap langkah. - [Memakai server MCP dengan model kompatibel OpenAI (penyedia apa pun)](https://broadrouter.com/id/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP bisa dipakai dengan LLM apa pun: client mengubah tool MCP menjadi function call. Cara memakai MCP dengan model terbuka di Cline, Continue, opencode, n8n, Python. - [Kapan context window 1M token sepadan (dan kapan tidak)](https://broadrouter.com/id/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): Apa yang muat dalam 1.048.576 token, berapa biayanya di DeepSeek V4.1 Flash, bagaimana penalaran dan caching mengubah hitungannya, dan kapan RAG tetap lebih baik. - [Biaya sebenarnya setup multi-agent: per token vs sesi harga tetap](https://broadrouter.com/id/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): Sistem multi-agent membakar token berkali lipat dibanding satu agen. Hitungan sesi harga tetap vs per token, pengali tersembunyi, dan kapan masing-masing menang. - [API speech-to-text multibahasa: 14 bahasa, satu endpoint](https://broadrouter.com/id/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): Transkripsikan Inggris, Spanyol, Prancis, Mandarin, Arab, dan 9 bahasa lain lewat satu endpoint kompatibel OpenAI berharga per menit, dan kapan memakai model Arab. - [AI Agent n8n dengan LLM murah: setup, pilihan model, biaya per run](https://broadrouter.com/id/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): Bangun workflow AI Agent n8n di model kompatibel OpenAI yang murah: model mana untuk tool, pengaturan yang menghemat token, dan hitungan biaya per 10.000 run. - [Model AI open-weight vs tertutup untuk bisnis: biaya, kontrol, risiko](https://broadrouter.com/id/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): LLM open-weight atau tertutup untuk bisnis Anda? Bandingkan biaya, kontrol, privasi, selisih kualitas, dan lock-in, plus tiga cara menjalankan model open. - [Pindah ke API kompatibel OpenAI: checklist migrasi](https://broadrouter.com/id/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): Mengganti base_url itu bagian mudahnya. Yang sering rusak: ID model, token penalaran, usage saat streaming, error, rate limit, dan kunci API. Checklist dengan kode. - [OpenRouter vs paket harga tetap untuk AI agent: mana yang lebih murah?](https://broadrouter.com/id/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter menagih per token plus biaya beli kredit; paket harga tetap menagih per sesi. Hitungan biaya agen yang selalu aktif, biaya, batas, dan pilihannya. - [Cara menjalankan coding agent semalaman tanpa pengawasan dengan aman](https://broadrouter.com/id/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): Setup praktis untuk coding agent tanpa pengawasan: sandbox, checkpoint git, file tugas, batas langkah, tinjauan pagi, dan sesi paket tetap untuk malam hari. - [Cara menyusun prompt agar prefiksnya ter-cache](https://broadrouter.com/id/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): Cara menyusun prompt LLM agar prefiks yang di-cache tetap identik: system prompt, skema tool, dokumen, riwayat yang hanya ditambah, plus tes untuk membuktikannya. - [Token penalaran: mengapa Anda membayar output lebih dari yang terlihat](https://broadrouter.com/id/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): Model penalaran menagih proses berpikir tersembunyi sebagai output. Cara mengukurnya, biaya per model, dan cara memangkasnya: reasoning_effort, max_tokens, routing. - [Cara menjalankan agen coding paralel dengan API LLM harga tetap](https://broadrouter.com/id/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Subagent, orchestrator, dan batch worker di DeepSeek V4.1 Flash unlimited: berapa sesi yang dibutuhkan, cara membaginya ke kunci, dan cara membatasi tumpang tindih. - [Self-hosting LLM vs API: biaya sebenarnya dan titik impasnya](https://broadrouter.com/id/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): Harga sewa GPU dari halaman resmi, rumus biaya per 1 juta token di GPU sendiri, utilisasi untuk mengalahkan API, dan biaya operasional yang sering terlupa. - [Perbandingan harga API speech-to-text: harga per menit di 2026](https://broadrouter.com/id/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): Harga per menit dan per 1.000 jam untuk OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq, dan BroadNet Transcribe dari halaman resmi, plus biaya add-on. - [JSON mode vs json_schema: output terstruktur yang andal](https://broadrouter.com/id/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON mode memberi JSON valid; json_schema memberi bentuk Anda. Cara memintanya di model kompatibel OpenAI, validasi dengan Pydantic, dan perbaiki output buruk. - [Token per kata bahasa Arab vs Inggris: mengapa bahasa Arab lebih mahal](https://broadrouter.com/id/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): Mengapa bahasa Arab kerap memakai lebih banyak token LLM daripada bahasa Inggris untuk makna yang sama, mengapa token per kata menyesatkan, dan skrip pengukurnya. - [Tool calling dengan model terbuka: skema, panggilan paralel, retry](https://broadrouter.com/id/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): Function calling di model terbuka yang kompatibel OpenAI: tips skema, tool call paralel, lima penyebab tool call gagal, dan loop Python yang memvalidasi dan retry. - [Cara transkripsi rekaman call center: pipeline siap bahasa Arab](https://broadrouter.com/id/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): Pisahkan panggilan stereo per kanal, potong giliran bicara, transkripsikan bahasa Arab dan Inggris, samarkan PII, lalu ringkas dengan LLM. Kode Python dan biaya. - [Model terbuka di agen coding: Aider, Cline, opencode, Codex](https://broadrouter.com/id/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Arahkan Aider, Cline, Roo Code, opencode, Continue, dan Codex CLI ke endpoint kompatibel OpenAI mana pun. Nama field tiap agen, plus pengaturan yang sering rusak. - [Voice note ke teks untuk bisnis: dari WhatsApp dan Telegram ke tiket](https://broadrouter.com/id/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): Ubah voice note pelanggan dari WhatsApp Business dan bot Telegram menjadi transkrip dan tiket support: unduh, transkripsi, ekstrak field dengan LLM, lalu routing. - [Paket LLM mingguan untuk proyek singkat: Orbit vs per jam vs bulanan](https://broadrouter.com/id/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): Kapan paket DeepSeek V4.1 Flash unlimited seminggu mengalahkan sesi per jam, paket bulanan, atau bayar per token: hitungan untuk hackathon, sprint, dan migrasi. - [Apa arti "unlimited" yang sebenarnya di paket API LLM](https://broadrouter.com/id/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Meteran token, jendela 5 jam, batas mingguan, rate limit, konkurensi: 7 pertanyaan untuk menguji apakah paket AI "unlimited" memang tanpa batas bagi agen Anda. - [Alternatif API Whisper: kapan beralih dan cara mengujinya](https://broadrouter.com/id/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): Kapan alternatif hosted untuk API Whisper OpenAI masuk akal: word error rate di audio Anda, bahasa, harga per jam, batas, dan privasi, lengkap dengan kode uji. - [API LLM zero data retention: checklist 10 pertanyaan](https://broadrouter.com/id/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): Arti "zero data retention" yang seharusnya untuk API LLM: prompt, output, log, pelatihan, subprosesor, dan audit. Sepuluh pertanyaan sebelum Anda mengirim data. - [Guides in Indonesian](https://broadrouter.com/id/guides/) - [Memakai OpenAI Python SDK dengan BroadNet Inference](https://broadrouter.com/id/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): Arahkan OpenAI Python SDK resmi ke BroadNet Inference cukup dengan mengganti base_url. Instal, atur kunci, kirim permintaan pertama, dan stream jawabannya. - [Memakai OpenAI Node.js SDK dengan BroadNet Inference](https://broadrouter.com/id/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): Arahkan OpenAI Node.js dan TypeScript SDK resmi ke BroadNet Inference dengan mengatur baseURL. Instal, kirim permintaan pertama, stream token, dan catat request ID. - [Coba BroadNet Inference di browser, tanpa akun](https://broadrouter.com/id/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): Coba speech-to-text bahasa Arab, pembacaan dokumen, dan chat Gemma 4 di playground browser BroadNet tanpa daftar atau kartu, lalu pindahkan permintaannya ke API. - [Permintaan pertama ke BroadNet Inference dengan curl](https://broadrouter.com/id/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): Kirim panggilan pertama Anda ke BroadNet Inference dengan curl: chat completion, balasan streaming, dan transkripsi audio. Lalu baca respons, usage, dan error-nya. - [Streaming chat completions dan menangani koneksi terputus](https://broadrouter.com/id/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): Cara kerja streaming di BroadNet Inference: Server-Sent Events, penanda [DONE], usage token di chunk terakhir, dan langkah jika stream berhenti lebih awal. - [Kenapa model reasoning mengembalikan jawaban kosong](https://broadrouter.com/id/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3, dan Qwen3.8 27B bernalar sebelum menjawab, dan penalaran itu ikut dihitung dalam max_tokens. Begini cara agar balasan tidak kosong. - [Kode error BroadNet Inference dan kapan harus mencoba ulang](https://broadrouter.com/id/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): Arti setiap status HTTP dari API BroadNet Inference, mana yang dicoba ulang dengan backoff, mana yang harus diperbaiki dulu, dan apa yang dikirim ke tim dukungan. - [Batasi kunci API ke server Anda dengan IP allow-list](https://broadrouter.com/id/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): Lindungi kunci BroadNet Inference dengan IP allow-list, pisahkan dari kredensial portal, dan cabut kunci yang bocor dalam hitungan detik. Dengan contoh curl. - [Cara kerja paket API LLM unlimited untuk agen](https://broadrouter.com/id/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): Paket DeepSeek V4.1 Flash unlimited tidak punya meteran token dan batas mingguan. Begini cara kerja sesi, agen paralel, kunci, dan penagihannya yang sebenarnya. - [Menjalankan Aider dengan DeepSeek V4.1 Flash unlimited](https://broadrouter.com/id/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Hubungkan agen coding Aider ke BroadNet Inference lewat pengaturan kompatibel OpenAI, lalu jalankan dengan paket DeepSeek V4.1 Flash unlimited berharga tetap. - [Memakai LangChain dengan BroadNet Inference](https://broadrouter.com/id/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): Hubungkan ChatOpenAI dari LangChain ke BroadNet Inference dengan base_url dan api_key, dan pahami apa yang disimpan dan dibuang LangChain dari model penalaran. - [Memakai Cline dengan BroadNet Inference (provider OpenAI Compatible)](https://broadrouter.com/id/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Siapkan ekstensi Cline di VS Code dengan BroadNet Inference lewat provider OpenAI Compatible: field, ID model, ukuran konteks, dan solusi masalah umum. - [Pakai Roo Code dengan BroadNet Inference (OpenAI Compatible)](https://broadrouter.com/id/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code dihentikan pada Mei 2026. Pengaturan untuk instalasi lama dan fork-nya, serta ke mana pindah (Cline atau fork Zoo Code) dengan BroadNet Inference. - [Memakai Continue dengan BroadNet Inference (config.yaml dan apiBase)](https://broadrouter.com/id/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): Tambahkan model BroadNet Inference ke ekstensi IDE Continue dengan provider openai dan apiBase di config.yaml, termasuk capability tool_use dan panjang konteks. - [Menambahkan BroadNet Inference ke opencode sebagai custom provider](https://broadrouter.com/id/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): Tambahkan BroadNet Inference ke agen coding opencode sebagai custom provider OpenAI-compatible di opencode.jsonc, lengkap dengan alias model, kapabilitas, dan limit. - [Menghubungkan Open WebUI ke BroadNet Inference](https://broadrouter.com/id/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Tambahkan BroadNet Inference ke Open WebUI sebagai koneksi kompatibel OpenAI, masukkan ID model ke allowlist, dan atasi masalah verifikasi, balasan kosong, dan 429. - [Menambahkan BroadNet Inference ke LibreChat sebagai custom endpoint](https://broadrouter.com/id/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): Tambahkan BroadNet Inference ke LibreChat sebagai custom endpoint kompatibel OpenAI di librechat.yaml, dengan kunci di .env, daftar model, model judul, dan solusi. - [Memakai BroadNet Inference di n8n (kredensial OpenAI, Base URL)](https://broadrouter.com/id/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): Arahkan kredensial OpenAI di n8n ke BroadNet Inference lewat field Base URL, lalu pakai node OpenAI Chat Model dengan mode ID dan Responses API dimatikan. - [Memakai Vercel AI SDK dengan BroadNet Inference](https://broadrouter.com/id/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): Panggil BroadNet Inference dari Vercel AI SDK dengan createOpenAICompatible dari @ai-sdk/openai-compatible: generateText, streamText, tools, dan reasoning effort. - [API speech-to-text bahasa Arab, langkah demi langkah](https://broadrouter.com/id/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): Transkripsikan audio bahasa Arab dengan BroadNet Transcribe Arabic lewat endpoint kompatibel OpenAI. Format, field language, contoh Python dan curl, serta harganya. - [Mengarahkan LiteLLM ke BroadNet Inference (SDK dan proxy)](https://broadrouter.com/id/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): Panggil BroadNet Inference lewat LiteLLM dengan prefiks model openai/ dan api_base, di Python SDK atau config.yaml proxy, dengan rate limit yang sesuai. - [Menjalankan OpenAI Agents SDK di BroadNet Inference](https://broadrouter.com/id/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Pakai OpenAI Agents SDK untuk Python dengan BroadNet Inference lewat AsyncOpenAI dan OpenAIChatCompletionsModel, lalu matikan tracing yang butuh kunci OpenAI. - [Transkripsi voice note WhatsApp ke teks dengan API](https://broadrouter.com/id/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): Konversi voice note WhatsApp .ogg atau rekaman ponsel .m4a ke WAV dengan ffmpeg, lalu transkripsikan dalam bahasa Arab atau 13 bahasa lain lewat satu panggilan API. - [Transkripsi voice note WhatsApp di n8n](https://broadrouter.com/id/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): Workflow n8n yang menerima voice note WhatsApp, mengunduhnya dari Meta, mengonversinya dengan ffmpeg, lalu mentranskripsinya dalam bahasa Arab atau 13 bahasa lain. - [Membaca faktur dan struk berbahasa Arab dengan model vision](https://broadrouter.com/id/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Kirim foto faktur atau struk ke model vision di BroadNet Inference dan dapatkan total, tanggal, dan penjual sebagai JSON, bertanya dalam bahasa Arab atau Inggris. - [Zed dengan BroadNet Inference sebagai provider kompatibel OpenAI](https://broadrouter.com/id/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Tambahkan BroadNet Inference ke editor Zed sebagai provider kompatibel OpenAI di settings.json, dengan ID model, batas token, capabilities, dan variabel kunci. - [Memakai Kilo Code dengan BroadNet Inference (OpenAI Compatible)](https://broadrouter.com/id/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Hubungkan Kilo Code ke BroadNet Inference sebagai custom provider OpenAI Compatible, lewat dialog VS Code atau kilo.json: ID model, limit, dan flag tool. - [Menambahkan BroadNet Inference ke goose sebagai custom provider](https://broadrouter.com/id/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): Hubungkan agen AI goose ke BroadNet Inference sebagai custom provider OpenAI Compatible, lewat goose Desktop, goose configure, atau sebuah file JSON. - [Pydantic AI dengan BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/id/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Jalankan agen Pydantic AI di BroadNet Inference dengan OpenAIChatModel dan OpenAICompatibleProvider, plus tool call, output terstruktur, dan model profile. - [CrewAI dengan BroadNet Inference sebagai endpoint OpenAI kustom](https://broadrouter.com/id/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): Arahkan agen CrewAI ke BroadNet Inference dengan LLM(custom_openai=True, base_url=...), lalu pilih model untuk tools, konteks panjang, dan paket harga tetap. - [Memakai LlamaIndex dengan BroadNet Inference lewat OpenAILike](https://broadrouter.com/id/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): Hubungkan LlamaIndex ke BroadNet Inference dengan class OpenAILike, lalu atur api_base, context_window, is_chat_model, dan is_function_calling_model dengan benar. - [BroadNet Inference di Dify lewat plugin OpenAI-API-compatible](https://broadrouter.com/id/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): Hubungkan Dify ke BroadNet Inference lewat plugin model provider OpenAI-API-compatible, lengkap dengan nilai field, ukuran konteks, dan pengaturan function calling. - [Jan dengan BroadNet Inference sebagai provider kompatibel OpenAI](https://broadrouter.com/id/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Tambahkan BroadNet Inference ke aplikasi desktop Jan sebagai custom provider kompatibel OpenAI, tambahkan ID model manual, dan aktifkan tools serta vision per model. - [Hubungkan SillyTavern ke BroadNet Inference (Custom OpenAI-compatible)](https://broadrouter.com/id/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): Hubungkan SillyTavern ke BroadNet Inference lewat sumber Chat Completion Custom (OpenAI-compatible): ID model yang benar, prompt post-processing, batas token. - [Pakai Spring AI dengan BroadNet Inference (OpenAI base-url)](https://broadrouter.com/id/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Arahkan starter chat OpenAI Spring AI ke BroadNet Inference dengan spring.ai.openai.base-url, ID model BroadNet, dan max-tokens, lalu panggil lewat ChatClient. - [Memanggil BroadNet Inference dari Go dengan SDK openai-go](https://broadrouter.com/id/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): Pakai SDK resmi openai-go v3 dengan BroadNet Inference lewat option.WithBaseURL: chat completions, streaming dengan usage, dan penanganan error di Go. - [Membangun agen LangGraph di BroadNet Inference](https://broadrouter.com/id/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): Jalankan StateGraph atau create_agent LangGraph di BroadNet Inference dengan ChatOpenAI, base_url, dan ID model BroadNet, plus tools, streaming, dan memori. - [Memakai pipeline dan agen Haystack dengan BroadNet Inference](https://broadrouter.com/id/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Arahkan OpenAIChatGenerator Haystack ke BroadNet Inference dengan api_base_url, ID model BroadNet, dan kunci Secret, untuk pipeline, tools, streaming, dan agen. - [Pakai Semantic Kernel dengan BroadNet Inference](https://broadrouter.com/id/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Hubungkan Microsoft Semantic Kernel ke BroadNet Inference: di Python lewat klien AsyncOpenAI, di C# lewat endpoint kustom, untuk chat, plugin, dan agen. - [Memakai agen Mastra dengan BroadNet Inference](https://broadrouter.com/id/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): Jalankan agen Mastra di BroadNet Inference dengan objek model berisi id, url, dan apiKey, termasuk prefiks broadnet/ ganda yang dibutuhkan ID modelnya. - [Mendapatkan output terstruktur dengan Instructor di BroadNet Inference](https://broadrouter.com/id/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Ekstrak objek Pydantic dari BroadNet Inference dengan Instructor, memakai from_openai atau from_provider dengan base_url kustom, dan pilih Mode yang tepat. - [Memakai DSPy dengan BroadNet Inference (atur model_type="chat")](https://broadrouter.com/id/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): Konfigurasikan dspy.LM untuk BroadNet Inference dengan prefiks openai/, api_base, dan model_type="chat", karena DSPy 3.5 mengirim model OpenAI ke Responses API. - [Mengevaluasi model BroadNet dengan promptfoo](https://broadrouter.com/id/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): Jalankan eval promptfoo dan uji model berdampingan di BroadNet Inference dengan provider openai:chat, apiBaseUrl, dan apiKeyEnvar, plus grader. - [Jalankan smolagents di BroadNet Inference dengan OpenAIModel](https://broadrouter.com/id/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Hubungkan smolagents dari Hugging Face ke BroadNet Inference lewat OpenAIModel dan api_base: CodeAgent di model apa pun, ToolCallingAgent di GLM 5.3 dan Qwen3.8 27B. - [Menghubungkan Flowise ke BroadNet Inference (OpenAI Custom Model)](https://broadrouter.com/id/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Pakai BroadNet Inference di chatflow Flowise lewat node OpenAI Custom Model, dengan Base Path, credential OpenAI API, dan nama model BroadNet yang diketik manual. - [Menghubungkan Langflow ke BroadNet Inference](https://broadrouter.com/id/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Pakai BroadNet Inference di Langflow lewat field OpenAI API Base komponen OpenAI dan nama model yang diketik manual, plus alasan OpenAI Compatible tak cocok. - [Glossary entries in Indonesian](https://broadrouter.com/id/glossary/) - [Agen AI (AI agent)](https://broadrouter.com/id/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): Agen AI adalah sistem di mana model bahasa mengarahkan langkahnya sendiri menuju suatu tujuan: model memutuskan tool mana yang dipanggil, membaca hasilnya, dan mengulanginya sampai menghasilkan jawaban akhir atau mencapai batas. Skrip tetap yang memanggil model di titik-titik tertentu biasanya disebut workflow. - [Kunci API (API key)](https://broadrouter.com/id/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): Kunci API adalah string rahasia yang mengidentifikasi dan mengotorisasi client yang memanggil API. API LLM biasanya mengharapkannya di header Authorization: Bearer, dan siapa pun yang memegang kunci itu bisa membuat permintaan yang ditagihkan ke akun pemiliknya. - [Harakat bahasa Arab (tashkeel)](https://broadrouter.com/id/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): Harakat bahasa Arab, atau tashkeel, adalah tanda kecil di atas atau di bawah huruf yang menunjukkan vokal pendek, konsonan ganda, dan ketiadaan vokal. Tanda ini opsional: sebagian besar teks Arab, termasuk berita, tidak memakainya, sehingga pembaca dan perangkat lunak menebak vokal dari konteks. - [Identifikasi dialek Arab (dialect identification)](https://broadrouter.com/id/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): Identifikasi dialek Arab adalah tugas menentukan varian bahasa Arab yang dipakai dalam teks atau rekaman, seperti Mesir, Teluk, Syam (Levant), Maghribi, atau Arab Standar Modern, di tingkat wilayah, negara, atau kota. Tugas ini dipakai untuk merutekan pesan, memilih prompt, dan membagi set evaluasi. - [Normalisasi teks Arab (Arabic text normalization)](https://broadrouter.com/id/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): Normalisasi teks Arab adalah langkah yang memetakan varian ejaan dari kata yang sama ke satu bentuk sebelum pencocokan: menghapus harakat dan tatweel, menyatukan bentuk alif, dan sering memetakan ى ke ي serta ة ke ه. Langkah ini diterapkan pada kunci pencarian dan evaluasi, bukan pada teks yang ditampilkan. - [Arabizi (Franco-Arabic)](https://broadrouter.com/id/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): Arabizi adalah bahasa Arab yang ditulis secara informal dengan huruf Latin, memakai angka untuk bunyi Arab yang tidak punya padanan Latin, misalnya 3 untuk ع. Arabizi umum di chat dan media sosial, kebanyakan berupa dialek, tidak punya ejaan baku, dan sering dicampur dengan bahasa Inggris atau Prancis. - [AWQ (activation-aware weight quantization)](https://broadrouter.com/id/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (activation-aware weight quantization) adalah metode pasca-pelatihan yang mengompresi bobot LLM ke presisi rendah, biasanya integer 4-bit. AWQ memakai statistik aktivasi dari set kalibrasi kecil untuk menemukan kanal bobot yang paling penting dan menskalakannya agar akurasinya tidak banyak hilang. Metode ini tidak membutuhkan backpropagation. - [Base URL](https://broadrouter.com/id/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): Base URL adalah alamat root tetap sebuah API, seperti https://api.example.com/v1, yang ditempatkan client di depan setiap path endpoint, seperti /chat/completions. Di SDK yang kompatibel dengan OpenAI, mengganti base URL bersama kunci API dan ID model akan mengarahkan kode yang sama ke penyedia lain. - [Batch API](https://broadrouter.com/id/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Batch API adalah endpoint LLM asinkron yang menerima banyak request sekaligus, biasanya sebagai file, memprosesnya dalam jangka waktu tertentu alih-alih seketika, lalu mengembalikan hasilnya bersama-sama. Penyedia seperti OpenAI dan Anthropic memberi harga request batch dengan diskon dibanding API standar mereka. - [BM25](https://broadrouter.com/id/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) adalah fungsi ranking untuk pencarian kata kunci yang menilai dokumen berdasarkan seberapa sering istilah kueri muncul di dalamnya, seberapa langka istilah itu di seluruh koleksi, dan panjang dokumennya. BM25 adalah similarity default di Lucene dan Elasticsearch. - [Byte pair encoding (BPE)](https://broadrouter.com/id/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Byte pair encoding (BPE) adalah algoritma tokenisasi subkata yang dimulai dari karakter atau byte tunggal lalu berulang kali menggabungkan pasangan bersebelahan yang paling sering muncul menjadi entri kosakata baru, hingga mencapai ukuran kosakata target. Sebagian besar tokenizer LLM saat ini memakai BPE atau varian byte-level-nya. - [Token input cache (cached input tokens)](https://broadrouter.com/id/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): Token input cache adalah token prompt yang baru saja diproses penyedia LLM dan bisa dipakai ulang dari cache alih-alih dihitung lagi, biasanya prefiks berulang seperti system prompt atau dokumen. Penyedia menagihnya dengan tarif lebih rendah daripada token input baru. - [Chain-of-thought (CoT)](https://broadrouter.com/id/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Chain-of-thought (CoT) adalah teknik prompting di mana model bahasa diberi contoh atau diminta menuliskan langkah-langkah penalaran perantara sebelum jawaban akhirnya. Teknik ini meningkatkan akurasi pada soal multi-langkah seperti aritmetika dan logika, dan model penalaran kini menghasilkannya secara otomatis. - [Character error rate (CER)](https://broadrouter.com/id/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Character error rate (CER) mengukur akurasi transkripsi atau OCR di tingkat karakter: jumlah substitusi, penghapusan, dan penyisipan yang dibutuhkan untuk mengubah output menjadi teks referensi, dibagi jumlah karakter dalam referensi. Makin rendah makin baik, dan 0 berarti cocok sempurna. - [Chat Completions API](https://broadrouter.com/id/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Chat Completions API adalah endpoint HTTP milik OpenAI, POST /v1/chat/completions, yang menerima nama model dan daftar pesan bertanda peran lalu mengembalikan pesan berikutnya dari model. Bentuk request dan response-nya menjadi format umum yang dipakai ulang oleh penyedia dan SDK yang kompatibel dengan OpenAI. - [Chat template](https://broadrouter.com/id/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): Chat template adalah aturan format, biasanya ditulis dengan Jinja dan disertakan bersama tokenizer model, yang mengubah daftar pesan chat beserta perannya menjadi satu urutan token yang diharapkan model, termasuk token kontrol khususnya. Memakai template yang salah bisa sangat menurunkan kualitas jawaban model. - [Chunking (RAG)](https://broadrouter.com/id/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Chunking adalah langkah dalam pipeline retrieval yang memecah dokumen menjadi potongan lebih kecil, disebut chunk, sebelum di-embed dan diindeks. Chunk size menentukan berapa banyak teks di setiap potongan, dan chunk overlap mengulang sebagian teks di antara chunk yang berdekatan agar gagasan yang terpotong di batas tidak hilang. - [Alih kode (code-switching)](https://broadrouter.com/id/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Alih kode (code-switching) adalah penggunaan dua bahasa atau lebih, atau beberapa varian dari satu bahasa, dalam percakapan atau kalimat yang sama, baik lisan maupun tulisan. Contohnya penutur bahasa Arab yang mencampur kata bahasa Inggris ke dalam dialek, atau berpindah antara Arab Standar Modern dan dialek mereka. - [Agen coding AI (AI coding agent)](https://broadrouter.com/id/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): Agen coding AI adalah alat perangkat lunak di mana model bahasa bekerja pada codebase dalam sebuah loop: membaca file, mengeditnya, menjalankan perintah seperti tes, membaca output, dan memutuskan langkah berikutnya sampai tugas selesai atau batas tercapai. - [Konkurensi (concurrency)](https://broadrouter.com/id/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): Konkurensi (concurrency) adalah jumlah permintaan yang sedang berjalan dari sebuah client pada saat yang sama. API LLM sering membatasinya per kunci atau per paket, terpisah dari batas permintaan per menit, karena setiap permintaan yang terbuka menahan kapasitas GPU sampai responsnya selesai. - [Context compaction (pemadatan konteks)](https://broadrouter.com/id/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): Context compaction adalah praktik meringkas percakapan yang mendekati batas context window lalu melanjutkan sesi dari ringkasan itu, sehingga agen tetap menyimpan keputusan dan status penting sambil membuang riwayat mentah seperti output tool lama. - [Context engineering (rekayasa konteks)](https://broadrouter.com/id/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): Context engineering adalah pekerjaan memilih dan merawat kumpulan token yang dilihat model bahasa di setiap request, termasuk system prompt, definisi tool, riwayat, dokumen yang diambil, dan catatan, agar model mendapat informasi bersinyal tinggi sesedikit mungkin yang ia butuhkan untuk mengerjakan tugas dengan baik. - [Context rot (degradasi konteks)](https://broadrouter.com/id/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): Context rot adalah hilangnya akurasi dan keandalan secara bertahap yang ditunjukkan model bahasa saat inputnya makin panjang, bahkan ketika tugasnya tetap sama dan input masih jauh di dalam context window. Istilah ini dipopulerkan oleh laporan teknis Chroma tahun 2025 yang menguji 18 model. - [Context window (jendela konteks)](https://broadrouter.com/id/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): Context window adalah jumlah token maksimum yang bisa diolah model bahasa dalam satu permintaan. Cakupannya meliputi prompt, riwayat percakapan, definisi tool, gambar, dan output yang dihasilkan, termasuk token penalaran. Apa pun yang tidak muat harus dipotong, diringkas, atau dibagi ke beberapa permintaan. - [Data residency (residensi data)](https://broadrouter.com/id/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): Data residency adalah persyaratan atau komitmen bahwa data disimpan dan diproses di negara atau wilayah tertentu. Pada API LLM, ini mencakup di mana prompt diproses selama inferensi dan di mana konten yang disimpan berada, dan sering kali tidak mencakup data akun dan penagihan. - [Dialek bahasa Arab (dialectal Arabic)](https://broadrouter.com/id/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): Dialek bahasa Arab adalah bahasa Arab lisan sehari-hari di suatu wilayah atau kota, seperti bahasa Arab Mesir, Teluk, Levant, atau Maghribi. Dialek berbeda dari Bahasa Arab Standar Modern dalam bunyi, kosakata, dan tata bahasa, tidak punya ejaan baku, dan makin sering ditulis di internet, sehingga lebih sulit bagi model bahasa. - [Speaker diarization (diarisasi pembicara)](https://broadrouter.com/id/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): Speaker diarization adalah tugas membagi rekaman audio menjadi segmen per pembicara untuk menjawab "siapa berbicara kapan". Setiap potongan ucapan diberi label ID pembicara anonim seperti SPEAKER_00, dan biasanya digabungkan dengan transkrip untuk menunjukkan siapa yang mengucapkan setiap baris. - [Distilasi (knowledge distillation)](https://broadrouter.com/id/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): Distilasi, atau knowledge distillation, adalah metode pelatihan di mana model murid yang lebih kecil belajar mereproduksi output model guru yang lebih besar, seperti distribusi probabilitas atau jawaban yang dihasilkannya, sehingga model murid mempertahankan sebagian besar kualitas guru dengan biaya menjalankan yang lebih rendah. - [Data processing agreement (DPA)](https://broadrouter.com/id/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): Data processing agreement (DPA, perjanjian pemrosesan data) adalah kontrak antara pengendali dan prosesor yang mengatur cara prosesor menangani data pribadi. GDPR Pasal 28(3) mewajibkannya dan merinci isinya: instruksi terdokumentasi, kerahasiaan, keamanan, aturan subprosesor, penghapusan atau pengembalian data, serta audit. - [Embeddings (vektor embedding)](https://broadrouter.com/id/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): Embeddings adalah daftar angka floating-point (vektor) yang dihasilkan model untuk merepresentasikan sepotong teks, gambar, atau input lain. Input dengan makna serupa mendapat vektor yang berdekatan, sehingga jarak antarvektor bisa dipakai untuk mengurutkan hasil pencarian, mengelompokkan dokumen, atau menemukan item terkait. - [Eval (evaluasi LLM)](https://broadrouter.com/id/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Eval (singkatan dari evaluation atau evaluasi) adalah pengujian yang dapat diulang terhadap model bahasa atau prompt: satu set input tetap, deskripsi seperti apa output yang baik, dan penilai (kode, manusia, atau model lain) yang memberi skor pada setiap output agar berbagai versi bisa dibandingkan. - [Exponential backoff](https://broadrouter.com/id/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): Exponential backoff adalah strategi percobaan ulang di mana client menunggu makin lama setelah setiap percobaan yang gagal, biasanya menggandakan jedanya sampai batas tertentu, sebelum mencoba lagi. Menambahkan jitter acak pada setiap jeda menyebarkan percobaan ulang dari banyak client agar tidak menghantam server pada saat yang sama. - [Few-shot prompting](https://broadrouter.com/id/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): Few-shot prompting adalah teknik di mana prompt berisi sejumlah kecil contoh input-output yang sudah dikerjakan sebelum input sebenarnya, sehingga model bahasa menyimpulkan tugas dan formatnya dari contoh itu. Bobot model tidak berubah; zero-shot tidak memakai contoh dan one-shot memakai satu. - [Fine-tuning](https://broadrouter.com/id/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): Fine-tuning adalah proses melanjutkan pelatihan model bahasa yang sudah dilatih sebelumnya dengan dataset yang lebih kecil dan khusus tugas, seperti contoh percakapan, sehingga bobotnya berubah dan model menjalankan tugas atau gaya itu dengan lebih andal tanpa perlu instruksi panjang atau contoh di setiap prompt. - [finish_reason (alasan berhenti)](https://broadrouter.com/id/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason adalah field di setiap choice dalam respons chat completion yang menyebutkan alasan model berhenti menghasilkan teks. Dalam format OpenAI nilainya stop, length, tool_calls, content_filter, dan function_call yang sudah usang. Messages API milik Anthropic melaporkan hal yang sama sebagai stop_reason. - [Paket LLM harga tetap (flat-rate)](https://broadrouter.com/id/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): Paket LLM harga tetap menagih harga tetap untuk satu periode, misalnya sebulan, untuk akses ke API model bahasa, alih-alih menagih per token. Paket seperti ini biasanya membatasi hal lain, seperti konkurensi, kecepatan, model, atau jendela pemakaian, jadi batasnya sama pentingnya dengan harganya. - [FP8](https://broadrouter.com/id/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 adalah format bilangan floating-point 8-bit yang dipakai dalam deep learning, didefinisikan dalam dua encoding: E4M3, dengan 4 bit eksponen dan 3 bit mantissa, dan E5M2, dengan 5 bit eksponen dan 2 bit mantissa. FP8 memangkas memori menjadi separuh dibanding format 16-bit dan berjalan native di GPU terbaru. - [Frequency penalty dan presence penalty](https://broadrouter.com/id/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): Frequency penalty dan presence penalty adalah pengaturan sampling yang menurunkan skor token yang sudah ada di output. Presence penalty berlaku begitu token pernah muncul; frequency penalty makin besar setiap kali token diulang. Di API OpenAI, keduanya berkisar dari -2.0 hingga 2.0, dengan default 0. - [GDPR](https://broadrouter.com/id/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): GDPR (General Data Protection Regulation, Regulasi (UE) 2016/679) adalah hukum UE tentang pemrosesan data pribadi. GDPR berlaku sejak 25 Mei 2018 dan mencakup organisasi yang berkedudukan di UE serta sebagian organisasi di luar UE. Untuk API LLM, GDPR mengatur kontrak prosesor, transfer internasional, dan minimisasi data. - [GGUF](https://broadrouter.com/id/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF adalah format file biner untuk menyimpan model bahasa yang akan dijalankan inferensinya dengan runtime berbasis GGML seperti llama.cpp. Satu file GGUF berisi bobot, sering kali terkuantisasi, ditambah metadata yang dibutuhkan untuk memuat model, termasuk detail tokenizer dan arsitektur, dan bisa di-memory-map agar cepat dimuat. - [GPU-hour (jam GPU)](https://broadrouter.com/id/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): GPU-hour (jam GPU) adalah satu prosesor grafis yang dipakai selama satu jam. Penyedia cloud menagih instance GPU sewaan per jam atau per detik, dan peneliti memakai GPU-hour untuk menyatakan berapa banyak komputasi yang dihabiskan sebuah pekerjaan. Anda membayar waktu mesin, entah mesin itu sibuk atau tidak. - [Grounding](https://broadrouter.com/id/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): Grounding adalah praktik menghubungkan output model bahasa ke sumber informasi yang spesifik dan dapat diverifikasi, seperti dokumen yang diberikan, hasil pencarian, atau catatan database, dan menginstruksikan model untuk menjawab berdasarkan sumber itu. Jawaban yang di-grounding bisa ditelusuri ke sumbernya dan lebih kecil kemungkinannya berisi fakta karangan. - [Guardrails (pagar pengaman AI)](https://broadrouter.com/id/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): Guardrails adalah pemeriksaan yang dijalankan aplikasi di sekitar panggilan model bahasa agar tetap dalam batas yang ditetapkan. Guardrail input menyaring apa yang masuk ke model, guardrail output memeriksa apa yang keluar, dan guardrail tool memvalidasi aksi sebelum dijalankan. Pemeriksaan yang gagal akan memblokir, menulis ulang, atau menandai request. - [Halusinasi AI (hallucination)](https://broadrouter.com/id/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): Halusinasi adalah output model bahasa yang terdengar masuk akal tetapi salah, dikarang, atau tidak didukung oleh input yang diberikan, misalnya kutipan sumber fiktif, angka yang keliru, atau API yang tidak ada. Halusinasi muncul karena model memprediksi teks yang paling mungkin, bukan memeriksa fakta. - [Hamzah (hamza)](https://broadrouter.com/id/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): Hamzah (ء) adalah huruf Arab untuk hentian glotal. Hamzah ditulis sendiri atau di atas sebuah dudukan: alif (أ إ), wawu (ؤ), atau ya (ئ), sedangkan آ menggabungkan alif dengan madd. Karena teks ketikan tidak konsisten soal hamzah, mesin pencari dan penilaian ucapan sering melebur bentuk-bentuk alif menjadi ا. - [Hybrid search (pencarian hibrida)](https://broadrouter.com/id/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): Hybrid search adalah metode retrieval yang menjalankan pencarian kata kunci, seperti BM25, dan pencarian vektor atas embedding pada query yang sama, lalu menggabungkan kedua daftar peringkat menjadi satu, sering dengan reciprocal rank fusion (RRF) atau jumlah berbobot dari skor yang dinormalisasi. - [Idempotency key](https://broadrouter.com/id/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Idempotency key adalah nilai unik, biasanya ID acak seperti UUID, yang dilampirkan klien pada request yang mengubah state, misalnya pembayaran atau pembelian. Jika request yang sama dikirim lagi dengan key yang sama, server mengembalikan hasil aslinya alih-alih menjalankan aksi itu dua kali. - [Token gambar (image tokens)](https://broadrouter.com/id/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): Token gambar adalah token input yang dipakai vision language model untuk merepresentasikan gambar dalam prompt. Penyedia menghitungnya dari ukuran gambar, biasanya dengan membaginya menjadi patch atau tile berukuran tetap setelah diubah ukurannya, lalu menagihnya seperti token input teks, sehingga gambar lebih besar lebih mahal. - [Penyedia inferensi (inference provider)](https://broadrouter.com/id/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): Penyedia inferensi adalah perusahaan yang menjalankan model machine learning, sering kali LLM open-weight, di GPU yang dimiliki, disewa jangka panjang, atau disewa per pemakaian, lalu menjual akses ke model itu lewat API, biasanya dihargai per token, per permintaan, atau per satuan waktu. Penyedia ini menangani penyajian, penskalaan, dan perangkat keras. - [Token input dan output](https://broadrouter.com/id/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): Token input adalah token yang dibaca model bahasa dalam sebuah permintaan, termasuk system prompt, riwayat percakapan, definisi tool, dan gambar. Token output adalah token yang dihasilkannya sebagai balasan, termasuk penalaran. API LLM biasanya menghargainya secara terpisah, dan output biasanya lebih mahal. - [INT4](https://broadrouter.com/id/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 adalah format bilangan integer 4-bit yang dapat merepresentasikan 16 level. Dalam inferensi LLM, istilah ini biasanya berarti kuantisasi 4-bit khusus bobot (W4A16): bobot disimpan sebagai INT4 dengan faktor skala dan diperluas ke 16-bit saat komputasi, sehingga memakai sekitar seperempat memori bobot 16-bit. - [IP allow list (daftar IP yang diizinkan)](https://broadrouter.com/id/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): IP allow list adalah kontrol keamanan yang hanya menerima request dari alamat IP atau rentang CIDR yang terdaftar dan menolak semua yang lain. Pada API, kontrol ini membatasi kerusakan akibat kunci yang bocor, karena kunci tidak berfungsi dari jaringan yang tidak ada di daftar. - [Jailbreak (LLM)](https://broadrouter.com/id/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): Jailbreak adalah input yang dirancang agar model bahasa mengabaikan pelatihan keamanannya atau aturan yang ditetapkan aplikasi yang memakainya, sehingga model menghasilkan konten atau melakukan aksi yang biasanya akan ditolak. OWASP memperlakukan jailbreaking sebagai salah satu bentuk prompt injection. - [JSON mode (mode JSON)](https://broadrouter.com/id/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): JSON mode adalah pengaturan API LLM, yaitu response_format {"type": "json_object"} di API bergaya OpenAI, yang membuat balasan model berupa JSON yang valid dan bisa di-parse. Berbeda dengan structured outputs yang memakai JSON Schema, mode ini tidak menjamin key, tipe, atau nilai apa yang ada di objek. - [JSON Schema (skema JSON)](https://broadrouter.com/id/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema adalah kosakata yang ditulis dalam JSON untuk mendeskripsikan dan memvalidasi struktur data JSON: tipe, field wajib, nilai yang diizinkan, dan susunan bertingkat. Versi terbarunya 2020-12. API LLM memakainya untuk mendefinisikan parameter tool dan bentuk persis yang harus diikuti output terstruktur. - [KV cache](https://broadrouter.com/id/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): KV cache adalah memori tempat model bahasa transformer menyimpan key dan value attention yang dihitung untuk setiap token yang sudah diproses, sehingga setiap token baru memakainya ulang alih-alih menghitung ulang seluruh urutan. Ukurannya bertambah seiring panjang konteks dan sering membatasi berapa banyak request yang bisa dilayani satu GPU. - [Latensi (latency)](https://broadrouter.com/id/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): Latensi, untuk API model bahasa, adalah waktu yang dibutuhkan sebuah permintaan sejak dikirim sampai token output terakhir diterima. Latensi adalah jumlah dari time to first token dan waktu generasi, dan biasanya dilaporkan sebagai persentil seperti p50 dan p95, bukan rata-rata. - [LLM-as-a-judge (LLM sebagai juri)](https://broadrouter.com/id/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge adalah metode evaluasi di mana model bahasa menilai output model lain, baik dengan memberi skor satu jawaban berdasarkan rubrik maupun memilih yang lebih baik dari dua jawaban. Metode ini mendekati penilaian manusia dengan biaya jauh lebih rendah, tetapi punya bias yang diketahui, seperti condong ke jawaban pertama atau lebih panjang. - [Benchmark LLM (LLM benchmark)](https://broadrouter.com/id/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): Benchmark LLM adalah sekumpulan tugas yang tetap dan biasanya publik, lengkap dengan metode penilaian, untuk mengukur dan membandingkan model bahasa pada satu keterampilan, seperti soal pengetahuan, coding, atau pengenalan ucapan. Skor benchmark menunjukkan kemampuan relatif pada uji itu, bukan performa model pada workload Anda sendiri. - [LLM router](https://broadrouter.com/id/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): LLM router adalah software atau layanan yang berada di antara aplikasi dan satu atau lebih API model bahasa, lalu memutuskan untuk setiap request model atau penyedia mana yang menanganinya, berdasarkan aturan seperti biaya, latensi, ketersediaan, atau jenis tugas. - [Logprobs (log probabilitas)](https://broadrouter.com/id/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Logprobs adalah log probabilitas yang diberikan model bahasa untuk token-token yang dihasilkannya. API yang mengembalikannya menunjukkan, untuk setiap token output, seberapa mungkin token itu menurut model dan, secara opsional, token alternatif paling mungkin di posisi tersebut. - [LoRA (low-rank adaptation)](https://broadrouter.com/id/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation) adalah metode fine-tuning yang hemat parameter: bobot model pra-latih dibekukan, lalu sepasang matriks low-rank kecil dilatih untuk layer tertentu. Pembaruan yang dipelajari bisa disimpan sebagai adapter kecil atau digabungkan ke bobot, tanpa menambah latensi inferensi. - [Lost in the middle](https://broadrouter.com/id/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle adalah kecenderungan model bahasa yang teramati untuk paling baik memakai informasi relevan saat informasi itu muncul di dekat awal atau akhir input panjang, dan paling buruk saat berada di tengah. Istilah ini berasal dari studi Liu dkk. tahun 2023. - [max_tokens](https://broadrouter.com/id/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens adalah parameter API LLM yang menetapkan jumlah token maksimum yang boleh dihasilkan model dalam satu respons. Saat generasi mencapai batas itu, proses berhenti dan respons ditandai terpotong, misalnya dengan finish_reason bernilai "length". - [MCP (Model Context Protocol)](https://broadrouter.com/id/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol) adalah protokol terbuka berbasis JSON-RPC 2.0 yang menstandarkan cara aplikasi LLM terhubung ke tool dan data eksternal. Server MCP menyediakan tools, resources, dan prompts; client MCP di dalam aplikasi host menemukannya dan meneruskan tool call dari model ke server. - [Mixture of experts (MoE)](https://broadrouter.com/id/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Mixture of experts (MoE) adalah desain jaringan saraf di mana satu layer berisi banyak sub-jaringan expert dan sebuah router yang dilatih mengirim setiap token hanya ke beberapa di antaranya. Model bahasa MoE punya jumlah parameter total yang besar, tetapi hanya memakai subset aktif yang kecil untuk setiap token. - [Deprecation model (model deprecation)](https://broadrouter.com/id/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): Deprecation model adalah ketika penyedia API mengumumkan bahwa sebuah model tidak akan ditawarkan lagi, biasanya dengan menyebut tanggal pensiun atau shutdown dan model pengganti yang direkomendasikan. Setelah tanggal itu, request ke model tersebut gagal, sehingga aplikasi harus bermigrasi ke ID model lain sebelumnya. - [Model fallback](https://broadrouter.com/id/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): Model fallback adalah pola keandalan di mana request LLM yang gagal dicoba ulang pada model atau penyedia lain dari sebuah daftar berurutan. Pemicunya bisa berupa error, rate limit, error panjang konteks, atau penolakan kebijakan konten. Pola ini berjalan di klien Anda atau di gateway di depan API. - [Lisensi model (model license)](https://broadrouter.com/id/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): Lisensi model adalah ketentuan hukum yang mengatur bagaimana bobot, kode, dan output sebuah model boleh dipakai, diubah, dan dibagikan. Model open-weight memakai lisensi perangkat lunak standar seperti Apache 2.0 atau MIT, atau lisensi model khusus yang bisa menambahkan syarat, misalnya batasan bagi layanan komersial besar. - [Bahasa Arab Standar Modern (MSA, Fusha)](https://broadrouter.com/id/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): Bahasa Arab Standar Modern (MSA) adalah bentuk baku bahasa Arab yang dipakai dalam tulisan, berita, pendidikan, dan pidato formal di seluruh dunia Arab. MSA bukan bahasa ibu sehari-hari siapa pun: orang berbicara dalam dialek daerah seperti Arab Mesir atau Teluk dan beralih ke MSA dalam situasi formal. - [Sistem multi-agen (multi-agent system)](https://broadrouter.com/id/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): Sistem multi-agen adalah aplikasi LLM di mana beberapa agen, masing-masing dengan prompt, tool, dan konteksnya sendiri, mengerjakan bagian-bagian dari satu tugas. Desain yang umum adalah orchestrator-worker: agen utama membuat rencana, menjalankan agen worker secara paralel, lalu menggabungkan hasilnya. - [AI multimodal (multimodal AI)](https://broadrouter.com/id/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): AI multimodal adalah model atau sistem yang bekerja dengan lebih dari satu jenis data, seperti teks, gambar, video, dan audio. Model multimodal bisa membaca foto dan pertanyaan sekaligus, atau menerima ucapan lalu menjawab dalam teks. Jenis yang paling umum adalah vision-language model. - [Named entity recognition (NER, pengenalan entitas bernama)](https://broadrouter.com/id/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): Named entity recognition (NER) adalah tugas menemukan nama dalam teks dan memberi label jenis pada masing-masing, seperti orang, organisasi, atau lokasi. Sistem klasik memberi tag pada setiap token; LLM biasanya diminta mengembalikan daftar entitas dalam JSON, yang lalu dicek terhadap teks sumber. - [Optical character recognition (OCR)](https://broadrouter.com/id/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): Optical character recognition (OCR) adalah teknologi yang mendeteksi dan mengubah teks cetak atau tulisan tangan di dalam gambar, hasil scan, atau PDF menjadi karakter yang bisa dibaca mesin. OCR klasik menghasilkan teks dan posisinya di halaman; vision-language model bisa membaca teks dan langsung menjawab pertanyaan tentangnya. - [Ogg Opus](https://broadrouter.com/id/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus adalah audio yang dienkode dengan codec Opus (RFC 6716) dan disimpan dalam container Ogg (RFC 7845). Format ini dirancang untuk ucapan dan audio interaktif pada bitrate rendah, dan dipakai WhatsApp untuk pesan suara, biasanya sebagai file .ogg atau .opus. - [Model open-weight (bobot terbuka)](https://broadrouter.com/id/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): Model open-weight adalah model machine learning yang bobot hasil pelatihannya dipublikasikan agar bisa diunduh dan dijalankan siapa saja, biasanya di bawah lisensi dengan syarat tertentu. Model ini berbeda dari AI yang sepenuhnya open source, yang juga merilis kode pelatihan dan informasi rinci tentang data pelatihan agar orang lain bisa mempelajari dan mereproduksinya. - [API kompatibel OpenAI](https://broadrouter.com/id/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): API kompatibel OpenAI adalah API HTTP yang menerima format permintaan dan respons yang sama dengan API OpenAI, seperti POST /v1/chat/completions, sehingga SDK dan tool OpenAI bisa memakainya hanya dengan mengganti base URL, kunci API, dan nama model. - [Parallel tool calls (pemanggilan tool paralel)](https://broadrouter.com/id/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Parallel tool calls adalah saat model mengembalikan lebih dari satu tool call dalam satu respons, misalnya membaca tiga file sekaligus. Dalam format OpenAI, panggilan ini datang sebagai beberapa entri di tool_calls, dan parallel_tool_calls bernilai false membatasi model ke nol atau satu panggilan per giliran. - [Bayar sesuai pemakaian (pay-as-you-go)](https://broadrouter.com/id/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): Bayar sesuai pemakaian (pay-as-you-go) adalah model harga di mana Anda hanya membayar yang Anda pakai, dihitung per satuan seperti token, permintaan, atau menit audio, tanpa langganan atau komitmen minimum. Sebagian besar API LLM menagih dengan cara ini, dari saldo prabayar atau tagihan berkala. - [PDPL (UU perlindungan data Arab Saudi dan UEA)](https://broadrouter.com/id/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) adalah nama umum untuk UU data pribadi Arab Saudi, yang diterbitkan dengan Dekret Kerajaan M/19 dan ditegakkan oleh SDAIA, serta untuk Dekret-Undang-Undang Federal UEA No. 45 Tahun 2021. Keduanya mengatur pemrosesan data pribadi dan menetapkan syarat transfer ke luar negeri, termasuk prompt yang dikirim ke API LLM asing. - [PII (personally identifiable information)](https://broadrouter.com/id/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): PII (personally identifiable information) adalah informasi yang bisa dipakai untuk membedakan atau melacak identitas seseorang, seperti nama, nomor identitas, atau data biometrik, ditambah informasi lain yang terkait atau bisa dikaitkan dengan orang tersebut. Istilah GDPR yang lebih luas adalah data pribadi. Dalam kerja LLM, redaksi PII menghapusnya dari prompt sebelum keluar dari sistem Anda. - [Prefill dan decode (fase inferensi LLM)](https://broadrouter.com/id/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill dan decode adalah dua fase inferensi LLM. Pada prefill, model memproses seluruh prompt secara paralel dan mengisi KV cache-nya; pada decode, model menghasilkan balasan satu token demi satu token sambil memakai ulang cache tersebut. Prefill terutama menentukan time to first token; decode menentukan kecepatan output. - [Kredit prabayar (prepaid credit)](https://broadrouter.com/id/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): Kredit prabayar adalah uang yang diisi ke akun API sebelum dipakai. Setiap request memotong biayanya dari saldo, dan saat saldo habis, request berbayar per pemakaian berhenti sampai Anda melakukan top up. Dengan begitu, pengeluaran dibatasi sebesar uang yang sudah Anda setor. - [Prompt engineering (rekayasa prompt)](https://broadrouter.com/id/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): Prompt engineering adalah proses menulis, menyusun, dan menguji instruksi yang diberikan ke model bahasa agar model konsisten menghasilkan output yang memenuhi kebutuhan Anda. Cakupannya meliputi pesan peran, instruksi yang jelas, contoh, format, dan pengukuran hasil terhadap kasus uji. - [Prompt injection (injeksi prompt)](https://broadrouter.com/id/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): Prompt injection adalah serangan di mana teks yang diberikan ke model bahasa, oleh pengguna atau tersembunyi di halaman web, email, file, atau hasil tool, mengubah perilaku model berlawanan dengan maksud developer. OWASP menempatkannya di urutan pertama (LLM01:2025) dalam Top 10 for LLM Applications. - [Kuantisasi (quantization)](https://broadrouter.com/id/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): Kuantisasi adalah proses menyimpan angka-angka jaringan saraf, terutama bobotnya, dalam format berpresisi lebih rendah seperti 8-bit atau 4-bit alih-alih 16-bit, sehingga model memakai lebih sedikit memori dan sering berjalan lebih cepat. Metode yang baik menjaga penurunan akurasi tetap kecil. - [RAG (retrieval-augmented generation)](https://broadrouter.com/id/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG (retrieval-augmented generation) adalah pola di mana aplikasi lebih dulu mengambil bagian teks yang relevan dengan pertanyaan dari penyimpanan eksternal, seperti indeks pencarian atau indeks vektor, lalu menambahkannya ke prompt agar model bahasa bisa menjawab berdasarkan teks itu, bukan hanya dari ingatannya. - [Rate limit (batas laju)](https://broadrouter.com/id/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): Rate limit adalah batas yang diterapkan API pada seberapa banyak client boleh memakainya dalam rentang waktu tertentu, misalnya permintaan per menit atau token per menit. Permintaan yang melebihi batas ditolak, biasanya dengan HTTP 429 Too Many Requests, sampai pemakaian kembali di bawah batas. - [ReAct (reason and act)](https://broadrouter.com/id/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (reason and act) adalah pola prompting yang diperkenalkan Yao dkk. pada 2022, di mana model bahasa menyelingi jejak penalaran dengan aksi seperti pencarian atau tool call. Model membaca setiap observasi, memperbarui rencananya, dan mengulanginya sampai bisa memberikan jawaban akhir. - [Token penalaran (reasoning tokens)](https://broadrouter.com/id/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): Token penalaran adalah token yang dihasilkan model penalaran untuk mengerjakan suatu masalah sebelum menulis jawaban yang terlihat. Token ini biasanya ditagih sebagai token output, dihitung dalam batas output dan context window, dan bisa disembunyikan atau diringkas oleh API. - [Request timeout (batas waktu permintaan)](https://broadrouter.com/id/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): Request timeout adalah waktu terlama sebuah client HTTP menunggu server sebelum membatalkan permintaan. Client biasanya membaginya menjadi connect timeout untuk membuka koneksi dan read timeout untuk menunggu data. Panggilan LLM memerlukan read timeout yang longgar karena jawaban panjang butuh waktu. - [Reranking (pemeringkatan ulang)](https://broadrouter.com/id/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Reranking adalah tahap retrieval kedua di mana model yang lebih akurat tetapi lebih lambat, biasanya cross-encoder, menilai setiap pasangan query dan dokumen kandidat dari pencarian tahap pertama lalu mengurutkannya ulang, sehingga hanya beberapa yang paling relevan diteruskan ke model bahasa. - [Teks right-to-left (RTL)](https://broadrouter.com/id/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): Teks right-to-left (RTL) adalah tulisan, seperti bahasa Arab atau Ibrani, yang ditampilkan dari kanan ke kiri. Komputer menyimpannya dalam urutan baca logis, dan Unicode Bidirectional Algorithm menentukan urutan tampilan saat teks RTL bercampur dengan teks kiri-ke-kanan, angka, nama model, atau kode. - [Sample rate (laju sampel)](https://broadrouter.com/id/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): Sample rate adalah jumlah sampel audio yang diambil per detik, diukur dalam hertz (Hz). Model pengenalan ucapan umumnya mengharapkan audio mono 16.000 Hz (16 kHz). Bersama bit depth dan jumlah kanal, sample rate menentukan ukuran file audio tanpa kompresi per menit. - [Parameter seed](https://broadrouter.com/id/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): Parameter seed adalah bilangan bulat yang Anda kirim bersama permintaan agar API berusaha mengambil sampel token yang sama saat permintaan yang sama diulang. OpenAI menyebutnya best effort: determinisme tidak dijamin, dan field system_fingerprint menunjukkan kapan konfigurasi backend berubah. - [Serverless inference (inferensi serverless)](https://broadrouter.com/id/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): Serverless inference adalah cara menjalankan model machine learning di mana platform otomatis memulai, menskalakan, dan menghentikan komputasi di balik sebuah endpoint, sering kali hingga nol saat tidak dipakai, dan menagih berdasarkan pemakaian, bukan mesin yang dipesan. Konsekuensinya adalah kemungkinan jeda cold start. - [SLA (service level agreement)](https://broadrouter.com/id/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): SLA (service level agreement) adalah komitmen kontraktual penyedia atas tingkat layanan yang terukur, biasanya persentase uptime bulanan, dengan kompensasi yang ditetapkan, biasanya kredit layanan, jika target tidak tercapai. SLA menetapkan apa yang dihitung sebagai downtime, apa yang dikecualikan, dan cara mengajukan klaim. - [Speculative decoding](https://broadrouter.com/id/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): Speculative decoding adalah teknik inferensi di mana model draft yang kecil dan cepat mengusulkan beberapa token berikutnya, lalu model target yang besar memverifikasinya dalam satu forward pass dan menyimpan token yang diterima. Teknik ini mempercepat generasi tanpa mengubah distribusi output. - [Speech-to-text (ucapan ke teks)](https://broadrouter.com/id/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): Speech-to-text, juga disebut automatic speech recognition (ASR), adalah teknologi yang mengubah audio ucapan menjadi teks tertulis. API speech-to-text menerima file atau stream audio dan mengembalikan transkrip, biasanya dihargai per menit audio. Akurasinya diukur dengan word error rate dan bervariasi menurut aksen, dialek, dan kualitas rekaman. - [Spend limit (batas pengeluaran)](https://broadrouter.com/id/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): Spend limit adalah batas jumlah yang boleh dibelanjakan sebuah akun, proyek, atau key API dalam satu periode, biasanya sebulan. Saat pemakaian mencapai batas, penyedia menolak permintaan berikutnya, sering dengan error HTTP 429 atau 400, sampai periode di-reset atau batasnya dinaikkan. - [Subtitle SRT](https://broadrouter.com/id/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) adalah format subtitle berbasis teks biasa. Setiap takarir terdiri dari nomor urut, waktu mulai dan selesai yang ditulis hh:mm:ss,mmm --> hh:mm:ss,mmm, satu atau lebih baris teks, dan satu baris kosong. WebVTT adalah format web serupa yang dipakai video HTML. - [Server-sent events (SSE)](https://broadrouter.com/id/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Server-sent events (SSE) adalah standar web yang memungkinkan server mengirim aliran event teks ke client melalui satu respons HTTP yang tetap terbuka lama, dengan content type text/event-stream. Setiap event terdiri dari satu atau beberapa baris field, seperti data:, dan diakhiri baris kosong. - [Stop sequences (urutan henti)](https://broadrouter.com/id/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): Stop sequence adalah string yang Anda kirim bersama permintaan agar model berhenti menghasilkan teks begitu ia menulis salah satunya. OpenAI menamai field-nya stop dan menerima hingga 4 urutan; Anthropic menamainya stop_sequences. Teks yang cocok biasanya tidak disertakan dalam output yang dikembalikan. - [Streaming (respons streaming)](https://broadrouter.com/id/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): Streaming adalah cara mengembalikan respons model bahasa dalam potongan-potongan kecil selagi masih dihasilkan, bukan sebagai satu balasan utuh di akhir. API yang kompatibel dengan OpenAI melakukan streaming lewat server-sent events jika permintaan mengatur stream ke true. - [Output terstruktur (structured output)](https://broadrouter.com/id/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): Output terstruktur adalah fitur API LLM yang membatasi balasan model pada format yang ditentukan developer, biasanya JSON Schema, agar kode bisa mem-parse-nya dengan andal. Fitur ini lebih ketat daripada JSON mode, yang menjamin JSON valid tetapi tidak menjamin bentuk tertentu. - [Subagent (sub-agen)](https://broadrouter.com/id/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): Subagent adalah agen terpisah yang dijalankan oleh agen utama untuk menangani satu tugas yang terfokus. Subagent berjalan dengan context window, system prompt, dan akses tool sendiri, mengerjakan tugasnya, lalu hanya mengembalikan hasil atau ringkasan ke agen utama, sehingga percakapan utama tetap ringkas. - [Subprosesor (subprocessor)](https://broadrouter.com/id/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): Subprosesor adalah pihak ketiga yang dilibatkan prosesor untuk memproses data pribadi atas nama pengendali data. Untuk API LLM, subprosesor bisa mencakup host GPU, penyedia inferensi mitra, dan kapasitas cadangan. Menurut GDPR Pasal 28, subprosesor memerlukan otorisasi pengendali dan kewajiban perlindungan data yang sama. - [System prompt (prompt sistem)](https://broadrouter.com/id/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): System prompt adalah teks instruksi yang dikirim aplikasi ke model bahasa sebelum percakapan, biasanya sebagai pesan dengan peran system atau developer. System prompt menetapkan peran, aturan, gaya bahasa, dan format output model, dan model dirancang untuk mengikutinya di atas pesan pengguna yang bertentangan. - [Ta marbutah (taa marbuta, ة)](https://broadrouter.com/id/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): Ta marbutah (ة, U+0629) adalah huruf Arab yang hanya muncul di akhir kata, sering menandai kata benda dan kata sifat feminin. Bentuknya mirip ha (ه, U+0647) dengan dua titik, sehingga keduanya sering tertukar dalam teks ketikan, dan banyak normalizer pencarian memetakan ة ke ه. - [Temperature (suhu sampling)](https://broadrouter.com/id/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): Temperature adalah parameter sampling yang mengatur seberapa acak output model bahasa. Model membagi skor token dengan temperature sebelum mengubahnya menjadi probabilitas: nilai rendah membuat token paling mungkin mendominasi, sedangkan nilai tinggi meratakan distribusi sehingga token yang kurang mungkin lebih sering terpilih. - [Text-to-speech (TTS)](https://broadrouter.com/id/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): Text-to-speech (TTS), atau sintesis ucapan, mengubah teks tertulis menjadi audio ucapan. API TTS menerima teks, sebuah suara, dan format output seperti MP3 atau WAV, lalu mengembalikan ucapan yang dihasilkan. TTS adalah kebalikan dari speech-to-text, yang mengubah audio menjadi teks. - [Throughput](https://broadrouter.com/id/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): Throughput, dalam penyajian LLM, adalah jumlah total pekerjaan yang diselesaikan sebuah sistem per satuan waktu di semua permintaan yang berjalan bersamaan, biasanya diukur dalam token output per detik atau permintaan per detik. Ini berbeda dengan kecepatan per stream, yaitu seberapa cepat satu permintaan menerima token-tokennya. - [Time to first token (TTFT)](https://broadrouter.com/id/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): Time to first token (TTFT) adalah waktu sejak permintaan dikirim ke model bahasa sampai token output pertama tiba. TTFT mencakup waktu jaringan, antrean di server, dan prefill, yaitu tahap saat model membaca seluruh prompt, sehingga nilainya naik seiring panjang prompt dan beban server. - [Timestamp (speech-to-text)](https://broadrouter.com/id/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): Timestamp dalam speech-to-text adalah waktu mulai dan selesai, dalam detik, yang dilekatkan pada bagian-bagian transkrip. Timestamp segmen mencakup satu frasa atau kalimat; timestamp kata mencakup setiap kata. Timestamp dibutuhkan untuk subtitle, pencarian di dalam rekaman, dan menyelaraskan teks dengan pembicara. - [Token](https://broadrouter.com/id/glossary/token/) (translation of https://broadrouter.com/glossary/token/): Token adalah satuan teks yang dibaca dan ditulis model bahasa: sebuah kata, bagian kata, tanda baca, atau urutan byte yang dihasilkan tokenizer model. API LLM menghitung, membatasi, dan biasanya menagih pemakaian dalam token. - [Tokenizer](https://broadrouter.com/id/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): Tokenizer adalah komponen yang mengubah teks menjadi urutan ID token berupa bilangan bulat yang dibaca model bahasa, dan mengubah ID yang dihasilkan kembali menjadi teks. Setiap model punya tokenizer sendiri, jadi teks yang sama bisa menghasilkan jumlah token berbeda di model yang berbeda. - [Token per detik (tokens per second, TPS)](https://broadrouter.com/id/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): Token per detik (TPS) adalah laju model bahasa menghasilkan token output, biasanya diukur untuk satu permintaan setelah token pertamanya tiba. Angka ini menentukan seberapa cepat balasan mengalir setelah dimulai. Jika dijumlahkan di semua permintaan bersamaan, ia menjadi throughput sistem; waktu tunggu sebelum token pertama diukur terpisah. - [Tool calling (function calling)](https://broadrouter.com/id/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): Tool calling, disebut juga function calling, adalah fitur API LLM di mana model mengembalikan permintaan terstruktur untuk menjalankan fungsi yang didefinisikan developer, lengkap dengan argumen JSON. Aplikasi yang menjalankan fungsi itu dan mengirim hasilnya kembali; model tidak pernah menjalankan kode sendiri. - [tool_choice](https://broadrouter.com/id/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice adalah parameter permintaan yang mengatur penggunaan tool. Di Chat Completions OpenAI, nilainya bisa auto (model yang memutuskan), required (wajib memanggil satu tool atau lebih), none (tanpa tool), atau nama fungsi tertentu yang dipaksakan. Opsi padanannya di Anthropic adalah auto, any, tool, dan none. - [Top-p (nucleus sampling)](https://broadrouter.com/id/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Top-p, disebut juga nucleus sampling, adalah metode decoding di mana model bahasa memilih token berikutnya hanya dari himpunan terkecil token paling mungkin yang total probabilitasnya mencapai ambang p, misalnya 0,9, lalu membuang ekor yang kecil kemungkinannya. - [Normalisasi Unicode (Unicode normalization)](https://broadrouter.com/id/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Normalisasi Unicode mengubah teks ke bentuk standar sehingga string yang terlihat sama memakai code point yang sama. Ada empat bentuk: NFD, NFC, NFKD, dan NFKC; bentuk K juga menyatukan karakter kompatibilitas, seperti presentation form dan ligatur Arab, menjadi huruf biasa. - [Voice activity detection (VAD)](https://broadrouter.com/id/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): Voice activity detection (VAD) adalah tahap pemrosesan sinyal yang mengklasifikasikan frame audio pendek sebagai ucapan atau bukan ucapan. Sistem suara memakainya untuk melewati bagian hening, memecah rekaman panjang di jeda, mendeteksi kapan pembicara selesai, dan menghindari transkripsi noise latar. - [Vector database (basis data vektor)](https://broadrouter.com/id/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): Vector database adalah penyimpanan data yang menyimpan vektor embedding, sering kali bersama teks sumber dan metadatanya, dan menjawab query nearest-neighbour: untuk sebuah vektor query, ia mengembalikan vektor tersimpan yang paling mirip, biasanya berdasarkan cosine distance, jarak Euclidean, atau inner product. - [Vision-language model (VLM)](https://broadrouter.com/id/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): Vision-language model (VLM) adalah model yang menerima gambar dan teks sekaligus sebagai input dan menghasilkan teks, sehingga bisa mendeskripsikan gambar, menjawab pertanyaan tentangnya, serta membaca dokumen, grafik, atau screenshot. VLM juga disebut model image-text-to-text atau model multimodal. - [vLLM](https://broadrouter.com/id/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM adalah library open-source untuk inferensi dan serving LLM, awalnya dikembangkan di Sky Computing Lab UC Berkeley. Ia mengelola memori KV cache dengan PagedAttention, mem-batch request yang masuk secara kontinu, mendukung banyak format kuantisasi, dan menyediakan server HTTP yang kompatibel dengan OpenAI. - [Ukuran kosakata (vocabulary size)](https://broadrouter.com/id/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): Vocabulary size (ukuran kosakata) adalah jumlah token berbeda di tokenizer sebuah model bahasa, masing-masing dengan ID dan embedding sendiri. Kosakata yang lebih besar bisa mengodekan teks umum dengan lebih sedikit token, tetapi memperbesar lapisan embedding dan output, dan hanya membantu untuk teks yang dipakai melatih kosakatanya. - [WAV](https://broadrouter.com/id/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format, atau WAVE) adalah container audio yang dibangun di atas format RIFF milik Microsoft dan IBM. Biasanya berisi sampel PCM linear tanpa kompresi, sehingga ukurannya mudah diprediksi: sample rate dikali byte per sampel dikali jumlah kanal, per detik. - [Word error rate (WER)](https://broadrouter.com/id/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): Word error rate (WER) adalah ukuran akurasi standar untuk pengenalan ucapan: jumlah substitusi, penghapusan, dan penyisipan kata yang dibutuhkan untuk mengubah transkrip menjadi referensi yang benar, dibagi jumlah kata dalam referensi. Makin rendah makin baik. - [Zero data retention (ZDR)](https://broadrouter.com/id/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): Zero data retention (ZDR) adalah kebijakan penanganan data di mana penyedia AI memproses prompt dan output hanya untuk melayani permintaan dan tidak menyimpan konten itu sesudahnya, baik di log, cache, maupun data pelatihan. Catatan tanpa konten seperti jumlah token dan durasi biasanya tetap disimpan. ### Filipino (fil, Filipino): 294 translated pages - [Articles in Filipino](https://broadrouter.com/fil/blog/) - [Pagbabago sa presyo ng LLM API, Oktubre 2026: 3 pagtaas, walang bawas](https://broadrouter.com/fil/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 pagbabago sa pricing page ng 5 sa 7 LLM API provider noong Oktubre 2026: 3 pagtaas, walang bawas, 9 bagong presyong model at 11 tinanggal. May archive link. - [Magkano ang isang run ng AI coding agent: hakbang, context, caching](https://broadrouter.com/fil/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): Bakit lumalaki ang gastos kada hakbang habang humahaba ang run ng coding agent, paano binabago ng caching ang total, at kailan mas sulit ang flat session. May math. - [Diacritics sa Arabic text gamit ang LLM: prompt, check at DER](https://broadrouter.com/fil/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): Ibalik ang Arabic diacritics (tashkeel) gamit ang LLM API: prompt na hindi gumagalaw sa letra, check na tumatanggi sa edit, DER at WER, at kailan gamitin ang CATT. - [Arabic grammar at spelling correction gamit ang LLM: paano gawin](https://broadrouter.com/fil/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Itama ang Arabic spelling at grammar gamit ang LLM API: minimal-edit na prompt, bantay laban sa rewrite, word-level na edit ayon sa error class at M2 scoring. - [Arabic meeting minutes mula sa recording: isang API pipeline](https://broadrouter.com/fil/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Gawing minutes na may desisyon at action item ang Arabic meeting recording: i-transcribe, linisin, i-summarise sa JSON gamit ang LLM, suriin ang pangalan at petsa. - [Arabic named entity extraction gamit ang LLM: JSON, petsa at check](https://broadrouter.com/fil/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Kunin ang Arabic na pangalan, organisasyon, lugar at Hijri na petsa sa JSON gamit ang LLM: prompt, span check na walang diacritics, date conversion, scoring script. - [Paghahambing ng Arabic OCR API: Google, Azure, Textract, Mistral, LLM](https://broadrouter.com/fil/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): Aling OCR API ang opisyal na sumusuporta sa Arabic, printed at sulat-kamay, ano ang ibinabalik ng bawat isa at magkano kada 1,000 page, kumpara sa vision LLM API. - [Arabic subtitles (SRT) mula sa speech-to-text API: ang kailangan mo](https://broadrouter.com/fil/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): Ano ang kailangan ng SRT file, aling speech-to-text API ang may timestamp o SRT para sa Arabic, at nasubok na pipeline para sa subtitle mula sa text-only na API. - [Arabic text classification gamit ang LLM: sentiment, intent, triage](https://broadrouter.com/fil/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): I-classify ang Arabic na review, mensahe at support ticket gamit ang LLM API: disenyo ng label, dialect-aware na prompt, JSON output, macro F1 test at gastos. - [Arabic text normalization para sa search at RAG: ano ang babaguhin](https://broadrouter.com/fil/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Tashkeel, alef, ta marbuta, hamza, Persian na letra, digit at Unicode: aling Arabic normalization rule ang tumutulong o nakakasama sa search, at tested na Python. - [Arabic text summarization gamit ang LLM: prompt, haba at scoring](https://broadrouter.com/fil/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): I-summarise ang Arabic na artikulo, report at dialect na mensahe gamit ang LLM API: prompt, kontrol sa bilang ng salita, check sa numero, at bakit 0 ang ROUGE. - [Alternatibo sa Artificial Analysis: saan ihahambing ang LLM sa 2026](https://broadrouter.com/fil/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Alternatibo sa Artificial Analysis: Arena, Epoch AI, Vals AI, OpenRouter rankings, Hugging Face at price data, ayon sa sinusukat ng bawat isa at kailan gamitin. - [GPU hardware at paraan ng serving ng BroadNet, ayon sa BroadNet](https://broadrouter.com/fil/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): Ang sinasabi ng BroadNet tungkol sa hardware sa likod ng BroadNet Inference: tatlong GPU tier mula B300 hanggang appliance, throughput kada GPU at aral ng test. - [I-cache ang LLM API response, makatipid: exact-match cache sa Python](https://broadrouter.com/fil/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Na-test na exact-match cache para sa LLM API response sa Python: ano ang nasa key, ano ang hindi dapat i-cache, TTL, privacy, at paano sukatin ang hit rate. - [Claude Code sa open model: proxy, native endpoint at mga limit](https://broadrouter.com/fil/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Anthropic Messages API ang sinasalita ng Claude Code. Paano ito patakbuhin sa open model gamit ang LiteLLM proxy o native endpoint, at ano ang nasisira sa daan. - [Gawing bahagi ng document AI workflow mo ang human review](https://broadrouter.com/fil/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): Paano idisenyo ang human review sa document-AI pipeline: validation rules, ano ang ipapasa sa tao, random na audit sample, audit trail at review rate. - [Data protection law ng Egypt (Law 151 of 2020) at mga LLM API](https://broadrouter.com/fil/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Personal data ng mga Egyptian sa LLM o speech-to-text API sa ibang bansa: ang sinasabi ng Law 151 of 2020 at Decision 816 of 2025 sa transfer, lisensya at breach. - [Bago mag-fine-tune: tukuyin muna ang problemang lulutasin](https://broadrouter.com/fil/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): Kailan sapat na ang prompting, RAG o evaluation at hindi na kailangan ng fine-tuning, anong training data ang kailangan, at paano i-test ang fine-tune bago humiling. - [Paano ayusin ang invalid JSON mula sa LLM: sanhi, repair, retry](https://broadrouter.com/fil/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): Bakit sirang JSON ang ibinabalik ng LLM (code fence, prose, naputol sa max_tokens, trailing comma) at isang tested na Python function na nag-aayos lang ng ligtas. - [Gaano kadalas nagbabago ang presyo ng LLM API? 3 taon ng archive](https://broadrouter.com/fil/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): Binasa namin ang 308 archived na pricing page ng 11 LLM API provider, 2023–2026. 80% ng model ang hindi nagbago ng presyo; DeepInfra ang may 75% ng pagbabago. - [Paano bawasan ang gastos sa LLM API: 10-step na checklist, sunud-sunod](https://broadrouter.com/fil/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): Pababain ang LLM API bill sa tamang ayos: sukatin, i-cache ang prefix, paikliin ang prompt, limitahan ang output, tamang model, batch, at kailan sulit ang flat rate. - [LLM API failover: lumipat sa pangalawang provider sa sarili mong code](https://broadrouter.com/fil/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): Paano mag-failover sa pagitan ng mga OpenAI-compatible na LLM provider: aling error, circuit breaker, mapping ng model ID, streaming, gastos at privacy. - [Pagbaba ng presyo ng LLM API 2023–2026: pinakamalalaki, may source](https://broadrouter.com/fil/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): Pinakamalalaking pagbaba ng presyo ng LLM API 2023–2026, may Internet Archive link bago at pagkatapos: o3, Codestral, Gemini 1.5, at mga pagtaas na di napansin. - [Bumababa lang ba ang presyo ng LLM API? 52 pagtaas at paano mag-budget](https://broadrouter.com/fil/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): Tumataas din ang presyo ng LLM API. 52 pagtaas sa parehong model ang nakita namin sa archived na pricing page, 2024–2026, median 40%. Lahat ng source at budget rule. - [Rate limit ng LLM API ayon sa provider: RPM, TPM at concurrency](https://broadrouter.com/fil/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): Rate limit ng entry tier ng 13 LLM API provider mula sa sarili nilang docs: OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks at iba pa, at ano ang RPM at TPM. - [LLM API SLA: ano ang pangako ng Azure, Bedrock, Google at OpenAI](https://broadrouter.com/fil/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): Ano ang nakasulat na pangako ng bawat LLM API provider: uptime percentage, ano ang itinuturing na downtime, service credit at latency target, mula sa mga SLA page. - [LLM batch processing nang walang Batch API: resumable na runner](https://broadrouter.com/fil/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Magpatakbo ng 10,000 hanggang 1,000,000 prompt sa chat-completions API nang walang Batch API: concurrency cap, rate pacing, retry, checkpoint at kuwenta ng oras. - [The Safety Map: paano hinahawakan ng 39 na LLM ang sensitibong data](https://broadrouter.com/fil/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): Sinuri ng BroadNet ang 39 na LLM mula sa 14 na lab sa credential, PII, health at financial data: ano ang nailipat, ano ang hindi, at leak sa tool call at reasoning. - [Paano i-test ang LLM sa safety ng sensitibong data: paraan ng BroadNet](https://broadrouter.com/fil/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): Ang paraan ng BroadNet sa pag-test kung nagle-leak ang LLM ng credential, PII, health o financial data bago i-deploy: mga kategorya, leak vs miss, 3+ run. - [Paglipat mula sa OpenAI API: code walkthrough sa Python, Node, cURL](https://broadrouter.com/fil/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): Mga pagbabago sa code mula OpenAI papunta sa OpenAI-compatible API: client setup, Responses sa Chat Completions, tools, streaming, embeddings at smoke test. - [Mga model na sinubukan at tinanggihan ng BroadNet, at bakit](https://broadrouter.com/fil/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): Inilalathala ng BroadNet ang mga model na sinubukan at tinanggihan nito, kasama ang dahilan: chat model na bumagsak sa credential-safety, at speech, OCR, embedding. - [Paano mag-design ng AI product para sa maraming wika](https://broadrouter.com/fil/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): Pag-design ng AI product para sa Arabic, English at Spanish: UI locale vs wika ng model, RTL at halo-halong text, numero, petsa, detection at gastos. - [Gastos ng open-weight vs closed model API: ano ang sinasabi ng data](https://broadrouter.com/fil/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): Mas mura ba sa API ang open-weight model kaysa GPT, Claude o Gemini? Per-token na presyo mula sa page ng mga provider, agwat ng host at gastos kada user. - [Kasaysayan ng presyo ng OpenAI vs Anthropic vs Google API, 2023–2026](https://broadrouter.com/fil/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): Presyo ng GPT, Claude at Gemini API na magkatabi mula sa archived na pricing page, 2023–2026: aling linya ang mumura, alin ang mumahal, may archive link bawat isa. - [Base URL ng OpenAI-compatible API: 14 na provider, sinuri](https://broadrouter.com/fil/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): Ang base_url ng 14 na OpenAI-compatible LLM API, mula sa sariling docs ng bawat provider, kasama ang mga endpoint at quirk na sumisira sa diretsong paglipat. - [OpenClaw at Hermes Agent sa murang model: setup at token math](https://broadrouter.com/fil/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): Ikonekta ang OpenClaw o Hermes Agent sa OpenAI-compatible na endpoint, alamin bakit nauubos ang token sa scheduled check, at paano babaan ang bill. - [Nagsisimula ang private endpoint sa malinaw na workload](https://broadrouter.com/fil/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): Ano ang dapat tukuyin bago humingi ng private LLM endpoint: workload, token kada segundo, concurrency, context, region, lisensya at budget, at paano ito i-test. - [Prompt injection sa coding agent: mga depensang talagang gumagana](https://broadrouter.com/fil/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): Saan umaabot ang prompt injection sa coding agent, at ang patong-patong na depensang itinuturo ng vendor at researcher: sandbox, egress limit, allow-list, approval. - [RAG chunk size at overlap: epekto sa retrieval at gastos](https://broadrouter.com/fil/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): Paano binabago ng chunk size at overlap ang recall ng RAG, laki ng index, gastos sa embedding at token kada query sa LLM, gamit ang published na resulta at kuwenta. - [Paano itakda ang max_tokens sa reasoning model: paraan ng pag-budget](https://broadrouter.com/fil/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Bakit walang laman ang sagot ng reasoning model kapag mababa ang max_tokens, at paano itakda ang budget mula sa sinukat na percentile, context window at retry. - [Paano mag-record ng audio sa browser at ipadala sa speech-to-text API](https://broadrouter.com/fil/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): Mag-record ng mikropono gamit ang MediaRecorder, gawing 16 kHz WAV ang WebM o MP4 ng Safari, at ipadala sa transcription API sa server na may hawak ng key. - [Bawasan ang prompt token: sinukat na tipid sa system prompt at tools](https://broadrouter.com/fil/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): Sinukat na tipid sa token sa system prompt, few-shot example, tool schema, format ng data, chat history at Arabic diacritics, gamit ang totoong tokenizer. - [Parehong model, ibang host: gaano kalaki ang agwat ng presyo?](https://broadrouter.com/fil/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): Sa 119 na model na ibinebenta ng ilang host, median na 1.76 beses ng pinakamura ang pinakamahal na list price. Mula sa pricing page ng 41 provider, Nobyembre 2026. - [Saudi at UAE PDPL at mga LLM API: checklist sa cross-border transfer](https://broadrouter.com/fil/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Personal data sa LLM API mula Saudi Arabia o UAE: ano ang sinasabi ng PDPL sa transfer, ano ang itatanong sa provider, at kailan mahalaga ang in-country inference. - [Text, speech o vision? Magsimula sa task](https://broadrouter.com/fil/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): Pumili sa text chat, image input o audio transcription ayon sa task, gamit ang mga model at limit na dinodokumento ng BroadNet, at kung ano ang wala, gaya ng TTS. - [Pag-transcribe ng audio na lampas 25 MB: hatiin, ipadala, pagdugtungin](https://broadrouter.com/fil/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): Pag-transcribe ng recording na lampas sa file limit ng speech-to-text API: paliitin, hatiin sa katahimikan gamit ang ffmpeg, ipadala nang sabay, pagdugtungin. - [Arabic translation ng app strings gamit ang LLM: placeholder, plural](https://broadrouter.com/fil/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): I-translate sa Arabic ang i18n JSON gamit ang LLM API: i-freeze ang placeholder, anim na Arabic plural form, ICU validation sa CI at isolation ng pangalan. - [Isalin ang Arabic audio sa English gamit ang API: tested na pipeline](https://broadrouter.com/fil/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): Gawing English text ang Arabic recording sa dalawang API call: transcribe na may language=ar, saka isalin gamit ang LLM, may number check, chunking, gastos, scoring. - [Live, on request, research: paano basahin ang model catalogue](https://broadrouter.com/fil/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): Paano malalaman kung aling model sa catalogue ang matatawag mo ngayon, alin ang kailangang i-request at alin ang sinubukan lang, gamit ang BroadNet at /v1/models. - [Uno sa vLLM: ano ang sinukat ng serving study ng BroadNet](https://broadrouter.com/fil/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): Isinama ng BroadNet ang Uno self-drafting adapter sa vLLM at sinukat sa H100: 2.58x ng plain throughput sa isang stream, 1.74x sa 32. Paraan, DOI at limitasyon. - [Uno sa Gemma 4 26B: ang mixture-of-experts na pag-aaral ng BroadNet](https://broadrouter.com/fil/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): Sinanay ng BroadNet ang unang Uno self-drafting adapter para sa mixture-of-experts model, Gemma 4 26B A4B. Ano ang sinukat, kailan sulit, at ang DOI. - [Token budget ng agent: saan napupunta ang token ng agent loop](https://broadrouter.com/fil/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): System prompt, tool schema, pagbasa ng file, history at reasoning: paano gumagastos ng token ang agent loop, paano sukatin ang bawat isa at ano ang unang babawasan. - [Arabic ASR word error rate (WER), ipinaliwanag](https://broadrouter.com/fil/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): Ano ang ibig sabihin ng word error rate, paano ito kinukuwenta, at bakit sensitibo ang Arabic WER sa diacritics, spelling at dialect. May Python at totoong numero. - [Arabic dialect speech recognition: Gulf, Egyptian, Levantine, Maghrebi](https://broadrouter.com/fil/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Bakit nahihirapan ang speech-to-text sa Gulf, Egyptian, Levantine at Maghrebi Arabic, at paano bumuo ng dialect test set para sukatin ang provider sa audio mo. - [Arabic–English translation gamit ang LLM: prompt, glossary, evaluation](https://broadrouter.com/fil/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): Mag-translate ng Arabic papuntang English at pabalik gamit ang LLM API: prompt na nagpapanatili ng tono, glossary, numero at pangalan, at evaluation ng kalidad. - [Gulf, Egyptian at Levantine Arabic: paano hinaharap ng LLM ang dialect](https://broadrouter.com/fil/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): Paano hinaharap ng LLM ang Gulf, Egyptian at Levantine Arabic: saan pumapalya, paano mag-prompt at mag-test, at ang ulat ng BroadNet sa Qwen3.8 27B. - [Arabic OCR vs vision LLM: pagbasa ng Arabic na dokumento at invoice](https://broadrouter.com/fil/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): Kailan gagamit ng classic OCR at kailan ng vision LLM para sa Arabic na invoice, resibo at form: mga bitag sa accuracy, JSON output, gastos kada page at validation. - [Arabic prompt engineering: register, RTL, numeral at halo-halong text](https://broadrouter.com/fil/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): Sumulat ng prompt na nagbibigay ng mahusay na Arabic mula sa LLM: register at dialect, numeral at petsa, malinis na RTL formatting at halo ng Arabic at English. - [Arabic RAG: normalisation, chunking at retrieval na gumagana](https://broadrouter.com/fil/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): RAG sa mga Arabic na dokumento: i-normalise ang alef, ya at ta marbuta, mag-chunk nang maayos, pumili ng multilingual embeddings at magdagdag ng keyword search. - [Arabic speech-to-text API: paano pumili sa 2026](https://broadrouter.com/fil/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): Paano pumili ng Arabic transcription API: word error rate, dialect, limit sa file, presyo kada oras ng audio, privacy, at test plan para sa sarili mong recording. - [Arabic WhatsApp chatbot architecture: mula voice note hanggang reply](https://broadrouter.com/fil/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): Gumaganang architecture para sa Arabic WhatsApp assistant: webhook, queue, transcription ng voice note, LLM reply, policy at gastos, may Python na puwedeng iangkop. - [Arabizi at LLM: paghawak ng Franco-Arabic na may numerong 3 at 7](https://broadrouter.com/fil/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): Arabic sa Latin na letra ang Arabizi, may digit gaya ng 3 para sa ع at 7 para sa ح. Paano ito i-detect, i-convert sa Arabic script gamit ang LLM, at sagutin. - [Pinakamahusay na LLM para sa Aider: paano pumili at ikonekta ang model](https://broadrouter.com/fil/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Paano pumili ng LLM para sa Aider: edit format, laki ng context at gastos kada oras, isang oras na test sa sarili mong repo, at ang eksaktong OPENAI_API_BASE. - [Pinakamahusay na open model sa coding agent: tools, context, gastos](https://broadrouter.com/fil/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B o Gemma 4 para sa coding agent mo? Pumili ayon sa tool calling, laki ng context, reasoning at gastos kada task. - [Cached input token: paano nito binabawasan ang LLM API bill mo](https://broadrouter.com/fil/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): Maliit na bahagi lang ng presyo ng input ang cached input. Paano sinisingil ang prompt caching, paano basahin ang cached_tokens, at paano ayusin ang prompt. - [Pinakamurang vision LLM API: gastos kada larawan at kada page](https://broadrouter.com/fil/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): Magkano ang vision LLM call kada larawan at page ng dokumento: token math para sa Gemma 4, Qwen3.8 27B at DeepSeek V4.1 Flash, at kung saan mas mura ang DeepInfra. - [Paano pumili ng LLM API provider: 20-point na checklist](https://broadrouter.com/fil/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): Dalawampung tanong bago pumili ng LLM API: presyo, limit, OpenAI compatibility, privacy, model, support at uptime, kasama ang scoring sheet at test plan. - [Pagpili ng LLM API para sa Arabic: ano ang dapat i-test](https://broadrouter.com/fil/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): Paano pumili ng LLM para sa Arabic na app: dialect, kalidad ng pagsulat, translation, Arabic na dokumento at larawan, gastos, at test set na kaya mo sa isang araw. - [Cline vs Roo Code vs Aider: para sa team sa open model (2026)](https://broadrouter.com/fil/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (archived na, itinuloy bilang Zoo Code) at Aider, pinaghambing sa mahalaga sa open model: tool call, gamit ng context, MCP, git at approval. - [Paano bilangin ang token bago magpadala ng LLM request](https://broadrouter.com/fil/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): Tantiyahin ang prompt token kada wika at model bago tawagin ang LLM API: bakit magkaiba ang tokenizer, sukat sa limang wika, at ligtas na margin sa max_tokens. - [10 paraan para bawasan ang gastos sa LLM ng coding agent, may tantiya](https://broadrouter.com/fil/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Sampung kongkretong paraan para pababain ang gastos sa token ng Cline, Aider, opencode o sarili mong coding agent, may math sa bawat tipid at kung paano ito sukatin. - [Alternatibo sa DeepSeek API: 5 paraan para tawagin ang DeepSeek](https://broadrouter.com/fil/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): Pagtawag sa DeepSeek sa official API, aggregator, cloud catalogue, sariling GPU o BroadNet Inference: mga trade-off sa presyo, privacy at limit. - [Presyo ng DeepSeek V4.1 Flash API: kada token, cached at mga plan](https://broadrouter.com/fil/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): Presyo ng DeepSeek V4.1 Flash API kada 1M input, cached at output token, flat na plan kada buwan, linggo at oras, at kumpara sa DeepSeek, DeepInfra at Fireworks. - [Paano tantiyahin ang gastos sa LLM API bago mag-launch ng feature](https://broadrouter.com/fil/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): Sunud-sunod na cost model para sa bagong AI feature: user × action × call × token, kasama ang cache, reasoning, retry, rate limit at halimbawang buwanang table. - [Paano mag-evaluate ng LLM bago lumipat ng provider](https://broadrouter.com/fil/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): Bumuo ng eval set mula sa totoong traffic, i-score gamit ang code check at debiased na LLM judge, at ihambing ang quality, gastos at latency bago mag-migrate. - [Data ng invoice, resibo at form sa JSON gamit ang vision LLM](https://broadrouter.com/fil/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): Production pipeline para sa invoice, resibo at form: paghanda ng larawan, JSON schema na prompt, Pydantic validation, arithmetic check, retry at human review queue. - [Flat-rate vs pay-per-token na LLM API: ang break-even math](https://broadrouter.com/fil/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): Kailan mas sulit ang flat-rate na LLM plan kaysa magbayad kada token? Apat na halimbawang bill sa DeepSeek V4.1 Flash, ang break-even at paano sukatin ang sa iyo. - [Libreng LLM API credit: gaano kalayo talaga ang panimulang credit](https://broadrouter.com/fil/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): Ano ang nabibili ng libreng LLM API credit: token, minuto ng audio at hourly key sa panimulang credit ng BroadNet, kumpara sa free tier ng Gemini at OpenRouter. - [GDPR at mga LLM API: 15 tanong para sa iyong provider](https://broadrouter.com/fil/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): GDPR checklist para sa LLM API: DPA, subprocessor, retention, training, international transfer at breach notice, kasama ang GDPR article sa likod ng bawat tanong. - [GLM 5.3 vs DeepSeek V4.1 Flash: alin ang gagamitin, at kailan](https://broadrouter.com/fil/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 vs DeepSeek V4.1 Flash sa presyo, context, tools, structured output, larawan at reasoning, may kuwenta ng gastos at simpleng patakaran sa pagpili. - [Paano ayusin ang 429 error sa LLM API: backoff, jitter at limiter](https://broadrouter.com/fil/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 vs 5xx, exponential backoff na may full jitter, concurrency limiter at queue para sa LLM API, may Python at Node code na puwedeng kopyahin sa production. - [Hourly LLM API key: flat na presyo kada oras, ipinaliwanag](https://broadrouter.com/fil/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): Bumibili ang hourly key ng BroadNet ng isang oras ng Gemma 4 VL 26B sa iisang stream sa fixed na presyo. Ano ang kasama, mga limit, para kanino at kailan sulit. - [Data retention ng LLM API: OpenAI, Anthropic, Gemini, Mistral](https://broadrouter.com/fil/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): Ano ang opisyal na sinasabi ng OpenAI, Anthropic, Gemini API, Mistral at BroadNet sa pagtatago at training sa API prompt, mula sa page nila noong Nobyembre 2026. - [Seguridad ng LLM API key: pag-iimbak, rotation at IP allow-list](https://broadrouter.com/fil/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): Ilayo ang LLM API key sa code at logs, bigyan ng sariling key ang bawat workload, mag-rotate nang regular, i-lock sa server IP at i-revoke ang leak sa ilang segundo. - [Presyo ng LLM API, ipinaliwanag: input, cached, output at reasoning](https://broadrouter.com/fil/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): Paano gumagana ang per-token na presyo ng LLM API: input, cached input, output, reasoning token, larawan at audio kada minuto, may mga halimbawang kuwenta. - [Sensitive data sa LLM API: mag-redact, magbawas, mag-log nang kaunti](https://broadrouter.com/fil/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): Paano ligtas na ipadala ang customer, health o financial data sa LLM API: i-classify ang field, mag-redact gamit ang placeholder, zero retention, malinis na log. - [Kontrol sa gastos ng LLM API: budget, limit sa key at loop protection](https://broadrouter.com/fil/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): Pigilan ang paglobo ng LLM API bill: prepaid balance, isang key kada app, IP allow-list, usage logging, budget sa code, limit sa loop at flat-price na key. - [Usage logging ng LLM API: gastos kada user at feature, walang PII](https://broadrouter.com/fil/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): I-log ang bawat tawag sa LLM API kasama ang token, gastos, latency at request ID, iugnay ang gastos sa user at feature, i-redact ang PII bago isulat. May Python. - [Gastos ng LLM kada user: unit economics ng AI feature sa SaaS](https://broadrouter.com/fil/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): Magkano ang AI feature kada aktibong user kada buwan, bakit mas mahalaga ang heavy user kaysa average, at paano presyuhan ang SaaS tier para sakop ang bill. - [Presyo ng open model sa mga inference provider: 6 na inihambing](https://broadrouter.com/fil/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): Paano nagpepresyo ng open model ang OpenRouter, Together AI, Fireworks, Groq, DeepInfra at BroadNet: kada token, cached, batch, GPU hour at flat plan, may source. - [LLM latency: TTFT vs tokens per second vs end-to-end na oras](https://broadrouter.com/fil/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): Ano ang sinusukat ng time to first token, tokens per second at end-to-end latency, paano i-time ang bawat isa sa client gamit ang code, at aling ayos ang para saan. - [Paliwanag sa LLM streaming: SSE chunk, delta, usage at cancel](https://broadrouter.com/fil/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Paano gumagana sa wire ang chat-completion streaming, paano buuin ang sagot mula sa delta, basahin ang usage sa huling chunk at mag-cancel sa Python, Node at curl. - [Long context vs RAG: kailan pinapalitan ng 1M token ang vector DB](https://broadrouter.com/fil/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): Kailan napapalitan ng 1M-token na context window ang vector database at kailan panalo pa rin ang RAG: gastos kada query, latency, recall at maintenance. - [Context management ng long-running agent na may 1M window](https://broadrouter.com/fil/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): Paano panatilihing tumpak at abot-kaya ang mahabang agent session sa 1M-token window: append-only na history, prompt na cache-friendly, kailan mag-compact. - [Matagal na tumatakbong agent: timeout, retry at checkpoint na matibay](https://broadrouter.com/fil/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): Paano patakbuhin nang ilang oras ang LLM agent: timeout kada call, streaming para makita ang stall, ligtas na retry, idempotent na tool at checkpoint kada step. - [Paano gamitin ang MCP server sa OpenAI-compatible na model](https://broadrouter.com/fil/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): Gumagana ang MCP sa kahit anong LLM: ginagawang function call ng client ang MCP tools. Gamitin ang MCP sa open model sa Cline, Continue, opencode, n8n o Python. - [Kailan sulit ang 1M-token na context window (at kailan hindi)](https://broadrouter.com/fil/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): Ano ang kasya sa 1,048,576 token, magkano ito sa DeepSeek V4.1 Flash, paano binabago ng reasoning at caching ang kuwenta, at kailan mas mainam pa rin ang RAG. - [Magkano talaga ang multi-agent na setup: per token vs flat session](https://broadrouter.com/fil/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): Ilang beses na mas maraming token ang gamit ng multi-agent kaysa sa isang agent. Math ng session: flat plan vs per token, nakatagong multiplier, at alin ang panalo. - [Multilingual speech-to-text API: 14 na wika, iisang endpoint](https://broadrouter.com/fil/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): English, Spanish, French, Chinese, Arabic at 9 pang wika sa iisang OpenAI-compatible na endpoint, presyo kada minuto, at kailan gamitin ang Arabic model. - [n8n AI Agent sa murang LLM: setup, pagpili ng model at gastos kada run](https://broadrouter.com/fil/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): Gumawa ng n8n AI Agent workflow sa murang OpenAI-compatible na model: aling model para sa tools, mga setting na nagbabawas ng token, at gastos kada 10,000 run. - [Open-weight vs closed na AI model sa negosyo: gastos, kontrol, risk](https://broadrouter.com/fil/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): Open-weight o closed na LLM para sa negosyo mo? Ikumpara ang gastos, kontrol, privacy, quality gap at lock-in, at ang tatlong paraan ng pagpapatakbo ng open model. - [Paglipat sa OpenAI-compatible API: checklist sa migration](https://broadrouter.com/fil/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): Madali ang pagpalit ng base_url. Ang madalas masira: model ID, reasoning token, usage sa streaming, error, rate limit at key. Checklist na may code. - [OpenRouter vs flat-rate plan para sa AI agent: alin ang mas mura?](https://broadrouter.com/fil/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): Kada token at may fee sa credits ang OpenRouter; kada session ang flat plan. Kuwenta para sa always-on na agent, fees at limit ng OpenRouter, at alin ang bagay. - [Paano patakbuhin nang ligtas ang coding agent nang magdamag](https://broadrouter.com/fil/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): Praktikal na setup para sa coding agent na walang nagbabantay: sandbox, git checkpoint, task file, step limit, review sa umaga, at flat plan session sa gabi. - [Paano ayusin ang prompt para ma-cache ang prefix](https://broadrouter.com/fil/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): Paano ilatag ang LLM prompt para manatiling pareho ang naka-cache na prefix: system prompt, tool schema, dokumento, append-only na history, at test na patunay. - [Reasoning token: bakit mas maraming output ang binabayaran mo](https://broadrouter.com/fil/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): Output token ang singil sa nakatagong pag-iisip ng reasoning model. Paano sukatin, magkano kada model, at paano bawasan sa reasoning_effort, max_tokens at routing. - [Paano magpatakbo ng sabay-sabay na coding agent sa flat-rate LLM API](https://broadrouter.com/fil/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Sub-agent, orchestrator at batch worker sa unlimited na DeepSeek V4.1 Flash. Ilang session ang kailangan, paano hatiin sa mga key, at paano limitahan ang overlap. - [Self-hosting ng LLM vs API: ang tunay na gastos at break-even](https://broadrouter.com/fil/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): Presyo ng GPU rental mula sa opisyal na page, formula ng gastos kada 1M token sa sariling GPU, utilization na kailangan para talunin ang API, at nakakalimutang ops. - [Presyo ng speech-to-text API, ikinumpara: kada minuto sa 2026](https://broadrouter.com/fil/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): Presyo kada minuto at kada 1,000 oras ng OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq at BroadNet Transcribe mula sa opisyal na page, pati add-on. - [JSON mode vs json_schema: structured output na maaasahan](https://broadrouter.com/fil/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): Valid na JSON ang bigay ng JSON mode; ang hugis mo ang bigay ng json_schema. Paano hingin ang bawat isa, i-validate gamit ang Pydantic, at ayusin ang maling output. - [Token kada salita sa Arabic vs English: bakit mas mahal ang Arabic](https://broadrouter.com/fil/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): Bakit madalas mas maraming LLM token ang Arabic kaysa English sa parehong kahulugan, bakit nakakalito ang token kada salita, at script para sukatin ang ratio mo. - [Tool calling sa open model: schema, parallel call, retry](https://broadrouter.com/fil/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): Function calling sa OpenAI-compatible na open model: tips sa schema, parallel tool call, limang paraan ng pagpalya, at Python loop na nagva-validate at nagre-retry. - [Paano i-transcribe ang call center recording: pipeline para sa Arabic](https://broadrouter.com/fil/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): Hatiin ang stereo na tawag kada channel, putulin sa speech turn, i-transcribe ang Arabic at English, i-redact ang PII at ibuod gamit ang LLM. May code at gastos. - [Open model sa coding agent: Aider, Cline, opencode, Codex](https://broadrouter.com/fil/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Ikonekta ang Aider, Cline, Roo Code, opencode, Continue at Codex CLI sa OpenAI-compatible endpoint. Eksaktong pangalan ng field at mga setting na nasisira. - [Voice note sa text para sa negosyo: WhatsApp at Telegram sa ticket](https://broadrouter.com/fil/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): Gawing transcript at support ticket ang voice note ng customer mula sa WhatsApp Business at Telegram bot: i-download, i-transcribe, kunin ang field gamit ang LLM. - [Lingguhang LLM plan para sa maiikling proyekto: Orbit vs oras vs buwan](https://broadrouter.com/fil/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): Kailan mas sulit ang isang linggong unlimited DeepSeek V4.1 Flash plan kaysa hourly session, buwanang plan o bayad kada token, may kuwenta sa hackathon at sprint. - [Ano talaga ang ibig sabihin ng "unlimited" sa isang LLM API plan](https://broadrouter.com/fil/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Token meter, rolling na 5-hour window, lingguhang cap, rate limit at concurrency. Pitong tanong para malaman kung talagang unlimited ang AI plan para sa agent mo. - [Alternatibo sa Whisper API: kailan lilipat, at paano ito i-test](https://broadrouter.com/fil/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): Kailan makatuwiran ang hosted na alternatibo sa OpenAI Whisper API: word error rate sa audio mo, mga wika, presyo kada oras, limit at privacy, may test code. - [Zero data retention sa LLM API: checklist na may 10 tanong](https://broadrouter.com/fil/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): Ano dapat ang ibig sabihin ng "zero data retention" sa LLM API: prompt, output, logs, training, subprocessor at audit. Sampung tanong bago ka magpadala ng data. - [Guides in Filipino](https://broadrouter.com/fil/guides/) - [Paano gamitin ang OpenAI Python SDK sa BroadNet Inference](https://broadrouter.com/fil/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): Ituro ang opisyal na OpenAI Python SDK sa BroadNet Inference sa isang palit lang ng base_url. Mag-install, i-set ang key, magpadala ng request at i-stream ang sagot. - [Paano gamitin ang OpenAI Node.js SDK sa BroadNet Inference](https://broadrouter.com/fil/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): Ituro ang opisyal na OpenAI Node.js at TypeScript SDK sa BroadNet Inference gamit ang baseURL. I-install, magpadala ng request, mag-stream at i-log ang request ID. - [Subukan ang BroadNet Inference sa browser, walang account](https://broadrouter.com/fil/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): Subukan ang Arabic speech-to-text, pagbasa ng dokumento at Gemma 4 chat sa browser playground ng BroadNet nang walang sign-up o card, saka ilipat ang request sa API. - [Paano gamitin ang curl sa BroadNet Inference: ang una mong request](https://broadrouter.com/fil/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): Ipadala ang una mong BroadNet Inference call gamit ang curl: chat completion, streamed na sagot at audio transcription, at basahin ang response, usage at errors. - [Paano mag-stream ng chat completions at ayusin ang naputol na stream](https://broadrouter.com/fil/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): Streaming sa BroadNet Inference: Server-Sent Events, ang [DONE] marker, token usage sa huling chunk, at ano ang gagawin kapag maagang natapos ang stream. - [Bakit walang laman ang sagot ng isang reasoning model](https://broadrouter.com/fil/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): Nag-iisip muna ang DeepSeek V4.1 Flash, GLM 5.3 at Qwen3.8 27B bago sumagot, at kasama ang reasoning sa max_tokens. Paano maiiwasan ang mga empty reply. - [Mga error code ng BroadNet Inference at kailan dapat mag-retry](https://broadrouter.com/fil/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): Ano ang ibig sabihin ng bawat HTTP status ng BroadNet Inference API, alin ang ire-retry nang may backoff, alin ang aayusin muna, at ano ang ipapadala sa support. - [I-lock ang API key sa mga server mo gamit ang IP allow-list](https://broadrouter.com/fil/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): Protektahan ang key ng BroadNet Inference gamit ang IP allow-list, ihiwalay ang inference key sa portal credential at i-revoke agad ang na-leak na key. May curl. - [Paano gumagana ang unlimited LLM API plans para sa mga agent](https://broadrouter.com/fil/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): Walang token meter at walang weekly limit ang unlimited na DeepSeek V4.1 Flash plans. Ganito talaga gumagana ang sessions, parallel agents, keys at billing. - [Aider sa unlimited na DeepSeek V4.1 Flash: paano i-set up](https://broadrouter.com/fil/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Ikonekta ang Aider coding agent sa BroadNet Inference gamit ang OpenAI-compatible na setting nito at patakbuhin sa flat-price na unlimited na DeepSeek V4.1 Flash. - [Paano gamitin ang LangChain sa BroadNet Inference](https://broadrouter.com/fil/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): Ikonekta ang ChatOpenAI ng LangChain sa BroadNet Inference gamit ang base_url at api_key, at alamin kung ano ang itinatago o tinatanggal nito sa reasoning model. - [Cline sa BroadNet Inference: setup ng OpenAI Compatible provider](https://broadrouter.com/fil/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): I-set up ang Cline VS Code extension sa BroadNet Inference gamit ang OpenAI Compatible provider: eksaktong field, model ID, laki ng context at mga karaniwang ayos. - [Roo Code sa BroadNet Inference (OpenAI Compatible)](https://broadrouter.com/fil/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Isinara ang Roo Code noong Mayo 2026. Settings para sa mga kasalukuyang install at fork, at saan lilipat (Cline o ang Zoo Code fork) gamit ang BroadNet Inference. - [Continue sa BroadNet Inference: config.yaml at apiBase](https://broadrouter.com/fil/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): Idagdag ang mga model ng BroadNet Inference sa Continue IDE extension gamit ang provider openai at apiBase sa config.yaml, kasama ang tool_use at context length. - [Idagdag ang BroadNet Inference sa opencode bilang custom provider](https://broadrouter.com/fil/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): Idagdag ang BroadNet Inference sa opencode coding agent bilang custom OpenAI-compatible provider sa opencode.jsonc, na may model alias, capabilities at limit. - [Paano ikonekta ang Open WebUI sa BroadNet Inference](https://broadrouter.com/fil/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Idagdag ang BroadNet Inference sa Open WebUI bilang OpenAI-compatible na connection, ilista ang model ID, at ayusin ang verify, blangkong sagot at rate limit. - [Idagdag ang BroadNet Inference sa LibreChat bilang custom endpoint](https://broadrouter.com/fil/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): Idagdag ang BroadNet Inference sa LibreChat bilang OpenAI-compatible na custom endpoint sa librechat.yaml, may key sa .env, model list, title model at mga ayos. - [BroadNet Inference sa n8n: OpenAI credential at Base URL](https://broadrouter.com/fil/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): Ituro ang OpenAI credential ng n8n sa BroadNet Inference gamit ang Base URL, saka gamitin ang OpenAI Chat Model node ayon sa ID nang naka-off ang Responses API. - [Paano gamitin ang Vercel AI SDK sa BroadNet Inference](https://broadrouter.com/fil/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): Tawagin ang BroadNet Inference sa Vercel AI SDK gamit ang createOpenAICompatible ng @ai-sdk/openai-compatible: generateText, streamText, tools, reasoning effort. - [Arabic speech-to-text API: gabay na hakbang-hakbang](https://broadrouter.com/fil/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): I-transcribe ang Arabic audio gamit ang BroadNet Transcribe Arabic sa OpenAI-compatible na endpoint: mga format, ang language field, Python at curl, at ang presyo. - [Paano i-route ang LiteLLM sa BroadNet Inference (SDK at proxy)](https://broadrouter.com/fil/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): Tawagin ang BroadNet Inference sa LiteLLM gamit ang openai/ prefix at api_base, sa Python SDK o sa config.yaml ng proxy, na may katugmang rate limit. - [Patakbuhin ang OpenAI Agents SDK sa BroadNet Inference](https://broadrouter.com/fil/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Gamitin ang OpenAI Agents SDK (Python) sa BroadNet Inference gamit ang AsyncOpenAI at OpenAIChatCompletionsModel, at i-off ang tracing na kailangan ng OpenAI key. - [Paano gawing text ang WhatsApp voice notes gamit ang API](https://broadrouter.com/fil/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): I-convert sa WAV gamit ang ffmpeg ang WhatsApp .ogg voice notes o .m4a recording ng phone, saka i-transcribe sa Arabic o 13 pang wika sa isang API call. - [Paano i-transcribe ang WhatsApp voice notes sa n8n](https://broadrouter.com/fil/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): Isang n8n workflow na tumatanggap ng WhatsApp voice note, dina-download ito mula sa Meta, kino-convert gamit ang ffmpeg at tina-transcribe sa Arabic o 13 pang wika. - [Basahin ang mga Arabic na invoice at resibo gamit ang vision model](https://broadrouter.com/fil/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Ipadala ang litrato ng invoice o resibo sa isang vision model sa BroadNet Inference at makuha ang total, petsa at vendor bilang JSON, may tanong sa Arabic o English. - [Zed at BroadNet Inference bilang OpenAI-compatible na provider](https://broadrouter.com/fil/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Idagdag ang BroadNet Inference sa Zed editor bilang OpenAI-compatible na provider sa settings.json, may model IDs, token limits, capabilities at key variable. - [Kilo Code at BroadNet Inference (OpenAI Compatible na provider)](https://broadrouter.com/fil/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Ikonekta ang Kilo Code sa BroadNet Inference bilang custom OpenAI Compatible na provider, sa VS Code dialog o sa kilo.json, may model IDs, limits at tool flags. - [Idagdag ang BroadNet Inference sa goose bilang custom provider](https://broadrouter.com/fil/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): Ikonekta ang goose AI agent sa BroadNet Inference bilang OpenAI Compatible na custom provider, mula sa goose Desktop, goose configure o isang JSON file. - [Pydantic AI sa BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/fil/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Patakbuhin ang Pydantic AI agents sa BroadNet Inference gamit ang OpenAIChatModel at OpenAICompatibleProvider, pati tool calls, structured output at model profile. - [CrewAI sa BroadNet Inference bilang custom na OpenAI endpoint](https://broadrouter.com/fil/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): Ituro ang CrewAI agents sa BroadNet Inference gamit ang LLM(custom_openai=True, base_url=...), saka pumili ng model para sa tools, mahabang context at flat na plan. - [LlamaIndex at BroadNet Inference gamit ang OpenAILike](https://broadrouter.com/fil/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): Ikonekta ang LlamaIndex sa BroadNet Inference sa OpenAILike class, at i-set nang tama ang api_base, context_window, is_chat_model at is_function_calling_model. - [BroadNet Inference sa Dify gamit ang OpenAI-API-compatible plugin](https://broadrouter.com/fil/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): Ikonekta ang Dify sa BroadNet Inference gamit ang OpenAI-API-compatible model provider plugin: eksaktong value ng field, laki ng context at function calling. - [Jan at BroadNet Inference bilang custom OpenAI-compatible provider](https://broadrouter.com/fil/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Idagdag ang BroadNet Inference sa Jan desktop app bilang custom OpenAI-compatible provider, mano-manong idagdag ang model IDs at i-on ang tools at vision kada model. - [SillyTavern sa BroadNet Inference (Custom OpenAI-compatible)](https://broadrouter.com/fil/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): Ikonekta ang SillyTavern sa BroadNet Inference sa Custom (OpenAI-compatible) na Chat Completion source, tamang model ID, prompt post-processing at token limit. - [Spring AI sa BroadNet Inference (OpenAI base-url)](https://broadrouter.com/fil/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Ituro ang OpenAI chat starter ng Spring AI sa BroadNet Inference gamit ang spring.ai.openai.base-url, BroadNet model ID at max-tokens, at tawagin mula sa ChatClient. - [Tawagin ang BroadNet Inference mula sa Go gamit ang openai-go SDK](https://broadrouter.com/fil/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): Gamitin ang opisyal na openai-go v3 SDK sa BroadNet Inference sa pamamagitan ng option.WithBaseURL, na may chat completions, streaming na may usage at mga error. - [Paano gumawa ng LangGraph agent sa BroadNet Inference](https://broadrouter.com/fil/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): Patakbuhin ang LangGraph StateGraph o create_agent sa BroadNet Inference gamit ang ChatOpenAI, base_url at BroadNet model ID, kasama ang tools, streaming at memory. - [Haystack pipelines at agents gamit ang BroadNet Inference](https://broadrouter.com/fil/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Ituro ang OpenAIChatGenerator ng Haystack sa BroadNet Inference gamit ang api_base_url, model ID at Secret key para sa pipelines, tools, streaming at agents. - [Semantic Kernel sa BroadNet Inference](https://broadrouter.com/fil/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Ikonekta ang Microsoft Semantic Kernel sa BroadNet Inference: Python gamit ang AsyncOpenAI client, o C# gamit ang custom endpoint, para sa chat, plugins at agents. - [Paano gamitin ang Mastra agents sa BroadNet Inference](https://broadrouter.com/fil/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): Patakbuhin ang Mastra agents sa BroadNet Inference gamit ang model object na may id, url at apiKey, kasama ang dobleng broadnet/ prefix na kailangan ng model ID. - [Structured output gamit ang Instructor sa BroadNet Inference](https://broadrouter.com/fil/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Kumuha ng Pydantic objects mula sa BroadNet Inference gamit ang Instructor, sa from_openai o from_provider na may custom base_url, at piliin ang tamang Mode. - [DSPy sa BroadNet Inference: i-set ang model_type="chat"](https://broadrouter.com/fil/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): I-configure ang dspy.LM para sa BroadNet Inference gamit ang openai/, api_base at model_type="chat", dahil sa Responses API dinadala ng DSPy 3.5 ang OpenAI models. - [I-evaluate ang mga BroadNet model gamit ang promptfoo](https://broadrouter.com/fil/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): Magpatakbo ng promptfoo evals at side-by-side na model test sa BroadNet Inference gamit ang openai:chat provider, apiBaseUrl at apiKeyEnvar, pati mga grader. - [smolagents sa BroadNet Inference gamit ang OpenAIModel](https://broadrouter.com/fil/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Ikonekta ang smolagents ng Hugging Face sa BroadNet Inference sa OpenAIModel at api_base: CodeAgent sa kahit anong model, ToolCallingAgent sa GLM 5.3 at Qwen3.8 27B. - [Flowise at BroadNet Inference: gamit ang OpenAI Custom Model](https://broadrouter.com/fil/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Gamitin ang BroadNet Inference sa mga Flowise chatflow sa OpenAI Custom Model node, may Base Path, OpenAI API credential at tinipang BroadNet model name. - [Paano ikonekta ang Langflow sa BroadNet Inference](https://broadrouter.com/fil/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Gamitin ang BroadNet Inference sa Langflow sa OpenAI API Base field ng OpenAI component at tinipang model name, at bakit hindi ito ililista ng OpenAI Compatible. - [Glossary entries in Filipino](https://broadrouter.com/fil/glossary/) - [AI agent](https://broadrouter.com/fil/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): Ang AI agent ay isang system kung saan ang language model mismo ang nagdidirekta ng mga hakbang nito tungo sa isang layunin: pinipili nito kung aling tools ang tatawagin, binabasa ang mga resulta at umuulit hanggang makabuo ng huling sagot o maabot ang limit. Workflow naman ang karaniwang tawag sa fixed na script na tumatawag sa model sa takdang punto. - [API key](https://broadrouter.com/fil/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): Ang API key ay isang lihim na string na kumikilala at nagpapahintulot sa client na tumatawag sa isang API. Kadalasang hinihingi ito ng mga LLM API sa Authorization: Bearer header, at ang sinumang may hawak ng key ay makakagawa ng mga request na sisingilin sa account na may-ari nito. - [Arabic diacritics (tashkeel)](https://broadrouter.com/fil/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): Ang Arabic diacritics, o tashkeel, ay maliliit na markang isinusulat sa itaas o ibaba ng mga titik para ipakita ang maiikling patinig, dobleng katinig at kawalan ng patinig. Opsyonal ang mga ito: wala ang mga ito sa karamihan ng Arabic text, pati sa balita, kaya hinuhulaan ng mambabasa at ng software ang patinig mula sa context. - [Arabic dialect identification](https://broadrouter.com/fil/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): Ang Arabic dialect identification ay ang task ng pagtukoy kung anong variety ng Arabic ang isang text o recording, gaya ng Egyptian, Gulf, Levantine, Maghrebi o Modern Standard Arabic, sa antas ng rehiyon, bansa o lungsod. Ginagamit ito para i-route ang mga mensahe, pumili ng prompt at hatiin ang mga evaluation set. - [Arabic text normalization](https://broadrouter.com/fil/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): Ang Arabic text normalization ay ang hakbang na nagmamapa sa iisang anyo ng mga variant ng spelling ng parehong salita bago ang matching: pagtatanggal ng diacritics at tatweel, pag-iisa ng mga anyo ng alef, at kadalasang pagmapa ng ى sa ي at ة sa ه. Ginagamit ito sa search key at evaluation, hindi sa text na ipinapakita mo. - [Arabizi](https://broadrouter.com/fil/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): Ang Arabizi ay Arabic na impormal na isinusulat sa Latin na titik, gamit ang mga numero para sa mga tunog ng Arabic na walang katumbas sa Latin, gaya ng 3 para sa ع. Karaniwan ito sa chat at social media, kadalasang dialect, walang standard na spelling, at madalas na hinahaluan ng English o French. - [AWQ (activation-aware weight quantization)](https://broadrouter.com/fil/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): Ang AWQ (activation-aware weight quantization) ay post-training method na nagko-compress sa weights ng LLM sa mababang precision, kadalasang 4-bit integers. Gumagamit ito ng activation statistics mula sa maliit na calibration set para hanapin ang pinakamahalagang weight channels at i-scale ang mga ito para mas kaunti ang mawalang accuracy. Hindi nito kailangan ng backpropagation. - [Base URL](https://broadrouter.com/fil/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): Ang base URL ay ang fixed na root address ng isang API, gaya ng https://api.example.com/v1, na idinudugtong ng client sa unahan ng bawat endpoint path, gaya ng /chat/completions. Sa mga OpenAI-compatible SDK, kapag pinalitan ang base URL, kasama ang API key at model ID, sa ibang provider na tatawag ang parehong code. - [Batch API](https://broadrouter.com/fil/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Ang batch API ay asynchronous na LLM endpoint na tumatanggap ng maraming request nang sabay, kadalasan bilang file, pinoproseso ang mga ito sa loob ng nakatakdang window sa halip na agad-agad, at ibinabalik nang sabay-sabay ang mga resulta. May discount ang presyo ng batch request ng mga provider gaya ng OpenAI at Anthropic kumpara sa standard API nila. - [BM25](https://broadrouter.com/fil/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): Ang BM25 (Okapi BM25) ay ranking function para sa keyword search na nagbibigay ng score sa dokumento batay sa kung gaano kadalas lumalabas dito ang mga termino ng query, kung gaano kabihira ang mga terminong iyon sa buong collection, at kung gaano kahaba ang dokumento. Ito ang default na similarity sa Lucene at Elasticsearch. - [Byte pair encoding (BPE)](https://broadrouter.com/fil/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Ang byte pair encoding (BPE) ay subword tokenization algorithm na nagsisimula sa iisang character o byte at paulit-ulit na mina-merge ang pinakamadalas na magkatabing pares tungo sa bagong entry ng vocabulary, hanggang maabot ang target na laki ng vocabulary. BPE o byte-level na bersyon nito ang gamit ng karamihan ng LLM tokenizer ngayon. - [Cached input token (cached input tokens)](https://broadrouter.com/fil/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): Ang cached input token ay prompt token na kamakailan lang naproseso ng LLM provider at magagamit ulit mula sa cache sa halip na kuwentahin muli, kadalasang paulit-ulit na prefix gaya ng system prompt o dokumento. Mas mababa ang singil dito kaysa sa bagong input token. - [Chain-of-thought (CoT)](https://broadrouter.com/fil/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Ang chain-of-thought (CoT) ay prompting technique kung saan ipinapakita o hinihiling sa language model na isulat ang mga intermediate na hakbang ng reasoning bago ang huling sagot. Pinapataas nito ang accuracy sa mga problemang maraming hakbang gaya ng arithmetic at logic, at kusa na itong ginagawa ng mga reasoning model. - [Character error rate (CER)](https://broadrouter.com/fil/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Sinusukat ng character error rate (CER) ang accuracy ng transcription o OCR sa antas ng character: ang mga substitution, deletion at insertion na kailangan para maging reference text ang output, hinati sa bilang ng character sa reference. Mas mababa, mas maganda, at 0 ang perpektong match. - [Chat Completions API](https://broadrouter.com/fil/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Ang Chat Completions API ay HTTP endpoint ng OpenAI, POST /v1/chat/completions, na tumatanggap ng pangalan ng model at listahan ng mga message na may role, at ibinabalik ang susunod na message ng model. Ang hugis ng request at response nito ang naging karaniwang format na ginagamit din ng mga OpenAI-compatible na provider at SDK. - [Chat template](https://broadrouter.com/fil/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): Ang chat template ay ang formatting rule, kadalasang nakasulat sa Jinja at kasama ng tokenizer ng model, na nagko-convert ng listahan ng chat messages na may role tungo sa iisang token sequence na inaasahan ng model, kasama ang mga special control token nito. Puwedeng lubhang sumama ang sagot ng model kapag mali ang template. - [Chunking (RAG)](https://broadrouter.com/fil/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Ang chunking ay ang hakbang sa retrieval pipeline na naghahati sa mga dokumento sa mas maliliit na piraso, na tinatawag na chunks, bago i-embed at i-index ang mga ito. Itinatakda ng chunk size kung gaano karaming text ang laman ng bawat piraso, at inuulit ng chunk overlap ang ilang text sa magkatabing chunk para hindi mawala ang ideyang naputol. - [Code-switching](https://broadrouter.com/fil/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Ang code-switching ay ang paggamit ng dalawa o higit pang wika, o ng mga variety ng iisang wika, sa iisang usapan o pangungusap, sa pagsasalita man o pagsulat. Halimbawa nito ang mga nagsasalita ng Arabic na naghahalo ng English sa dialect, o nagpapalipat-lipat sa Modern Standard Arabic at sa dialect nila. - [AI coding agent](https://broadrouter.com/fil/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): Ang AI coding agent ay software tool kung saan gumagana ang isang language model sa codebase nang paulit-ulit sa isang loop: nagbabasa ng mga file, nag-e-edit ng mga ito, nagpapatakbo ng mga command gaya ng tests, binabasa ang output at nagpapasya ng susunod na hakbang hanggang matapos ang task o maabot ang limit. - [Concurrency](https://broadrouter.com/fil/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): Ang concurrency ay ang bilang ng request na sabay-sabay na in flight mula sa isang client. Madalas itong nililimitahan ng mga LLM API kada key o plan, hiwalay sa requests per minute, dahil humahawak ng GPU capacity ang bawat bukas na request hanggang matapos ang response nito. - [Context compaction](https://broadrouter.com/fil/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): Ang context compaction ay ang pagbubuod ng usapang malapit nang maabot ang limit ng context window at pagtutuloy ng session mula sa buod na iyon, para mapanatili ng agent ang mahahalagang desisyon at state habang tinatanggal ang raw na history gaya ng lumang tool output. - [Context engineering](https://broadrouter.com/fil/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): Ang context engineering ay ang trabaho ng pagpili at pag-maintain ng set ng mga token na nakikita ng language model sa bawat request, kasama ang system prompt, tool definitions, history, mga retrieved na dokumento at notes, para matanggap ng model ang pinakamaliit na set ng high-signal na impormasyong kailangan nito para magawa nang maayos ang task. - [Context rot](https://broadrouter.com/fil/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): Ang context rot ay ang unti-unting pagbaba ng accuracy at pagiging maaasahan ng mga language model habang humahaba ang input nila, kahit pareho pa rin ang task at malayo pa sa limit ng context window ang input. Pinasikat ang termino ng isang technical report ng Chroma noong 2025 na sumubok sa 18 model. - [Context window](https://broadrouter.com/fil/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): Ang context window ay ang pinakamaraming token na kayang hawakan ng isang language model sa iisang request. Sakop nito ang prompt, history ng usapan, tool definitions, mga larawan at ang nabuong output, kasama ang anumang reasoning tokens. Ang hindi kasya ay kailangang bawasan, ibuod o hatiin sa ilang request. - [Data residency](https://broadrouter.com/fil/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): Ang data residency ay requirement o commitment na iniimbak at pino-process ang data sa isang partikular na bansa o rehiyon. Sa isang LLM API, sakop nito kung saan pino-process ang mga prompt habang nag-i-inference at kung saan nakalagay ang anumang naka-store na content, at kadalasang hindi kasama ang account at billing data. - [Dialectal Arabic (mga dialect ng Arabic)](https://broadrouter.com/fil/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): Ang dialectal Arabic ay ang pang-araw-araw na sinasalitang Arabic ng isang rehiyon o lungsod, gaya ng Egyptian, Gulf, Levantine o Maghrebi Arabic. Naiiba ang mga dialect sa Modern Standard Arabic sa tunog, salita at grammar, walang standard na spelling, at lalong isinusulat online, kaya mas mahirap ang mga ito para sa mga language model. - [Speaker diarization](https://broadrouter.com/fil/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): Ang speaker diarization ay ang task ng paghahati ng audio recording sa mga segment ayon sa nagsasalita, na sumasagot sa tanong na "sino ang nagsalita at kailan". Nilalagyan nito ang bawat bahagi ng pananalita ng anonymous na speaker ID gaya ng SPEAKER_00, at kadalasang isinasama sa transcript para makita kung sino ang nagsabi ng bawat linya. - [Distillation (knowledge distillation)](https://broadrouter.com/fil/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): Ang distillation, o knowledge distillation, ay isang paraan ng training kung saan natututo ang mas maliit na student model na gayahin ang output ng mas malaking teacher model, gaya ng mga probability distribution o nabuong sagot nito, para mapanatili ng student ang malaking bahagi ng kalidad ng teacher sa mas mababang gastos sa pagpapatakbo. - [Data processing agreement (DPA)](https://broadrouter.com/fil/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): Ang data processing agreement (DPA) ay kontrata sa pagitan ng controller at processor na nagtatakda kung paano hahawakan ng processor ang personal data. Hinihingi ito ng GDPR Article 28(3), na naglilista rin ng laman nito: dokumentadong instruction, confidentiality, security, patakaran sa subprocessor, pagbura o pagbabalik ng data, at audit. - [Embeddings](https://broadrouter.com/fil/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): Ang embeddings ay mga listahan ng floating-point na numero (vector) na binubuo ng isang model para kumatawan sa isang piraso ng text, larawan o ibang input. Magkakalapit ang vector ng mga input na magkatulad ang kahulugan, kaya magagamit ang layo ng mga vector para i-rank ang search results, i-cluster ang mga dokumento o hanapin ang magkakaugnay na item. - [Eval (LLM evaluation)](https://broadrouter.com/fil/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Ang eval (maikli para sa evaluation) ay isang paulit-ulit na test ng language model o prompt: fixed na set ng mga input, paglalarawan kung ano ang magandang output, at grader (code, tao o ibang model) na nagbibigay ng score sa bawat output para maikumpara ang mga bersyon. - [Exponential backoff](https://broadrouter.com/fil/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): Ang exponential backoff ay estratehiya sa pag-retry kung saan mas matagal na naghihintay ang client pagkatapos ng bawat pumalyang attempt, kadalasang dinodoble ang delay hanggang sa isang cap, bago sumubok ulit. Ikinakalat ng random na jitter sa bawat paghihintay ang mga retry ng maraming client para hindi sila sabay-sabay na tumama sa server. - [Few-shot prompting](https://broadrouter.com/fil/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): Ang few-shot prompting ay isang technique kung saan may kasamang ilang halimbawang pares ng input at output ang prompt bago ang totoong input, para mahinuha ng language model ang task at format mula sa mga ito. Hindi nagbabago ang weights ng model; walang example ang zero-shot at isa ang one-shot. - [Fine-tuning](https://broadrouter.com/fil/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): Ang fine-tuning ay ang proseso ng pagpapatuloy ng training ng isang pre-trained na language model sa mas maliit na dataset para sa isang partikular na task, gaya ng mga halimbawang conversation, para magbago ang weights nito at mas maaasahan nitong magawa ang task o style na iyon nang hindi kailangan ng mahahabang instruction o example sa bawat prompt. - [finish_reason](https://broadrouter.com/fil/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): Ang finish_reason ay field sa bawat choice ng chat completion response na nagsasabi kung bakit huminto ang model sa pagbuo. Sa format ng OpenAI, ang mga value ay stop, length, tool_calls, content_filter at ang deprecated na function_call. Sa Messages API ng Anthropic, stop_reason ang katumbas nito. - [Flat-rate LLM plan (unlimited na plan)](https://broadrouter.com/fil/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): Ang flat-rate LLM plan ay naniningil ng fixed na presyo para sa isang period, gaya ng isang buwan, para sa access sa language model API sa halip na maningil kada token. Kadalasang may ibang nililimitahan ang ganitong plan, gaya ng concurrency, bilis, mga model o usage window, kaya kasinghalaga ng presyo ang mga limit. - [FP8](https://broadrouter.com/fil/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): Ang FP8 ay isang 8-bit floating-point na number format na ginagamit sa deep learning, na may dalawang encoding: E4M3, na may 4 na exponent bit at 3 mantissa bit, at E5M2, na may 5 exponent bit at 2 mantissa bit. Kalahati ang memory nito kumpara sa mga 16-bit format at native itong tumatakbo sa mga bagong GPU. - [Frequency at presence penalty](https://broadrouter.com/fil/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): Ang frequency at presence penalty ay mga sampling setting na nagbababa sa score ng mga token na nasa output na. Isang beses inilalapat ang presence penalty kapag lumabas na ang token; lumalaki naman ang frequency penalty sa bawat pag-ulit. Sa API ng OpenAI, parehong -2.0 hanggang 2.0 ang range, at 0 ang default. - [GDPR](https://broadrouter.com/fil/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): Ang GDPR (General Data Protection Regulation, Regulation (EU) 2016/679) ay ang batas ng EU sa pagproseso ng personal data. Ipinapatupad ito mula Mayo 25, 2018 at sakop nito ang mga organisasyong nasa EU at ilang nasa labas nito. Sa mga LLM API, kinokontrol nito ang mga kontrata sa processor, international transfer at data minimisation. - [GGUF](https://broadrouter.com/fil/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): Ang GGUF ay isang binary file format para sa pag-iimbak ng mga language model para sa inference gamit ang mga GGML-based na runtime gaya ng llama.cpp. Hawak ng iisang GGUF file ang weights, na kadalasang quantized, at ang metadata na kailangan para i-load ang model, kasama ang tokenizer at architecture, at puwede itong i-memory-map para mabilis mag-load. - [GPU-hour](https://broadrouter.com/fil/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): Ang GPU-hour ay isang graphics processor na ginamit nang isang oras. Sinisingil ng mga cloud provider ang mga inuupahang GPU instance kada oras o kada segundo, at ginagamit ng mga researcher ang GPU-hours para sabihin kung gaano karaming compute ang nagamit ng isang job. Binabayaran mo ang oras ng makina, abala man ito o hindi. - [Grounding](https://broadrouter.com/fil/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): Ang grounding ay ang pagkonekta ng output ng language model sa mga partikular at nabe-verify na source ng impormasyon, gaya ng mga ibinigay na dokumento, search result o database record, at pag-utos dito na sumagot mula sa mga iyon. Natutunton pabalik sa source ang mga grounded na sagot at mas maliit ang tsansang may gawa-gawang fact. - [Guardrails (LLM guardrails)](https://broadrouter.com/fil/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): Ang guardrails ay mga check na pinapatakbo ng application sa paligid ng isang language model call para manatili ito sa itinakdang limit. Sinasala ng input guardrails ang pumapasok sa model, chine-check ng output guardrails ang lumalabas, at vina-validate ng tool guardrails ang mga aksyon bago tumakbo. Kapag pumalya ang check, bina-block, binabago o fina-flag ang request. - [AI hallucination](https://broadrouter.com/fil/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): Ang hallucination ay output ng language model na kapani-paniwala pero mali, gawa-gawa o hindi suportado ng input na ibinigay dito, gaya ng inimbentong citation, maling numero o API na hindi umiiral. Nangyayari ito dahil hinuhulaan ng model ang malamang na text sa halip na mag-check ng facts. - [Hamza](https://broadrouter.com/fil/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): Ang hamza (ء) ang letra ng glottal stop sa Arabic. Isinusulat ito nang mag-isa o nakapatong sa isang upuan: alef (أ إ), waw (ؤ) o yeh (ئ), at pinagsasama ng آ ang alef at madda. Dahil hindi pare-pareho ang pag-type nito, madalas ginagawang ا ng mga search engine at speech scoring ang mga anyong may alef. - [Hybrid search](https://broadrouter.com/fil/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): Ang hybrid search ay paraan ng retrieval na nagpapatakbo ng keyword search, gaya ng BM25, at vector search sa embeddings para sa iisang query, saka pinagsasama ang dalawang ranked list sa isa, kadalasan gamit ang reciprocal rank fusion (RRF) o weighted sum ng mga normalized score. - [Idempotency key](https://broadrouter.com/fil/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Ang idempotency key ay natatanging value, kadalasang random na ID gaya ng UUID, na ikinakabit ng client sa request na nagbabago ng state, gaya ng bayad o pagbili. Kapag ipinadala ulit ang parehong request gamit ang parehong key, ibinabalik ng server ang orihinal na resulta sa halip na gawin ulit ang aksyon. - [Image tokens](https://broadrouter.com/fil/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): Ang image tokens ay mga input token na ginagamit ng vision language model para i-represent ang isang larawan sa prompt. Kinukuwenta ito ng mga provider mula sa laki ng larawan, kadalasan sa paghati nito sa mga patch o tile na pare-pareho ang laki pagkatapos i-resize, at sinisingil gaya ng text input token, kaya mas mahal ang mas malalaking larawan. - [Inference provider](https://broadrouter.com/fil/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): Ang inference provider ay kumpanyang nagpapatakbo ng machine learning models, kadalasang open-weight na LLM, sa mga GPU na pag-aari, naka-lease o inuupahan nito, at nagbebenta ng access sa mga ito sa pamamagitan ng API, kadalasang may presyo kada token, request o yunit ng oras. Ito ang humahawak sa serving, scaling at hardware. - [Input at output token (input and output tokens)](https://broadrouter.com/fil/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): Ang input token ay ang mga token na binabasa ng language model sa isang request, kasama ang system prompt, history ng usapan, tool definitions at mga larawan. Ang output token ay ang mga token na binubuo nito bilang sagot, kasama ang anumang reasoning. Kadalasang magkahiwalay ang presyo ng dalawa, at mas mahal ang output. - [INT4](https://broadrouter.com/fil/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): Ang INT4 ay 4-bit na integer number format na kayang mag-represent ng 16 na level. Sa LLM inference, kadalasan itong tumutukoy sa 4-bit weight-only quantization (W4A16): iniimbak ang weights bilang INT4 kasama ng mga scale factor at pinalalawak sa 16-bit habang nagkukuwenta, kaya mga sangkapat lang ng memory ng 16-bit na weights ang gamit. - [IP allow list](https://broadrouter.com/fil/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): Ang IP allow list ay security control na tumatanggap lang ng request mula sa mga nakalistang IP address o CIDR range at tinatanggihan ang lahat ng iba. Sa isang API, nililimitahan nito ang pinsala ng na-leak na key, dahil hindi gumagana ang key mula sa mga network na wala sa listahan. - [Jailbreak (LLM)](https://broadrouter.com/fil/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): Ang jailbreak ay input na sadyang ginawa para balewalain ng isang language model ang safety training nito o ang mga rule na itinakda ng application na gumagamit dito, para makagawa ito ng content o aksyon na karaniwan nitong tinatanggihan. Itinuturing ng OWASP ang jailbreaking na isang anyo ng prompt injection. - [JSON mode](https://broadrouter.com/fil/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): Ang JSON mode ay isang setting ng LLM API, response_format {"type": "json_object"} sa mga OpenAI-style na API, na ginagawang valid at nae-parse na JSON ang sagot ng model. Hindi tulad ng structured outputs na may JSON Schema, hindi nito ginagarantiya kung anong mga key, type o value ang laman ng object. - [JSON Schema](https://broadrouter.com/fil/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): Ang JSON Schema ay isang vocabulary, nakasulat sa JSON, para ilarawan at i-validate ang istruktura ng JSON data: mga type, required na field, pinapayagang value at nesting. 2020-12 ang kasalukuyang bersyon. Ginagamit ito ng mga LLM API para tukuyin ang tool parameters at ang eksaktong hugis na dapat sundin ng structured output. - [KV cache](https://broadrouter.com/fil/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): Ang KV cache ay ang memory kung saan iniimbak ng transformer language model ang attention keys at values ng bawat token na naproseso na nito, kaya nire-reuse ito ng bawat bagong token sa halip na kuwentahin ulit ang buong sequence. Lumalaki ito kasabay ng context length at madalas nitong nililimitahan kung ilang request ang kaya ng isang GPU. - [Latency](https://broadrouter.com/fil/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): Ang latency, para sa isang language model API, ay ang tagal ng isang request mula sa pagpapadala hanggang matanggap ang huling output token. Kabuuan ito ng time to first token at ng generation time, at kadalasang iniuulat bilang percentile gaya ng p50 at p95 sa halip na average. - [LLM-as-a-judge](https://broadrouter.com/fil/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): Ang LLM-as-a-judge ay paraan ng evaluation kung saan isang language model ang nagma-marka sa output ng ibang model, sa pag-score ng iisang sagot batay sa rubric o sa pagpili ng mas mahusay sa dalawang sagot. Malapit ito sa paghusga ng tao sa mas mababang gastos pero may mga kilalang bias, gaya ng pagpabor sa una o mas mahabang sagot. - [LLM benchmark](https://broadrouter.com/fil/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): Ang LLM benchmark ay fixed at kadalasang public na set ng mga task na may paraan ng pag-score, na ginagamit para sukatin at ikumpara ang mga language model sa iisang kakayahan, gaya ng kaalaman, coding o speech recognition. Ipinapakita ng score ang relatibong galing sa test na iyon, hindi kung paano gagana ang model sa sarili mong workload. - [LLM router](https://broadrouter.com/fil/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): Ang LLM router ay software o serbisyo na nakapuwesto sa pagitan ng isang application at ng isa o higit pang language model API, at nagpapasya, para sa bawat request, kung aling model o provider ang hahawak dito, gamit ang mga rule gaya ng gastos, latency, availability o uri ng task. - [Logprobs](https://broadrouter.com/fil/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Ang logprobs ay ang mga log probability na ibinigay ng language model sa mga token na nabuo nito. Kapag ibinabalik ito ng API, makikita mo sa bawat output token kung gaano ito kalamang sa tingin ng model at, kung gusto mo, ang pinaka-malamang na alternatibong token sa posisyong iyon. - [LoRA (low-rank adaptation)](https://broadrouter.com/fil/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): Ang LoRA (low-rank adaptation) ay parameter-efficient na paraan ng fine-tuning na nagfi-freeze sa weights ng isang pre-trained na model at nagte-train ng isang pares ng maliliit na low-rank matrix para sa mga piling layer. Puwedeng i-store ang natutunang update bilang maliit na adapter o i-merge sa weights, nang walang dagdag na inference latency. - [Lost in the middle](https://broadrouter.com/fil/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Ang lost in the middle ay ang napansing hilig ng mga language model na pinakamahusay gamitin ang mahalagang impormasyon kapag nasa simula o dulo ito ng mahabang input, at pinakamahina kapag nasa gitna. Pinangalanan ito sa isang 2023 na pag-aaral nina Liu et al. - [max_tokens](https://broadrouter.com/fil/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): Ang max_tokens ay isang parameter ng LLM API na nagtatakda ng pinakamaraming token na puwedeng isulat ng model sa isang response. Kapag naabot ang limit, humihinto ang pagbuo ng sagot at minamarkahang naputol ang response, halimbawa may finish_reason na "length". - [MCP (Model Context Protocol)](https://broadrouter.com/fil/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): Ang MCP (Model Context Protocol) ay isang open protocol na nakabatay sa JSON-RPC 2.0 na nag-i-standardise kung paano kumokonekta ang mga LLM application sa external na tools at data. Naglalantad ang MCP server ng tools, resources at prompts; hinahanap ito ng MCP client sa host application at ipinapasa sa server ang mga tool call ng model. - [Mixture of experts (MoE)](https://broadrouter.com/fil/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Ang mixture of experts (MoE) ay disenyo ng neural network kung saan may maraming expert sub-network ang isang layer at ipinapadala ng isang natutong router ang bawat token sa iilan lang sa mga ito. Malaki ang kabuuang bilang ng parameter ng isang MoE language model pero maliit na active na bahagi lang ang ginagamit kada token. - [Model deprecation](https://broadrouter.com/fil/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): Ang model deprecation ay kapag inanunsyo ng isang API provider na hihinto na ang pag-aalok ng isang model, kadalasang may binanggit na retirement o shutdown date at inirerekomendang kapalit. Pagkatapos ng petsang iyon, pumapalya ang mga request sa model, kaya kailangang lumipat ang mga application sa ibang model ID bago iyon. - [Model fallback](https://broadrouter.com/fil/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): Ang model fallback ay reliability pattern kung saan nire-retry ang pumalyang LLM request sa ibang model o provider mula sa isang nakaayos na listahan. Puwedeng i-trigger ito ng error, rate limit, context-length error o pagtanggi dahil sa content policy. Tumatakbo ito sa client mo o sa isang gateway sa harap ng mga API. - [Model license (lisensya ng model)](https://broadrouter.com/fil/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): Ang model license ay ang legal na terms kung paano puwedeng gamitin, baguhin at ibahagi ang weights, code at outputs ng isang model. Gumagamit ang mga open-weight model ng karaniwang software license gaya ng Apache 2.0 o MIT, o ng custom model license na puwedeng magdagdag ng kondisyon, gaya ng limit sa malalaking commercial service. - [Modern Standard Arabic (MSA)](https://broadrouter.com/fil/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): Ang Modern Standard Arabic (MSA) ay ang standardized na anyo ng Arabic na ginagamit sa pagsulat, balita, edukasyon at pormal na pananalita sa buong mundong Arabo. Hindi ito katutubong pang-araw-araw na wika ninuman: mga rehiyonal na dialect gaya ng Egyptian o Gulf Arabic ang sinasalita, at lumilipat sa MSA sa pormal na sitwasyon. - [Multi-agent system](https://broadrouter.com/fil/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): Ang multi-agent system ay isang LLM application kung saan ilang agent, bawat isa may sariling prompt, tools at context, ang gumagawa sa mga bahagi ng iisang task. Karaniwang disenyo ang orchestrator-worker: nagpaplano ang lead agent, nagpapatakbo ng mga worker agent nang sabay-sabay at pinagsasama ang mga resulta nila. - [Multimodal AI](https://broadrouter.com/fil/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): Ang multimodal AI ay isang model o system na gumagana sa higit sa isang uri ng data, gaya ng text, larawan, video at audio. Puwedeng sabay na basahin ng multimodal model ang isang litrato at isang tanong, o tumanggap ng speech at sumagot sa text. Ang vision-language model ang pinakakaraniwang uri. - [Named entity recognition (NER)](https://broadrouter.com/fil/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): Ang named entity recognition (NER) ay ang task ng paghahanap ng mga pangalan sa text at paglalagay ng uri sa bawat isa, gaya ng tao, organisasyon o lokasyon. Tina-tag ng mga klasikong system ang bawat token; sa mga LLM naman, madalas hinihingi ang listahan ng entities bilang JSON, na sinusuri pagkatapos laban sa source. - [Optical character recognition (OCR)](https://broadrouter.com/fil/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): Ang optical character recognition (OCR) ay teknolohiyang tumutukoy at nagko-convert ng naka-print o sulat-kamay na text sa isang larawan, scan o PDF tungo sa machine-readable na character. Text at posisyon nito sa page ang output ng klasikong OCR; ang vision-language model naman ay kayang basahin ang text at direktang sagutin ang mga tanong tungkol dito. - [Ogg Opus](https://broadrouter.com/fil/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ang Ogg Opus ay audio na naka-encode gamit ang Opus codec (RFC 6716) at naka-store sa isang Ogg container (RFC 7845). Ginawa ito para sa speech at interactive na audio sa mababang bitrate, at ito ang format na ginagamit ng WhatsApp para sa voice notes, kadalasan bilang .ogg o .opus na file. - [Open-weight model](https://broadrouter.com/fil/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): Ang open-weight model ay machine learning model na inilathala ang trained weights para i-download at patakbuhin ng kahit sino, kadalasan sa ilalim ng lisensyang may kondisyon. Iba ito sa ganap na open-source AI, na naglalabas din ng training code at detalyadong impormasyon tungkol sa training data para mapag-aralan at ma-reproduce ng iba ang model. - [OpenAI-compatible API](https://broadrouter.com/fil/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): Ang OpenAI-compatible API ay isang HTTP API na tumatanggap ng parehong request at response format gaya ng OpenAI API, halimbawa POST /v1/chat/completions, kaya magagamit ito ng mga OpenAI SDK at tool sa pagpalit lang ng base URL, API key at pangalan ng model. - [Parallel tool calls](https://broadrouter.com/fil/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Ang parallel tool calls ay kapag nagbalik ang isang model ng higit sa isang tool call sa iisang response, halimbawa tatlong pagbasa ng file nang sabay. Sa OpenAI format, dumarating ang mga ito bilang ilang entry sa tool_calls, at nililimitahan ng parallel_tool_calls na naka-false ang model sa zero o isang tawag kada turn. - [Pay-as-you-go](https://broadrouter.com/fil/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): Ang pay-as-you-go ay paraan ng pagpepresyo kung saan ang nagamit mo lang ang binabayaran mo, sinusukat kada yunit gaya ng token, request o minuto ng audio, nang walang subscription o minimum na commitment. Ganito naniningil ang karamihan ng LLM API, mula sa prepaid na balance o sa pana-panahong invoice. - [PDPL (data protection law ng Saudi at UAE)](https://broadrouter.com/fil/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): Ang PDPL (Personal Data Protection Law) ay karaniwang pangalan ng batas sa personal data ng Saudi Arabia (Royal Decree M/19, ipinapatupad ng SDAIA) at ng Federal Decree-Law No. 45 of 2021 ng UAE. Pareho nilang kinokontrol ang pagproseso ng personal data at nagtatakda ng kondisyon sa transfer sa ibang bansa, kasama ang mga prompt sa dayuhang LLM API. - [PII (personally identifiable information)](https://broadrouter.com/fil/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): Ang PII (personally identifiable information) ay impormasyong magagamit para makilala o ma-trace ang pagkakakilanlan ng isang tao, gaya ng pangalan, ID number o biometric record, pati ibang impormasyong naka-link o mali-link sa taong iyon. Personal data ang mas malawak na termino ng GDPR. Sa LLM, tinatanggal ng PII redaction ito sa mga prompt bago lumabas sa system mo. - [Prefill at decode](https://broadrouter.com/fil/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Ang prefill at decode ay ang dalawang yugto ng LLM inference. Sa prefill, pinoproseso ng model ang buong prompt nang parallel at pinupuno ang KV cache nito; sa decode, binubuo nito ang sagot nang isang token bawat hakbang, gamit ulit ang cache na iyon. Ang prefill ang halos nagtatakda ng time to first token; ang decode, ng bilis ng output. - [Prepaid credit](https://broadrouter.com/fil/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): Ang prepaid credit ay perang idinadagdag sa isang API account bago gamitin. Ibinabawas sa balance ang gastos ng bawat request, at kapag naubos ang balance, titigil ang mga metered request hanggang mag-top-up ka. Nililimitahan nito ang gastos sa halagang naibayad mo na. - [Prompt engineering](https://broadrouter.com/fil/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): Ang prompt engineering ay ang proseso ng pagsulat, pag-aayos at pag-test ng mga instruction na ibinibigay sa isang language model para palagi itong makagawa ng output na tumutugon sa mga requirement mo. Sakop nito ang mga role message, malinaw na instruction, mga example, format, at pagsukat ng resulta gamit ang mga test case. - [Prompt injection](https://broadrouter.com/fil/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): Ang prompt injection ay isang atake kung saan binabago ng text na ibinigay sa isang language model, mula sa user o nakatago sa web page, email, file o tool result, ang kilos ng model laban sa layunin ng developer. Inilalagay ito ng OWASP sa unang puwesto (LLM01:2025) sa Top 10 for LLM Applications nito. - [Quantization](https://broadrouter.com/fil/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): Ang quantization ay ang proseso ng pag-iimbak ng mga numero ng isang neural network, pangunahin ang weights nito, sa format na mas mababa ang precision gaya ng 8-bit o 4-bit sa halip na 16-bit, para mas kaunti ang memory na gamit ng model at kadalasang mas bumilis ito. Maliit lang ang nababawas sa accuracy sa mahuhusay na method. - [RAG (retrieval-augmented generation)](https://broadrouter.com/fil/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): Ang RAG (retrieval-augmented generation) ay isang pattern kung saan kinukuha muna ng application ang mga bahaging may kinalaman sa tanong mula sa isang external na store, gaya ng search o vector index, saka idinadagdag sa prompt para makasagot ang language model mula sa text na iyon, hindi lang mula sa memorya nito. - [Rate limit](https://broadrouter.com/fil/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): Ang rate limit ay hangganang itinatakda ng API sa dami ng paggamit ng isang client sa isang yugto ng oras, gaya ng request kada minuto o token kada minuto. Tinatanggihan ang mga request na lampas sa limit, kadalasan na may HTTP 429 Too Many Requests, hanggang bumaba ulit ang paggamit. - [ReAct (reason and act)](https://broadrouter.com/fil/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): Ang ReAct (reason and act) ay isang prompting pattern, ipinakilala nina Yao et al. noong 2022, kung saan pinagsasalit ng language model ang reasoning traces at mga aksyon gaya ng search o tool call. Binabasa nito ang bawat observation, ina-update ang plano, at umuulit hanggang makapagbigay ng huling sagot. - [Reasoning tokens](https://broadrouter.com/fil/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): Ang reasoning tokens ay mga token na binubuo ng isang reasoning model para pag-isipan ang isang problema bago nito isulat ang nakikitang sagot. Kadalasang sinisingil ang mga ito bilang output token, binibilang sa output limit at sa context window, at puwedeng itago o ibuod ng API. - [Request timeout](https://broadrouter.com/fil/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): Ang request timeout ay ang pinakamatagal na hihintayin ng isang HTTP client sa server bago nito iwanan ang request. Karaniwang hinahati ito ng mga client sa connect timeout para sa pagbukas ng connection at read timeout para sa paghihintay ng data. Kailangan ng maluwag na read timeout ang mga LLM call dahil matagal mabuo ang mahahabang sagot. - [Reranking](https://broadrouter.com/fil/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Ang reranking ay pangalawang yugto ng retrieval kung saan ang mas tumpak pero mas mabagal na model, kadalasang cross-encoder, ang nagbibigay ng score sa bawat pares ng query at kandidatong dokumento mula sa unang search at muling inaayos ang pagkakasunod, para iilang pinakanauugnay lang ang maipasa sa language model. - [Right-to-left (RTL) text](https://broadrouter.com/fil/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): Ang right-to-left (RTL) text ay pagsulat, gaya ng Arabic o Hebrew, na ipinapakita mula kanan pakaliwa. Iniimbak ito ng computer sa lohikal na pagkakasunod ng pagbasa, at ang Unicode Bidirectional Algorithm ang nagpapasya ng ayos sa screen kapag nahahalo ang RTL text sa left-to-right text, numero, pangalan ng model o code. - [Sample rate](https://broadrouter.com/fil/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): Ang sample rate ay ang bilang ng audio sample na kinukuha kada segundo, na sinusukat sa hertz (Hz). Karaniwang 16,000 Hz (16 kHz) na mono audio ang inaasahan ng mga speech recognition model. Kasama ng bit depth at bilang ng channel, ang sample rate ang nagtatakda kung gaano kalaki kada minuto ang hindi naka-compress na audio file. - [Seed parameter](https://broadrouter.com/fil/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): Ang seed parameter ay integer na ipinapadala mo kasama ng request para subukan ng API na kunin ang parehong mga token kapag inulit ang parehong request. Best effort lang ito ayon sa OpenAI: hindi garantisado ang determinism, at ipinapakita ng system_fingerprint field kapag nagbago ang configuration ng backend. - [Serverless inference](https://broadrouter.com/fil/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): Ang serverless inference ay paraan ng pagpapatakbo ng machine learning model kung saan awtomatikong sinisimulan, sina-scale at pinapatay ng platform ang compute sa likod ng isang endpoint, madalas hanggang zero kapag idle, at paggamit ang sinisingil sa halip na mga naka-reserve na makina. Ang kapalit ay posibleng delay dahil sa cold start. - [SLA (service level agreement)](https://broadrouter.com/fil/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): Ang SLA (service level agreement) ay pangako sa kontrata ng isang provider sa nasusukat na antas ng serbisyo, kadalasang porsiyento ng buwanang uptime, na may takdang remedyo, kadalasang service credit, kapag hindi ito naabot. Itinatakda ng SLA kung ano ang binibilang na downtime, ano ang hindi kasama at paano mag-claim. - [Speculative decoding](https://broadrouter.com/fil/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): Ang speculative decoding ay inference technique kung saan nagmumungkahi ang isang maliit at mabilis na draft model ng ilang susunod na token, at sabay-sabay itong bine-verify ng malaking target model sa iisang forward pass, saka pinapanatili ang mga tinanggap. Pinapabilis nito ang generation nang hindi binabago ang distribution ng output. - [Speech-to-text (STT)](https://broadrouter.com/fil/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): Ang speech-to-text, na tinatawag ding automatic speech recognition (ASR), ay teknolohiyang ginagawang nakasulat na text ang binigkas na audio. Tumatanggap ang speech-to-text API ng audio file o stream at nagbabalik ng transcript, kadalasang may presyo kada minuto ng audio. Sinusukat ang accuracy gamit ang word error rate, at nag-iiba ito ayon sa accent, dialect at kalidad ng recording. - [Spend limit](https://broadrouter.com/fil/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): Ang spend limit ay hangganan sa kung magkano ang puwedeng gastusin ng isang API account, project o key sa isang yugto, kadalasang isang buwan. Kapag naabot ito, tinatanggihan ng provider ang mga susunod na request, madalas na may HTTP 429 o 400 na error, hanggang mag-reset ang yugto o itaas ang limit. - [SRT subtitles](https://broadrouter.com/fil/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): Ang SRT (SubRip) ay plain-text na subtitle format. Binubuo ang bawat caption ng sequence number, oras ng simula at dulo na nakasulat bilang hh:mm:ss,mmm --> hh:mm:ss,mmm, isa o higit pang linya ng text, at isang blangkong linya. Ang WebVTT ang katulad na web format na ginagamit ng HTML video. - [Server-sent events (SSE)](https://broadrouter.com/fil/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Ang server-sent events (SSE) ay web standard kung saan nagpapadala ang server ng tuloy-tuloy na text event sa client sa iisang matagal na HTTP response na may content type na text/event-stream. Ang bawat event ay isa o higit pang field line, gaya ng data:, na tinatapos ng blank line. - [Stop sequences](https://broadrouter.com/fil/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): Ang stop sequences ay mga string na ipinapadala mo kasama ng request para tumigil ang model sa pagbuo ng sagot sa sandaling maisulat nito ang isa sa mga iyon. Sa OpenAI, stop ang field at hanggang 4 na sequence ang tinatanggap; sa Anthropic, stop_sequences. Karaniwang hindi kasama sa ibinabalik na output ang tumugmang text. - [Streaming](https://broadrouter.com/fil/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): Ang streaming ay paraan ng pagbabalik ng sagot ng language model nang paunti-unting chunk habang binubuo pa ito, sa halip na iisang sagot sa dulo. Sa mga OpenAI-compatible na API, dumadaan ang streaming sa server-sent events kapag naka-set sa true ang stream sa request. - [Structured output](https://broadrouter.com/fil/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): Ang structured output ay feature ng LLM API na naglilimita sa sagot ng model sa format na itinakda ng developer, kadalasang JSON Schema, para maaasahang ma-parse ito ng code. Mas mahigpit ito kaysa sa JSON mode, na gumagarantiya ng valid na JSON pero hindi ng partikular na hugis. - [Subagent](https://broadrouter.com/fil/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): Ang subagent ay hiwalay na agent na sinisimulan ng isang main agent para hawakan ang isang nakatutok na task. Tumatakbo ito nang may sariling context window, system prompt at access sa tools, ginagawa ang trabaho, at resulta o buod lang ang ibinabalik sa main agent, kaya nananatiling maliit ang pangunahing usapan. - [Subprocessor](https://broadrouter.com/fil/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): Ang subprocessor ay third party na kinukuha ng isang processor para magproseso ng personal data para sa controller. Sa isang LLM API, puwedeng kasama sa mga subprocessor ang mga GPU host, partner na inference provider at backup capacity. Sa ilalim ng GDPR Article 28, kailangan nila ng awtorisasyon ng controller at ng parehong obligasyon sa data protection. - [System prompt](https://broadrouter.com/fil/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): Ang system prompt ay ang instruction na ipinapadala ng isang application sa language model bago ang usapan, kadalasan bilang mensaheng may system o developer role. Itinatakda nito ang role, mga patakaran, tono at format ng output ng model, at dapat itong sundin ng model kahit salungat ang mensahe ng user. - [Taa marbuta](https://broadrouter.com/fil/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): Ang taa marbuta (ة, U+0629) ay Arabic na letrang lumalabas lang sa dulo ng salita, at madalas na tanda ng pambabaeng pangngalan at pang-uri. Kamukha ito ng heh (ه, U+0647) na may dalawang tuldok, kaya nahahalo ang dalawa sa tina-type na text, at maraming search normalizer ang nagpapalit ng ة sa ه. - [Temperature](https://broadrouter.com/fil/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): Ang temperature ay isang sampling parameter na kumokontrol kung gaano ka-random ang output ng isang language model. Hinahati ng model ang score ng bawat token sa temperature bago gawing probability: sa mababang value, nangingibabaw ang pinaka-malamang na token; sa mataas na value, pumapantay ang distribution kaya mas madalas mapili ang hindi gaanong malamang na token. - [Text-to-speech (TTS)](https://broadrouter.com/fil/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): Ang text-to-speech (TTS), o speech synthesis, ay ang paggawa ng binigkas na audio mula sa nakasulat na text. Tumatanggap ang TTS API ng text, boses at output format gaya ng MP3 o WAV, at nagbabalik ng nabuong pananalita. Kabaligtaran ito ng speech-to-text, na ginagawang text ang audio. - [Throughput](https://broadrouter.com/fil/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): Ang throughput, sa LLM serving, ay ang kabuuang trabahong natatapos ng isang system kada yunit ng oras sa lahat ng sabay-sabay na request, kadalasang sinusukat bilang output tokens per second o requests per second. Iba ito sa bilis kada stream, na kung gaano kabilis natatanggap ng iisang request ang mga token nito. - [Time to first token (TTFT)](https://broadrouter.com/fil/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): Ang time to first token (TTFT) ay ang tagal mula sa pagpapadala ng request sa isang language model hanggang dumating ang unang output token. Kasama rito ang oras sa network, pila sa server at prefill, ang hakbang kung saan binabasa ng model ang buong prompt, kaya humahaba ito kasabay ng haba ng prompt at ng load. - [Timestamps (speech-to-text)](https://broadrouter.com/fil/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): Ang mga timestamp sa speech-to-text ay oras ng simula at dulo, sa segundo, na nakakabit sa mga bahagi ng transcript. Sakop ng segment timestamp ang isang parirala o pangungusap; sakop ng word timestamp ang bawat salita. Kailangan ang mga ito para sa subtitle, paghahanap sa loob ng recording at pagtutugma ng text sa mga nagsasalita. - [Token](https://broadrouter.com/fil/glossary/token/) (translation of https://broadrouter.com/glossary/token/): Ang token ay ang yunit ng text na binabasa at isinusulat ng isang language model: isang salita, bahagi ng salita, bantas o byte sequence na ginawa ng tokenizer ng model. Sa token binibilang, nililimitahan at kadalasang sinisingil ng mga LLM API ang paggamit. - [Tokenizer](https://broadrouter.com/fil/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): Ang tokenizer ay ang bahaging ginagawang sunod ng integer token ID na binabasa ng language model ang text, at ibinabalik sa text ang mga nabuong ID. May sariling tokenizer ang bawat model, kaya puwedeng magkaiba ang bilang ng token ng iisang text sa magkakaibang model. - [Tokens per second (TPS)](https://broadrouter.com/fil/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): Ang tokens per second (TPS) ay ang bilis ng pagbuo ng language model ng mga output token, kadalasang sinusukat para sa isang request pagkatapos dumating ang unang token. Ito ang bilis ng pag-stream ng sagot kapag nagsimula na. Kapag pinagsama sa lahat ng sabay na request, nagiging throughput ito ng system; hiwalay na sinusukat ang paghihintay bago ang unang token. - [Tool calling (function calling)](https://broadrouter.com/fil/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): Ang tool calling, tinatawag ding function calling, ay feature ng LLM API kung saan nagbabalik ang model ng structured na request para magpatakbo ng function na tinukoy ng developer, kasama ang JSON arguments. Ang application ang nagpapatakbo ng function at nagpapadala pabalik ng resulta; hindi kailanman nagpapatakbo ng code ang model mismo. - [tool_choice](https://broadrouter.com/fil/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): Ang tool_choice ay request parameter na kumokontrol sa paggamit ng tool. Sa Chat Completions ng OpenAI, puwede itong auto (ang model ang magpapasya), required (kailangang tumawag ng isa o higit pang tool), none (walang tool), o isang pinangalanang function na ipipilit. Ang katumbas na option ng Anthropic ay auto, any, tool at none. - [Top-p (nucleus sampling)](https://broadrouter.com/fil/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Ang top-p, na tinatawag ding nucleus sampling, ay isang paraan ng decoding kung saan kumukuha lang ang language model ng susunod na token mula sa pinakamaliit na set ng pinaka-malamang na token na umaabot sa threshold p, gaya ng 0.9, ang pinagsamang probability, at itinatapon ang hindi malamang na buntot. - [Unicode normalization](https://broadrouter.com/fil/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Ang Unicode normalization ay ang pag-convert ng text sa isang standard na form para pareho ang code point ng mga string na magkamukha. Apat ang form: NFD, NFC, NFKD at NFKC; pinag-iisa rin ng mga K form ang mga compatibility character, gaya ng Arabic presentation forms at ligature, sa ordinaryong letra. - [Voice activity detection (VAD)](https://broadrouter.com/fil/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): Ang voice activity detection (VAD) ay isang signal-processing step na nag-uuri sa maiikling frame ng audio bilang speech o non-speech. Ginagamit ito ng mga speech system para laktawan ang katahimikan, hatiin ang mahahabang recording sa mga pause, malaman kung tapos nang magsalita ang speaker, at iwasang i-transcribe ang ingay sa background. - [Vector database](https://broadrouter.com/fil/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): Ang vector database ay isang data store na naglalaman ng mga embedding vector, kadalasang kasama ang source text at metadata, at sumasagot sa nearest-neighbour query: bigyan mo ito ng query vector at ibabalik nito ang mga naka-store na vector na pinakakahawig nito, karaniwang batay sa cosine distance, Euclidean distance o inner product. - [Vision-language model (VLM)](https://broadrouter.com/fil/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): Ang vision-language model (VLM) ay isang model na tumatanggap ng larawan at text nang sabay bilang input at naglalabas ng text, kaya kaya nitong ilarawan ang mga larawan, sagutin ang tanong tungkol dito, at basahin ang mga dokumento, chart o screenshot. Tinatawag din ang mga VLM na image-text-to-text o multimodal na model. - [vLLM](https://broadrouter.com/fil/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): Ang vLLM ay isang open-source na library para sa LLM inference at serving, na unang binuo sa Sky Computing Lab ng UC Berkeley. Pinamamahalaan nito ang KV cache memory gamit ang PagedAttention, tuloy-tuloy na bina-batch ang mga papasok na request, sumusuporta sa maraming quantization format at may OpenAI-compatible na HTTP server. - [Vocabulary size (vocab size)](https://broadrouter.com/fil/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): Ang vocabulary size ay ang bilang ng magkakaibang token sa tokenizer ng isang language model, bawat isa may sariling ID at embedding. Kayang i-encode ng mas malaking vocabulary ang karaniwang text sa mas kaunting token, pero pinalalaki nito ang embedding at output layer, at nakakatulong lang ito sa text na pinag-trainan ng vocabulary. - [WAV](https://broadrouter.com/fil/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): Ang WAV (Waveform Audio File Format, o WAVE) ay isang audio container na nakabatay sa RIFF format ng Microsoft at IBM. Kadalasang uncompressed na linear PCM sample ang laman nito, kaya predictable ang laki: sample rate na pinarami sa bytes per sample at sa bilang ng channel, bawat segundo. - [Word error rate (WER)](https://broadrouter.com/fil/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): Ang word error rate (WER) ay ang karaniwang sukatan ng accuracy sa speech recognition: ang bilang ng substitution, deletion at insertion ng salita na kailangan para gawing tamang reference ang isang transcript, hinati sa bilang ng salita sa reference. Mas mababa, mas maganda. - [Zero data retention (ZDR)](https://broadrouter.com/fil/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): Ang zero data retention (ZDR) ay patakaran kung saan pinoproseso lang ng AI provider ang mga prompt at output para i-serve ang isang request, at hindi na nito itinatago ang content na iyon pagkatapos, sa log, cache o training data. Kadalasang itinatago pa rin ang mga record na walang content, gaya ng bilang ng token at timing. ### Malay (ms, Bahasa Melayu): 294 translated pages - [Articles in Malay](https://broadrouter.com/ms/blog/) - [Perubahan harga API LLM Oktober 2026: 3 kenaikan, tiada penurunan](https://broadrouter.com/ms/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 perubahan halaman harga di 5 daripada 7 penyedia API LLM pada Oktober 2026: 3 kenaikan, tiada penurunan, 9 model baharu dan 11 dibuang. Pautan arkib disertakan. - [Berapakah kos satu larian ejen pengekodan AI: langkah, konteks, cache](https://broadrouter.com/ms/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): Mengapa larian ejen pengekodan makin mahal setiap langkah apabila makin lama, bagaimana cache mengubah jumlahnya, dan bila sesi kadar tetap lebih murah. Ada kiraan. - [Tambah baris pada teks Arab dengan LLM: prompt, semakan dan DER](https://broadrouter.com/ms/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): Pulihkan baris (tashkeel) bahasa Arab dengan API LLM: prompt yang mengekalkan setiap huruf, semakan yang menolak suntingan, skor DER dan WER, dan bila guna CATT. - [Pembetulan tatabahasa dan ejaan bahasa Arab dengan LLM: panduan](https://broadrouter.com/ms/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Betulkan ejaan dan tatabahasa Arab dengan API LLM: prompt suntingan minimum, kawalan terhadap tulis semula, suntingan mengikut kelas ralat dan pemarkahan gaya M2. - [Minit mesyuarat bahasa Arab daripada rakaman: saluran API](https://broadrouter.com/ms/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Tukar rakaman mesyuarat bahasa Arab kepada minit dengan keputusan dan item tindakan: transkripsi, bersihkan, ringkaskan ke JSON dengan LLM, semak nama dan tarikh. - [Pengekstrakan entiti bernama bahasa Arab dengan LLM: JSON dan tarikh](https://broadrouter.com/ms/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Ekstrak nama, organisasi, tempat dan tarikh Hijrah dalam teks Arab ke JSON dengan LLM: prompt, semakan rentang tanpa diakritik, penukaran tarikh dan skrip skor. - [Perbandingan API OCR Arab: Google, Azure, Textract, Mistral, LLM](https://broadrouter.com/ms/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): API OCR mana yang rasmi menyokong bahasa Arab, bercetak dan tulisan tangan, apa yang dipulangkan dan kos setiap 1,000 halaman, berbanding API LLM penglihatan. - [Sari kata bahasa Arab (SRT) daripada API speech-to-text: keperluannya](https://broadrouter.com/ms/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): Keperluan fail SRT, API speech-to-text mana yang memberi cap masa atau SRT untuk bahasa Arab, dan saluran teruji untuk sari kata daripada API teks sahaja. - [Pengelasan teks bahasa Arab dengan LLM: sentimen, niat, triaj](https://broadrouter.com/ms/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): Kelaskan ulasan, mesej dan tiket sokongan bahasa Arab dengan API LLM: reka bentuk label, prompt peka dialek, output JSON, ujian macro F1 dan kos pada volum tinggi. - [Penormalan teks Arab untuk carian dan RAG: apa diubah, apa tidak](https://broadrouter.com/ms/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Tashkeel, alef, ta marbuta, hamzah, huruf Parsi, digit dan Unicode: peraturan penormalan Arab mana yang membantu carian, mana yang merosakkan, dan kod Python teruji. - [Peringkasan teks bahasa Arab dengan LLM: prompt, panjang dan skor](https://broadrouter.com/ms/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): Ringkaskan artikel, laporan dan mesej dialek Arab dengan API LLM: corak prompt, kawalan bilangan perkataan, semakan nombor, dan mengapa ROUGE lalai memberi 0. - [Alternatif Artificial Analysis: tempat membandingkan LLM pada 2026](https://broadrouter.com/ms/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Alternatif Artificial Analysis: Arena, Epoch AI, Vals AI, kedudukan OpenRouter, Hugging Face dan data harga, mengikut apa yang diukur dan bila setiap satu sesuai. - [Perkakasan GPU dan pendekatan penyajian BroadNet, menurut BroadNet](https://broadrouter.com/ms/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): Apa kata BroadNet tentang perkakasan di sebalik BroadNet Inference: tiga peringkat GPU dari B300 hingga peranti klien, daya pemprosesan GPU dan pengajaran ujian. - [Cache respons API LLM untuk jimat kos: cache padanan tepat Python](https://broadrouter.com/ms/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Cache padanan tepat yang diuji untuk respons API LLM dalam Python: isi kunci, apa yang tidak boleh di-cache, TTL, privasi dan cara mengukur kadar hit dahulu. - [Claude Code dengan model terbuka: proksi, endpoint asli dan had](https://broadrouter.com/ms/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code menggunakan Anthropic Messages API. Cara menjalankannya pada model terbuka melalui proksi LiteLLM atau endpoint asli, dan apa yang rosak sepanjang jalan. - [Jadikan semakan manusia sebahagian aliran kerja AI dokumen anda](https://broadrouter.com/ms/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): Reka langkah semakan manusia dalam saluran AI dokumen: peraturan pengesahan, apa yang dihantar kepada manusia, sampel audit rawak, jejak audit dan kadar semakan. - [Undang-undang perlindungan data Mesir (Law 151 of 2020) dan API LLM](https://broadrouter.com/ms/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Data peribadi rakyat Mesir ke API LLM atau speech-to-text di luar negara: kata Law 151 of 2020 dan Decision 816 of 2025 tentang pemindahan, lesen dan pelanggaran. - [Sebelum penalaan halus: tentukan masalah yang hendak diselesaikan](https://broadrouter.com/ms/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): Bila prompt, RAG atau penilaian menyelesaikan masalah tanpa penalaan halus, data latihan yang diperlukan, dan cara menguji fine-tune sebelum memintanya. - [Cara membaiki JSON tidak sah daripada LLM: punca, baiki, cuba semula](https://broadrouter.com/ms/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): Mengapa LLM memulangkan JSON rosak (pagar kod, prosa, terpotong pada max_tokens, koma di hujung) dan fungsi Python teruji yang membaiki yang selamat sahaja. - [Berapa kerap harga API LLM berubah? 3 tahun halaman harga berarkib](https://broadrouter.com/ms/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): Kami membaca 308 halaman harga berarkib daripada 11 penyedia API LLM, 2023–2026. 80% model tidak pernah berubah harga; DeepInfra membuat 75% perubahan harga senarai. - [Cara mengurangkan kos API LLM: senarai semak 10 langkah, ikut urutan](https://broadrouter.com/ms/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): Kurangkan bil API LLM dengan urutan yang betul: ukur, cache prefix, ringkaskan prompt, hadkan output, pilih model yang sesuai, batch, dan bila kadar tetap berbaloi. - [Failover API LLM: beralih ke penyedia kedua dalam kod anda sendiri](https://broadrouter.com/ms/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): Cara failover antara penyedia LLM serasi OpenAI: ralat yang patut mencetuskan failover, pemutus litar, pemetaan ID model, penstriman, kos dan privasi. - [Penurunan harga API LLM 2023–2026: yang terbesar, dengan sumber](https://broadrouter.com/ms/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): Penurunan harga API LLM terbesar 2023–2026, dengan pautan Internet Archive sebelum dan selepas: o3, Codestral, Gemini 1.5, serta kenaikan yang jarang disedari. - [Adakah harga API LLM hanya turun? 52 kenaikan dan cara membajet](https://broadrouter.com/ms/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): Harga API LLM memang naik: 52 kenaikan harga senarai bagi model yang sama dalam halaman harga berarkib, 2024–2026, median 40%. Semua sumber dan peraturan bajet. - [Had kadar API LLM mengikut penyedia: RPM, TPM dan keserentakan](https://broadrouter.com/ms/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): Had kadar peringkat permulaan 13 penyedia API LLM menurut dokumentasi mereka: OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks dll., serta maksud RPM dan TPM. - [Perbandingan SLA API LLM: janji Azure, Bedrock, Google dan OpenAI](https://broadrouter.com/ms/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): Komitmen bertulis setiap penyedia API LLM: peratusan masa operasi, apa yang dikira sebagai gangguan, kredit perkhidmatan dan sasaran kependaman, dari halaman SLA. - [Pemprosesan kelompok LLM tanpa Batch API: pelari boleh sambung](https://broadrouter.com/ms/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Jalankan 10,000 hingga 1,000,000 prompt melalui API chat-completions tanpa Batch API: had serentak, kadar permintaan, cubaan semula, checkpoint dan kiraan masa. - [The Safety Map: bagaimana 39 LLM mengendalikan data sensitif](https://broadrouter.com/ms/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet meninjau 39 LLM daripada 14 makmal: kelayakan, PII, data kesihatan dan kewangan, apa yang berpindah, apa yang tidak, dan kebocoran tool call dan penaakulan. - [Cara menguji keselamatan data sensitif LLM: kaedah BroadNet](https://broadrouter.com/ms/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): Kaedah BroadNet untuk menguji sama ada LLM membocorkan kelayakan, PII, data kesihatan atau kewangan sebelum digunakan: kategori, bocor vs terlepas, 3+ larian. - [Berpindah dari API OpenAI: panduan kod dalam Python, Node, cURL](https://broadrouter.com/ms/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): Perubahan kod untuk berpindah dari OpenAI ke API serasi OpenAI: persediaan klien, Responses ke Chat Completions, tools, penstriman, embeddings dan ujian asap. - [Model yang dinilai dan ditolak oleh BroadNet, dan sebabnya](https://broadrouter.com/ms/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet menerbitkan model yang diuji dan ditolaknya, dengan sebab: model chat yang gagal get keselamatan kelayakan; model pertuturan, OCR, embedding yang kalah. - [Mereka bentuk pengalaman AI untuk pelbagai bahasa](https://broadrouter.com/ms/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): Reka produk AI untuk pengguna bahasa Arab, Inggeris dan Sepanyol: lokal UI lwn bahasa model, RTL dan teks bercampur, nombor, tarikh, pengesanan dan kos. - [Kos API model open-weight vs tertutup: apa kata data harga](https://broadrouter.com/ms/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): Adakah model open-weight lebih murah melalui API berbanding GPT, Claude atau Gemini? Harga setiap token daripada penyedia, jurang hos dan kos setiap pengguna. - [Sejarah harga API OpenAI vs Anthropic vs Google, 2023–2026](https://broadrouter.com/ms/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): Harga API GPT, Claude dan Gemini bersebelahan daripada halaman harga berarkib, 2023–2026: barisan mana lebih murah, mana lebih mahal, setiapnya dengan pautan arkib. - [Base URL API serasi OpenAI: 14 penyedia, telah disemak](https://broadrouter.com/ms/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): base_url bagi 14 API LLM serasi OpenAI, dibaca daripada dokumentasi setiap penyedia, berserta endpoint dan keanehan yang merosakkan pertukaran terus. - [OpenClaw dan Hermes Agent pada model murah: persediaan dan kos token](https://broadrouter.com/ms/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): Sambungkan OpenClaw atau Hermes Agent ke endpoint serasi OpenAI, fahami mengapa semakan berjadual membakar token, dan kurangkan bil: konteks ringan, model murah. - [Endpoint persendirian bermula dengan beban kerja yang jelas](https://broadrouter.com/ms/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): Apa yang perlu ditakrifkan sebelum meminta endpoint LLM persendirian: beban kerja, token sesaat, keserentakan, konteks, rantau, lesen, bajet, dan cara mengujinya. - [Prompt injection ejen pengekodan: pertahanan yang benar-benar teguh](https://broadrouter.com/ms/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): Di mana prompt injection sampai ke ejen pengekodan, dan pertahanan berlapis yang didokumenkan vendor dan penyelidik: sandbox, had egress, senarai benar, kelulusan. - [Saiz chunk dan overlap RAG: kesannya pada retrieval dan kos](https://broadrouter.com/ms/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): Cara saiz chunk dan overlap mengubah recall RAG, saiz indeks, kos embedding dan token setiap pertanyaan ke LLM, dengan hasil yang diterbitkan dan kiraan sebenar. - [Menetapkan max_tokens untuk model penaakulan: kaedah bajet](https://broadrouter.com/ms/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Mengapa model penaakulan memulangkan jawapan kosong pada max_tokens rendah, dan cara menetapkan bajet daripada persentil diukur, tetingkap konteks dan cubaan semula. - [Rakam audio dalam pelayar dan hantar ke API speech-to-text](https://broadrouter.com/ms/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): Rakam mikrofon dengan MediaRecorder, tukar WebM atau MP4 Safari kepada WAV 16 kHz, dan hantar ke API transkripsi melalui pelayan kecil yang menyimpan kunci. - [Kurangkan token prompt: penjimatan terukur untuk system prompt, tools](https://broadrouter.com/ms/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): Penjimatan token terukur daripada memangkas system prompt, contoh few-shot, skema tool, format data, sejarah chat dan baris Arab, dikira dengan tokenizer sebenar. - [Model sama, hos berbeza: sejauh mana harga API LLM berbeza](https://broadrouter.com/ms/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): Bagi 119 model yang dijual beberapa hos, harga senarai termahal ialah median 1.76 kali ganda yang termurah. Data daripada halaman harga 41 penyedia, November 2026. - [PDPL Arab Saudi dan UAE serta API LLM: senarai semak rentas sempadan](https://broadrouter.com/ms/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Menghantar data peribadi ke API LLM dari Arab Saudi atau UAE: kata peraturan pemindahan PDPL, soalan untuk penyedia, dan bila inferens dalam negara penting. - [Teks, ucapan atau penglihatan? Mulakan dengan tugas](https://broadrouter.com/ms/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): Pilih chat teks, input imej atau transkripsi audio mengikut tugas, dengan model dan had yang didokumenkan BroadNet, serta apa yang tiada, seperti TTS. - [Transkripsi audio panjang melebihi 25 MB: pecah, hantar, cantum](https://broadrouter.com/ms/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): Cara mentranskripsi rakaman yang melebihi had fail API speech-to-text: kecilkan fail, pecah pada senyap dengan ffmpeg, hantar serentak dan cantum teksnya. - [Terjemah rentetan aplikasi ke bahasa Arab: pemegang tempat dan jamak](https://broadrouter.com/ms/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): Terjemah fail JSON i18n ke bahasa Arab dengan API LLM: bekukan pemegang tempat, jana enam bentuk jamak Arab, sahkan ICU dalam CI dan asingkan nama yang disisip. - [Terjemah audio bahasa Arab ke teks Inggeris dengan API: saluran diuji](https://broadrouter.com/ms/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): Tukar rakaman bahasa Arab kepada teks Inggeris dengan dua panggilan API: transkripsi dengan language=ar, kemudian terjemah dengan LLM, dengan semakan nombor dan kos. - [Live, atas permintaan, penyelidikan: cara membaca katalog model](https://broadrouter.com/ms/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): Cara mengetahui model katalog yang boleh dipanggil hari ini, yang perlu pertanyaan dan yang hanya diuji, dengan katalog BroadNet dan /v1/models sebagai contoh. - [Uno dalam vLLM: apa yang diukur oleh kajian penyajian BroadNet](https://broadrouter.com/ms/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet membina penyesuai Uno self-drafting dalam vLLM dan mengukurnya pada H100: 2.58x daya pemprosesan biasa pada satu strim, 1.74x pada 32. Kaedah, DOI dan had. - [Uno pada Gemma 4 26B: kajian mixture-of-experts BroadNet](https://broadrouter.com/ms/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet melatih penyesuai Uno self-drafting pertama untuk model mixture-of-experts, Gemma 4 26B A4B. Apa yang diukur, bila ia berbaloi, dan DOInya. - [Bajet token ejen: ke mana token gelung ejen pergi](https://broadrouter.com/ms/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): Prompt sistem, skema alat, bacaan fail, sejarah dan penaakulan: cara gelung ejen membelanjakan token, cara mengukur setiap bahagian dan apa yang dipangkas dahulu. - [Kadar ralat perkataan (WER) ASR bahasa Arab: penjelasan](https://broadrouter.com/ms/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): Maksud kadar ralat perkataan, cara mengiranya, dan mengapa WER bahasa Arab sensitif terhadap diakritik, ejaan dan dialek. Dengan Python dan angka sebenar. - [Pengecaman pertuturan dialek Arab: Teluk, Mesir, Levant, Maghreb](https://broadrouter.com/ms/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Mengapa speech-to-text bergelut dengan bahasa Arab Teluk, Mesir, Levant dan Maghreb, dan cara membina set ujian dialek untuk menilai penyedia pada audio anda. - [Terjemahan Arab–Inggeris dengan LLM: prompt, glosari, penilaian](https://broadrouter.com/ms/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): Terjemah bahasa Arab ke Inggeris dan sebaliknya dengan API LLM: prompt yang mengekalkan nada, glosari istilah tetap, nombor dan nama, serta cara menilai kualiti. - [Bahasa Arab Teluk, Mesir dan Levant: cara LLM mengendalikan dialek](https://broadrouter.com/ms/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): Cara LLM mengendalikan bahasa Arab Teluk, Mesir dan Levant: di mana ia gagal, cara prompt untuk dialek, cara mengujinya, dan laporan BroadNet bagi Qwen3.8 27B. - [OCR bahasa Arab vs LLM vision: membaca dokumen dan invois Arab](https://broadrouter.com/ms/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): Bila guna OCR klasik dan bila LLM vision untuk invois, resit dan borang bahasa Arab: perangkap ketepatan, output JSON, kos setiap halaman dan saluran pengesahan. - [Kejuruteraan prompt bahasa Arab: laras, RTL, angka dan teks bercampur](https://broadrouter.com/ms/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): Tulis prompt yang menghasilkan bahasa Arab yang baik daripada LLM: pilih laras dan dialek, kawal angka dan tarikh, format RTL yang kemas dan campuran Arab–Inggeris. - [RAG bahasa Arab: penormalan, chunking dan carian yang berkesan](https://broadrouter.com/ms/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): RAG atas dokumen bahasa Arab: normalkan alif, ya dan ta marbuta, pecahkan teks dengan munasabah, pilih embedding berbilang bahasa dan tambah carian kata kunci. - [API speech-to-text bahasa Arab: cara memilih pada 2026](https://broadrouter.com/ms/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): Cara memilih API transkripsi bahasa Arab: kadar ralat perkataan, dialek, had fail, harga sejam audio dan privasi, serta pelan ujian untuk rakaman anda sendiri. - [Seni bina chatbot WhatsApp bahasa Arab: nota suara hingga balasan](https://broadrouter.com/ms/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): Seni bina berfungsi untuk pembantu WhatsApp bahasa Arab: webhook, baris gilir, transkripsi nota suara, balasan LLM, dasar dan kos, dengan kod Python. - [Arabizi dan LLM: mengendalikan Franco-Arab dengan nombor 3 dan 7](https://broadrouter.com/ms/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): Arabizi menulis bahasa Arab dalam huruf Rumi dengan digit seperti 3 untuk ع dan 7 untuk ح. Cara mengesannya, menukarnya ke skrip Arab dengan LLM, dan membalas. - [LLM terbaik untuk Aider: cara memilih model dan menyambungkannya](https://broadrouter.com/ms/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Cara memilih LLM terbaik untuk Aider: disiplin format suntingan, saiz konteks dan kos sejam, ujian sejam pada repo anda, dan persediaan OPENAI_API_BASE yang tepat. - [Model terbuka terbaik untuk ejen pengekodan: alat, konteks dan kos](https://broadrouter.com/ms/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B atau Gemma 4 untuk ejen pengekodan anda? Pilih mengikut panggilan alat, saiz konteks, penaakulan dan kos setiap tugas. - [Token input cache: cara mengurangkan bil API LLM](https://broadrouter.com/ms/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): Input cache berharga sebahagian kecil input biasa. Cara prompt caching dibilkan, cara membaca cached_tokens dalam usage, dan cara menyusun prompt untuk mengenainya. - [API LLM penglihatan termurah: kos setiap imej dan setiap halaman](https://broadrouter.com/ms/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): Kos panggilan LLM penglihatan setiap imej dan halaman dokumen: kiraan token untuk Gemma 4, Qwen3.8 27B dan DeepSeek V4.1 Flash, serta di mana DeepInfra lebih murah. - [Cara memilih penyedia API LLM: senarai semak 20 perkara](https://broadrouter.com/ms/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): Dua puluh soalan sebelum memilih API LLM: harga, had, keserasian OpenAI, privasi, model, sokongan dan masa operasi, serta helaian skor dan pelan ujian. - [Memilih API LLM untuk bahasa Arab: apa yang perlu diuji](https://broadrouter.com/ms/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): Cara memilih LLM untuk aplikasi bahasa Arab: kefahaman dialek, kualiti penulisan, terjemahan, dokumen dan imej Arab, kos, serta set ujian yang siap dalam sehari. - [Cline vs Roo Code vs Aider untuk pasukan pada model terbuka (2026)](https://broadrouter.com/ms/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (kini diarkibkan, diteruskan sebagai Zoo Code) dan Aider dibandingkan untuk model terbuka: panggilan alat, konteks, MCP, git dan kelulusan. - [Cara mengira token sebelum menghantar permintaan LLM](https://broadrouter.com/ms/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): Anggarkan token prompt mengikut bahasa dan model sebelum memanggil API LLM: mengapa tokenizer berbeza, ukuran dalam lima bahasa, dan margin max_tokens yang selamat. - [10 cara mengurangkan kos LLM ejen pengekodan, dengan anggaran jimat](https://broadrouter.com/ms/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Sepuluh cara konkrit untuk mengurangkan perbelanjaan token Cline, Aider, opencode atau ejen pengekodan anda sendiri, dengan kiraan penjimatan dan cara mengukurnya. - [Alternatif API DeepSeek: 5 cara memanggil model DeepSeek](https://broadrouter.com/ms/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): Memanggil model DeepSeek melalui API rasmi, agregator, katalog awan, GPU sendiri atau BroadNet Inference: pertukaran dari segi harga, privasi dan had. - [Harga API DeepSeek V4.1 Flash: setiap token, cache dan pelan](https://broadrouter.com/ms/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): Harga API DeepSeek V4.1 Flash setiap 1M token input, cache dan output, pelan tetap bulanan, mingguan dan sejam, berbanding DeepSeek, DeepInfra dan Fireworks. - [Cara menganggar kos API LLM sebelum melancarkan ciri baharu](https://broadrouter.com/ms/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): Model kos langkah demi langkah untuk ciri AI baharu: pengguna × tindakan × panggilan × token, dengan cache, penaakulan, cubaan semula, had kadar dan contoh bulanan. - [Cara menilai LLM sebelum bertukar penyedia](https://broadrouter.com/ms/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): Bina set eval daripada trafik sebenar, beri skor dengan semakan kod dan hakim LLM yang dinyahbias, dan bandingkan kualiti, kos serta kependaman sebelum berhijrah. - [Ekstrak data invois, resit dan borang ke JSON dengan LLM visi](https://broadrouter.com/ms/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): Saluran paip produksi untuk invois, resit dan borang: penyediaan imej, prompt skema JSON, pengesahan Pydantic, semakan aritmetik, cubaan semula dan semakan manusia. - [API LLM kadar tetap vs bayar setiap token: kiraan pulang modal](https://broadrouter.com/ms/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): Bilakah pelan LLM kadar tetap lebih murah daripada bayar setiap token? Empat contoh bil bulanan DeepSeek V4.1 Flash, volum pulang modal dan cara mengukur volum anda. - [Kredit API LLM percuma: sejauh mana kredit permulaan sebenarnya](https://broadrouter.com/ms/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): Apa yang dibeli kredit API LLM percuma: token, minit audio dan kunci sejam pada kredit permulaan BroadNet, berbanding peringkat percuma Gemini dan OpenRouter. - [GDPR dan API LLM: 15 soalan untuk ditanya kepada penyedia anda](https://broadrouter.com/ms/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): Senarai semak GDPR API LLM: DPA, subpemproses, penyimpanan, latihan, pemindahan antarabangsa dan notis pelanggaran, dengan perkara GDPR di sebalik setiap soalan. - [GLM 5.3 vs DeepSeek V4.1 Flash: yang mana patut digunakan, dan bila](https://broadrouter.com/ms/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 vs DeepSeek V4.1 Flash dari segi harga, konteks, tools, output berstruktur, imej dan penaakulan, dengan contoh kos dan peraturan mudah untuk memilih. - [Cara mengendalikan ralat 429 API LLM: backoff, jitter dan pengehad](https://broadrouter.com/ms/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 vs 5xx, backoff eksponen dengan full jitter, pengehad keserentakan dan baris gilir untuk API LLM, dengan kod Python dan Node yang boleh disalin ke pengeluaran. - [Kunci API LLM sejam: harga rata setiap jam, diterangkan](https://broadrouter.com/ms/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): Kunci sejam BroadNet membeli satu jam Gemma 4 VL 26B pada satu strim dengan harga tetap. Apa yang termasuk, hadnya, untuk siapa dan bila ia berbaloi. - [Pengekalan data API LLM: OpenAI, Anthropic, Gemini, Mistral](https://broadrouter.com/ms/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): Apa kata OpenAI, Anthropic, Gemini API, Mistral dan BroadNet secara rasmi tentang menyimpan prompt API dan melatih model dengannya, dipetik pada November 2026. - [Keselamatan kunci API LLM: penyimpanan, putaran dan senarai IP](https://broadrouter.com/ms/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): Jauhkan kunci API LLM daripada kod dan log, beri setiap beban kerja kunci sendiri, putar ikut jadual, kunci pada IP pelayan, batalkan kebocoran dalam beberapa saat. - [Harga API LLM dijelaskan: input, cache, output dan penaakulan](https://broadrouter.com/ms/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): Cara harga API LLM setiap token berfungsi: input, input cache, output, token penaakulan, imej dan audio seminit, dengan contoh pengiraan daripada senarai kadar. - [Menghantar data sensitif ke API LLM: sunting, kurangkan, kurangkan log](https://broadrouter.com/ms/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): Cara menghantar data pelanggan, kesihatan atau kewangan ke API LLM dengan selamat: kelaskan medan, sunting dengan pemegang tempat, sifar pengekalan, log bersih. - [Kawalan perbelanjaan API LLM: bajet, had kunci dan perlindungan gelung](https://broadrouter.com/ms/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): Halang bil API LLM daripada melambung: baki prabayar, satu kunci setiap aplikasi, senarai IP dibenarkan, log penggunaan, bajet dalam kod, had gelung, harga tetap. - [Log penggunaan API LLM: kos setiap pengguna dan ciri, PII disunting](https://broadrouter.com/ms/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): Log setiap panggilan API LLM dengan token, kos, kependaman dan ID permintaan, kaitkan kos dengan pengguna dan ciri, dan sunting PII sebelum ditulis. Kod Python. - [Kos LLM setiap pengguna: ekonomi unit ciri AI SaaS](https://broadrouter.com/ms/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): Kira kos ciri AI bagi setiap pengguna aktif sebulan, mengapa pengguna berat lebih penting daripada purata, dan cara menetapkan harga peringkat SaaS. - [Harga model terbuka di penyedia inferens: 6 dibandingkan](https://broadrouter.com/ms/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): Cara OpenRouter, Together AI, Fireworks, Groq, DeepInfra dan BroadNet meletakkan harga model terbuka: setiap token, cache, batch, jam GPU dan pelan kadar tetap. - [Kependaman LLM: TTFT vs token sesaat vs masa hujung ke hujung](https://broadrouter.com/ms/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): Apa yang diukur oleh TTFT, token sesaat dan kependaman hujung ke hujung, cara mengukur setiap satu dari klien dengan kod, dan pembaikan yang sesuai bagi setiap satu. - [Penstriman LLM diterangkan: chunk SSE, delta, usage dan pembatalan](https://broadrouter.com/ms/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Cara penstriman chat completion berfungsi, membina semula jawapan daripada delta, membaca usage dalam chunk terakhir dan membatalkan dalam Python, Node dan curl. - [Konteks panjang vs RAG: bila 1M token menggantikan DB vektor](https://broadrouter.com/ms/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): Bila tetingkap konteks 1M token boleh menggantikan pangkalan data vektor dan bila RAG masih menang: kos setiap pertanyaan, kependaman, recall, penyelenggaraan. - [Pengurusan konteks untuk ejen berjalan lama dengan tetingkap 1M](https://broadrouter.com/ms/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): Cara memastikan sesi ejen panjang tepat dan berpatutan dengan tetingkap 1M token: sejarah tambah sahaja, prompt mesra cache, bila memampatkan, apa dibuang dahulu. - [Ejen yang berjalan lama: tamat masa, cubaan semula dan checkpoint](https://broadrouter.com/ms/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): Pastikan ejen LLM berjalan berjam-jam: tamat masa setiap panggilan, penstriman untuk kesan terhenti, cubaan semula selamat, alat idempoten dan checkpoint. - [Gunakan pelayan MCP dengan model serasi OpenAI (mana-mana penyedia)](https://broadrouter.com/ms/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP berfungsi dengan mana-mana LLM: klien menukar tool MCP kepada function call. Cara guna MCP dengan model terbuka dalam Cline, Continue, opencode, n8n atau Python. - [Bila tetingkap konteks 1M token berbaloi (dan bila tidak)](https://broadrouter.com/ms/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): Apa yang muat dalam 1,048,576 token, kosnya pada DeepSeek V4.1 Flash, cara penaakulan dan cache mengubah kiraan, dan bila RAG masih menjadi pilihan lebih baik. - [Kos sebenar persediaan berbilang ejen: setiap token vs sesi tetap](https://broadrouter.com/ms/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): Sistem berbilang ejen membakar beberapa kali ganda token satu ejen. Kiraan sesi kadar tetap vs setiap token, pengganda tersembunyi, dan bila setiap satu menang. - [API speech-to-text berbilang bahasa: 14 bahasa, satu endpoint](https://broadrouter.com/ms/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): Transkripsi bahasa Inggeris, Sepanyol, Perancis, Cina, Arab dan 9 lagi melalui satu endpoint serasi OpenAI berharga seminit, dan bila guna model bahasa Arab. - [n8n AI Agent pada LLM murah: persediaan, pilihan model, kos larian](https://broadrouter.com/ms/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): Bina aliran kerja n8n AI Agent pada model serasi OpenAI kos rendah: model untuk alat, tetapan yang mengurangkan token, dan kiraan kos setiap 10,000 larian. - [Model AI open-weight vs tertutup bagi perniagaan: kos, kawalan, risiko](https://broadrouter.com/ms/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): LLM open-weight atau tertutup untuk perniagaan anda? Bandingkan kos, kawalan, privasi, jurang kualiti dan kebergantungan vendor, serta tiga cara menjalankannya. - [Beralih ke API serasi OpenAI: senarai semak migrasi](https://broadrouter.com/ms/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): Menukar base_url ialah bahagian yang mudah. ID model, token penaakulan, penggunaan strim, ralat, had kadar dan kunci yang selalu rosak. Senarai semak dengan kod. - [OpenRouter vs pelan kadar tetap untuk ejen AI: mana lebih murah?](https://broadrouter.com/ms/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter mengebil setiap token serta yuran kredit; pelan tetap mengebil setiap sesi. Kos ejen sentiasa aktif, yuran dan had OpenRouter, dan mana yang sesuai. - [Menjalankan ejen pengekodan semalaman tanpa pengawasan dengan selamat](https://broadrouter.com/ms/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): Persediaan praktikal untuk ejen pengekodan tanpa pengawasan: sandbox, checkpoint git, fail tugas, had langkah, semakan pagi dan sesi pelan tetap pada waktu malam. - [Cara menyusun prompt supaya prefix di-cache](https://broadrouter.com/ms/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): Cara menyusun prompt LLM supaya prefix yang di-cache kekal sama: system prompt, skema tool, dokumen, sejarah tambah-sahaja, serta ujian yang membuktikannya. - [Token penaakulan: mengapa output yang dibayar melebihi yang dilihat](https://broadrouter.com/ms/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): Model penaakulan membilkan pemikiran tersembunyi sebagai output. Cara mengukur, kos setiap model dan cara mengurangkannya: reasoning_effort, max_tokens, penghalaan. - [Cara menjalankan ejen pengekodan selari pada API LLM kadar tetap](https://broadrouter.com/ms/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Sub-ejen, orkestrator dan pekerja kelompok pada DeepSeek V4.1 Flash tanpa had. Bilangan sesi diperlukan, pembahagian antara kunci, dan had pertindihan. - [Hos LLM sendiri vs API: kos sebenar dan titik pulang modal](https://broadrouter.com/ms/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): Harga sewa GPU daripada halaman rasmi, formula kos setiap 1M token pada GPU sendiri, penggunaan yang diperlukan untuk menewaskan API, dan kos operasi yang dilupakan. - [Harga API speech-to-text dibandingkan: harga seminit pada 2026](https://broadrouter.com/ms/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): Harga seminit dan setiap 1,000 jam bagi OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq dan BroadNet Transcribe daripada halaman rasmi, serta add-on. - [JSON mode vs json_schema: output berstruktur yang boleh diharap](https://broadrouter.com/ms/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON mode memberi JSON yang sah; json_schema memberi bentuk anda. Cara memintanya pada model serasi OpenAI, mengesahkan dengan Pydantic, dan membaiki output rosak. - [Token setiap perkataan Arab vs Inggeris: mengapa Arab lebih mahal](https://broadrouter.com/ms/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): Mengapa bahasa Arab sering guna lebih banyak token LLM daripada Inggeris bagi makna sama, mengapa token setiap perkataan mengelirukan, dan skrip untuk mengukurnya. - [Tool calling pada model terbuka: skema, panggilan selari, cuba semula](https://broadrouter.com/ms/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): Function calling pada model terbuka yang serasi OpenAI: tip skema, panggilan alat selari, lima punca kegagalan, dan gelung Python yang mengesah dan mencuba semula. - [Cara mentranskripsi rakaman pusat panggilan: saluran sedia bahasa Arab](https://broadrouter.com/ms/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): Pisahkan panggilan stereo mengikut saluran, potong giliran ucapan, transkripsi bahasa Arab dan Inggeris, sunting PII dan ringkaskan dengan LLM. Kod Python dan kos. - [Model terbuka dalam ejen pengekodan: Aider, Cline, opencode, Codex](https://broadrouter.com/ms/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Halakan Aider, Cline, Roo Code, opencode, Continue dan Codex CLI ke endpoint serasi OpenAI. Nama medan tepat bagi setiap satu, serta tetapan yang selalu rosak. - [Nota suara ke teks untuk perniagaan: WhatsApp dan Telegram ke tiket](https://broadrouter.com/ms/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): Tukar nota suara pelanggan daripada WhatsApp Business dan bot Telegram kepada transkrip dan tiket sokongan: muat turun, transkripsi, ekstrak medan dengan LLM. - [Pelan LLM mingguan untuk projek pendek: Orbit vs sejam vs sebulan](https://broadrouter.com/ms/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): Bilakah pelan DeepSeek V4.1 Flash tanpa had seminggu lebih baik daripada sesi sejam, pelan bulanan atau bayar setiap token, dengan kiraan untuk hackathon dan sprint. - [Apakah maksud sebenar "tanpa had" dalam pelan API LLM](https://broadrouter.com/ms/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Meter token, tetingkap bergulir 5 jam, had mingguan, had kadar dan keserentakan. Tujuh soalan untuk mengetahui sama ada pelan AI "tanpa had" benar-benar tanpa had. - [Alternatif API Whisper: bila perlu bertukar, dan cara mengujinya](https://broadrouter.com/ms/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): Bilakah alternatif berhos kepada API Whisper OpenAI masuk akal: kadar ralat perkataan pada audio anda, bahasa, harga sejam, had dan privasi, dengan kod ujian. - [API LLM dengan pengekalan data sifar: senarai semak 10 soalan](https://broadrouter.com/ms/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): Maksud sebenar "pengekalan data sifar" bagi API LLM: prompt, output, log, latihan, subpemproses dan audit. Sepuluh soalan untuk ditanya sebelum menghantar data. - [Guides in Malay](https://broadrouter.com/ms/guides/) - [Cara guna OpenAI Python SDK dengan BroadNet Inference](https://broadrouter.com/ms/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): Halakan OpenAI Python SDK rasmi ke BroadNet Inference dengan satu perubahan base_url. Pasang, tetapkan kunci, buat permintaan pertama dan strim jawapan. - [Cara guna OpenAI Node.js SDK dengan BroadNet Inference](https://broadrouter.com/ms/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): Halakan OpenAI Node.js dan TypeScript SDK rasmi ke BroadNet Inference dengan menetapkan baseURL. Pasang, buat permintaan pertama, strim token dan log ID permintaan. - [Cuba BroadNet Inference dalam pelayar, tanpa akaun](https://broadrouter.com/ms/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): Uji pertuturan ke teks bahasa Arab, pembacaan dokumen dan sembang Gemma 4 dalam playground pelayar BroadNet tanpa daftar atau kad, lalu pindahkan permintaan ke API. - [Permintaan pertama ke BroadNet Inference dengan curl](https://broadrouter.com/ms/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): Buat panggilan pertama ke BroadNet Inference dengan curl: chat completion, jawapan distrim dan transkripsi audio, kemudian baca respons, usage dan ralat. - [Strim chat completions dan kendalikan sambungan yang terputus](https://broadrouter.com/ms/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): Cara penstriman berfungsi di BroadNet Inference: Server-Sent Events, penanda [DONE], penggunaan token dalam chunk terakhir, dan tindakan jika strim berakhir awal. - [Mengapa model penaakulan memulangkan jawapan kosong](https://broadrouter.com/ms/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3 dan Qwen3.8 27B berfikir dahulu sebelum menjawab, dan penaakulan dikira dalam max_tokens. Cara mengelakkan balasan kosong. - [Kod ralat BroadNet Inference dan bila perlu cuba semula](https://broadrouter.com/ms/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): Maksud setiap status HTTP BroadNet Inference API, yang mana perlu dicuba semula dengan backoff, yang mana dibaiki dahulu, dan apa yang dihantar kepada sokongan. - [Kunci API kepada pelayan anda dengan senarai IP dibenarkan](https://broadrouter.com/ms/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): Lindungi kunci BroadNet Inference dengan senarai IP dibenarkan, asingkan kunci inferens daripada kelayakan portal dan batalkan kunci bocor dalam beberapa saat. - [Cara pelan API LLM tanpa had berfungsi untuk ejen](https://broadrouter.com/ms/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): Pelan DeepSeek V4.1 Flash tanpa had tiada meter token dan tiada had mingguan. Begini sebenarnya cara sesi, ejen selari, kunci dan pengebilan berfungsi. - [Jalankan Aider dengan DeepSeek V4.1 Flash tanpa had](https://broadrouter.com/ms/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Sambungkan ejen pengekodan Aider kepada BroadNet Inference melalui tetapan serasi OpenAI dan jalankannya pada pelan DeepSeek V4.1 Flash tanpa had berharga tetap. - [Cara guna LangChain dengan BroadNet Inference](https://broadrouter.com/ms/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): Sambungkan ChatOpenAI LangChain ke BroadNet Inference dengan base_url dan api_key, dan ketahui apa yang disimpan dan dibuang LangChain daripada model penaakulan. - [Guna Cline dengan BroadNet Inference (penyedia OpenAI Compatible)](https://broadrouter.com/ms/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Sediakan sambungan VS Code Cline dengan BroadNet Inference melalui penyedia OpenAI Compatible: medan tepat, ID model, saiz konteks dan pembaikan biasa. - [Guna Roo Code dengan BroadNet Inference (OpenAI Compatible)](https://broadrouter.com/ms/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code ditutup pada Mei 2026. Tetapan untuk pemasangan sedia ada dan fork, serta ke mana hendak berpindah (Cline atau fork Zoo Code) dengan BroadNet Inference. - [Guna Continue dengan BroadNet Inference (config.yaml dan apiBase)](https://broadrouter.com/ms/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): Tambah model BroadNet Inference ke sambungan IDE Continue dengan provider openai dan apiBase dalam config.yaml, termasuk keupayaan tool_use dan panjang konteks. - [Tambah BroadNet Inference ke opencode sebagai penyedia tersuai](https://broadrouter.com/ms/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): Tambah BroadNet Inference ke ejen pengekodan opencode sebagai penyedia serasi OpenAI tersuai dalam opencode.jsonc, dengan alias model, capabilities dan had. - [Cara sambungkan Open WebUI kepada BroadNet Inference](https://broadrouter.com/ms/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Tambah BroadNet Inference dalam Open WebUI sebagai sambungan serasi OpenAI, senaraikan ID model, dan baiki isu verify, balasan kosong dan had kadar. - [Tambah BroadNet Inference ke LibreChat sebagai custom endpoint](https://broadrouter.com/ms/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): Tambah BroadNet Inference ke LibreChat sebagai custom endpoint serasi OpenAI dalam librechat.yaml, dengan kunci dalam .env, senarai model, model tajuk dan pembaikan. - [Guna BroadNet Inference dalam n8n (kelayakan OpenAI, Base URL)](https://broadrouter.com/ms/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): Halakan kelayakan OpenAI n8n ke BroadNet Inference melalui medan Base URL, kemudian guna nod OpenAI Chat Model mengikut ID dengan Responses API dimatikan. - [Cara guna Vercel AI SDK dengan BroadNet Inference](https://broadrouter.com/ms/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): Panggil BroadNet Inference dalam Vercel AI SDK dengan createOpenAICompatible dari @ai-sdk/openai-compatible: generateText, streamText, alatan, reasoning effort. - [API speech-to-text bahasa Arab, langkah demi langkah](https://broadrouter.com/ms/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): Transkripsikan audio bahasa Arab dengan BroadNet Transcribe Arabic melalui endpoint serasi OpenAI. Format, medan language, contoh Python dan curl, serta harganya. - [Halakan LiteLLM ke BroadNet Inference (SDK dan proksi)](https://broadrouter.com/ms/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): Panggil BroadNet Inference melalui LiteLLM dengan awalan model openai/ dan api_base, dalam Python SDK atau config.yaml proksi, dengan had kadar yang sepadan. - [Jalankan OpenAI Agents SDK pada BroadNet Inference](https://broadrouter.com/ms/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Guna OpenAI Agents SDK (Python) dengan BroadNet Inference melalui AsyncOpenAI dan OpenAIChatCompletionsModel, dan matikan tracing yang memerlukan kunci OpenAI. - [Cara tukar nota suara WhatsApp kepada teks dengan API](https://broadrouter.com/ms/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): Tukar nota suara WhatsApp .ogg atau rakaman telefon .m4a ke WAV dengan ffmpeg dan transkripsikannya dalam bahasa Arab atau 13 bahasa lain dengan satu panggilan API. - [Cara transkripsikan nota suara WhatsApp dalam n8n](https://broadrouter.com/ms/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): Aliran kerja n8n yang menerima nota suara WhatsApp, memuat turunnya daripada Meta, menukarnya dengan ffmpeg dan mentranskripsikannya dalam bahasa Arab atau 13 lagi. - [Baca invois dan resit Arab dengan model vision](https://broadrouter.com/ms/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Hantar foto invois atau resit kepada model vision di BroadNet Inference dan dapatkan jumlah, tarikh dan vendor sebagai JSON. Soalan boleh dalam Arab atau Inggeris. - [Gunakan Zed dengan BroadNet Inference sebagai penyedia serasi OpenAI](https://broadrouter.com/ms/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Tambah BroadNet Inference pada editor Zed sebagai penyedia serasi OpenAI dalam settings.json, dengan ID model, had token, keupayaan dan pembolehubah kunci. - [Guna Kilo Code dengan BroadNet Inference (penyedia OpenAI Compatible)](https://broadrouter.com/ms/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Sambungkan Kilo Code ke BroadNet Inference sebagai penyedia tersuai OpenAI Compatible dalam dialog VS Code atau kilo.json, dengan ID model, had dan bendera alatan. - [Tambah BroadNet Inference ke goose sebagai penyedia tersuai](https://broadrouter.com/ms/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): Sambungkan ejen AI goose kepada BroadNet Inference sebagai penyedia tersuai OpenAI Compatible, melalui goose Desktop, goose configure atau fail JSON. - [Guna Pydantic AI dengan BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/ms/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Jalankan ejen Pydantic AI pada BroadNet Inference dengan OpenAIChatModel dan OpenAICompatibleProvider, serta tool calls, output berstruktur dan model profile. - [Guna CrewAI dengan BroadNet Inference sebagai endpoint OpenAI tersuai](https://broadrouter.com/ms/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): Halakan ejen CrewAI ke BroadNet Inference dengan LLM(custom_openai=True, base_url=...), kemudian pilih model untuk tools, konteks panjang dan pelan kadar tetap. - [Guna LlamaIndex dengan BroadNet Inference melalui OpenAILike](https://broadrouter.com/ms/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): Sambungkan LlamaIndex ke BroadNet Inference dengan kelas OpenAILike, dan tetapkan api_base, context_window, is_chat_model dan is_function_calling_model dengan betul. - [Tambah BroadNet Inference ke Dify dengan pemalam OpenAI-API-compatible](https://broadrouter.com/ms/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): Sambungkan Dify kepada BroadNet Inference melalui pemalam penyedia model OpenAI-API-compatible, dengan nilai medan tepat, saiz konteks dan tetapan function calling. - [Guna Jan dengan BroadNet Inference sebagai penyedia serasi OpenAI](https://broadrouter.com/ms/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Tambah BroadNet Inference ke aplikasi desktop Jan sebagai penyedia serasi OpenAI, masukkan ID model secara manual dan hidupkan alatan serta vision setiap model. - [Sambung SillyTavern ke BroadNet Inference (Custom OpenAI-compatible)](https://broadrouter.com/ms/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): Sambung SillyTavern ke BroadNet Inference dengan sumber Chat Completion Custom (OpenAI-compatible), ID model yang betul, prompt post-processing dan had token. - [Guna Spring AI dengan BroadNet Inference (OpenAI base-url)](https://broadrouter.com/ms/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Halakan starter chat OpenAI Spring AI ke BroadNet Inference dengan spring.ai.openai.base-url, ID model BroadNet dan max-tokens, kemudian panggil dari ChatClient. - [Panggil BroadNet Inference daripada Go dengan SDK openai-go](https://broadrouter.com/ms/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): Guna SDK rasmi openai-go v3 dengan BroadNet Inference melalui option.WithBaseURL, termasuk chat completions, penstriman dengan usage dan pengendalian ralat. - [Bina ejen LangGraph pada BroadNet Inference](https://broadrouter.com/ms/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): Jalankan StateGraph atau create_agent LangGraph pada BroadNet Inference dengan ChatOpenAI, base_url dan ID model BroadNet, serta alatan, penstriman dan memori. - [Guna pipeline dan ejen Haystack dengan BroadNet Inference](https://broadrouter.com/ms/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Halakan OpenAIChatGenerator Haystack ke BroadNet Inference dengan api_base_url, ID model BroadNet dan kunci Secret, untuk pipeline, alatan, strim dan ejen. - [Guna Semantic Kernel dengan BroadNet Inference](https://broadrouter.com/ms/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Sambung Microsoft Semantic Kernel ke BroadNet Inference dalam Python dengan klien AsyncOpenAI, atau dalam C# dengan endpoint tersuai, untuk chat, pemalam dan ejen. - [Guna ejen Mastra dengan BroadNet Inference](https://broadrouter.com/ms/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): Jalankan ejen Mastra pada BroadNet Inference dengan objek model yang menetapkan id, url dan apiKey, termasuk awalan broadnet/ berganda yang diperlukan ID model. - [Dapatkan output berstruktur dengan Instructor pada BroadNet Inference](https://broadrouter.com/ms/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Ekstrak objek Pydantic daripada BroadNet Inference dengan Instructor, melalui from_openai atau from_provider dengan base_url tersuai, dan pilih Mode yang betul. - [Guna DSPy dengan BroadNet Inference (tetapkan model_type="chat")](https://broadrouter.com/ms/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): Konfigurasikan dspy.LM untuk BroadNet Inference dengan awalan openai/, api_base dan model_type="chat", kerana DSPy 3.5 menghantar model OpenAI ke Responses API. - [Nilai model BroadNet dengan promptfoo](https://broadrouter.com/ms/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): Jalankan eval promptfoo dan ujian model sebelah-menyebelah pada BroadNet Inference dengan penyedia openai:chat, apiBaseUrl dan apiKeyEnvar, serta grader. - [Jalankan smolagents di BroadNet Inference dengan OpenAIModel](https://broadrouter.com/ms/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Sambung smolagents Hugging Face ke BroadNet Inference dengan OpenAIModel dan api_base: CodeAgent pada sebarang model, ToolCallingAgent pada GLM 5.3 dan Qwen3.8 27B. - [Sambungkan Flowise kepada BroadNet Inference (OpenAI Custom Model)](https://broadrouter.com/ms/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Guna BroadNet Inference dalam chatflow Flowise melalui nod OpenAI Custom Model, dengan Base Path, kelayakan OpenAI API dan nama model BroadNet yang ditaip. - [Sambungkan Langflow kepada BroadNet Inference](https://broadrouter.com/ms/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Guna BroadNet Inference dalam Langflow dengan medan OpenAI API Base komponen OpenAI dan nama model yang ditaip, serta sebab OpenAI Compatible tidak menyenaraikannya. - [Glossary entries in Malay](https://broadrouter.com/ms/glossary/) - [Ejen AI (AI agent)](https://broadrouter.com/ms/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): Ejen AI ialah sistem di mana model bahasa mengarahkan langkahnya sendiri ke arah sesuatu matlamat: ia memutuskan alatan yang hendak dipanggil, membaca hasilnya dan mengulang sehingga menghasilkan jawapan akhir atau mencapai had. Skrip tetap yang memanggil model pada titik yang ditetapkan biasanya dipanggil aliran kerja (workflow). - [Kunci API (API key)](https://broadrouter.com/ms/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): Kunci API ialah rentetan rahsia yang mengenal pasti dan membenarkan klien yang memanggil sesuatu API. API LLM biasanya menjangkakannya dalam pengepala Authorization: Bearer, dan sesiapa yang memegang kunci itu boleh membuat permintaan yang dibilkan kepada akaun pemiliknya. - [Baris Arab (tashkeel, Arabic diacritics)](https://broadrouter.com/ms/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): Baris Arab, atau tashkeel, ialah tanda kecil yang ditulis di atas atau di bawah huruf untuk menunjukkan vokal pendek, konsonan berganda dan ketiadaan vokal. Ia pilihan: kebanyakan teks Arab, termasuk berita, meninggalkannya, jadi pembaca dan perisian meneka vokal daripada konteks. - [Pengecaman dialek Arab (Arabic dialect identification)](https://broadrouter.com/ms/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): Pengecaman dialek Arab ialah tugas menentukan varian bahasa Arab yang digunakan dalam sesuatu teks atau rakaman, seperti Mesir, Teluk, Levant, Maghribi atau Bahasa Arab Standard Moden, pada peringkat rantau, negara atau bandar. Ia digunakan untuk menghalakan mesej, memilih prompt dan membahagikan set penilaian. - [Penormalan teks Arab (Arabic text normalization)](https://broadrouter.com/ms/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): Penormalan teks Arab ialah langkah yang memetakan varian ejaan bagi perkataan yang sama kepada satu bentuk sebelum padanan: membuang baris dan tatweel, menyatukan bentuk alif, dan selalunya memetakan ى kepada ي dan ة kepada ه. Ia digunakan pada kunci carian dan penilaian, bukan pada teks yang anda paparkan. - [Arabizi](https://broadrouter.com/ms/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): Arabizi ialah bahasa Arab yang ditulis secara tidak formal dalam huruf Rumi, menggunakan digit untuk bunyi Arab yang tiada padanan Rumi, seperti 3 untuk ع. Ia biasa dalam chat dan media sosial, kebanyakannya dialek, tiada ejaan standard, dan sering bercampur dengan bahasa Inggeris atau Perancis. - [AWQ (activation-aware weight quantization)](https://broadrouter.com/ms/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (activation-aware weight quantization) ialah kaedah pasca latihan yang memampatkan pemberat LLM kepada ketepatan rendah, biasanya integer 4-bit. Ia menggunakan statistik pengaktifan daripada set kalibrasi kecil untuk mencari saluran pemberat paling penting dan menskalakannya supaya kurang ketepatan hilang. Ia tidak memerlukan backpropagation. - [Base URL (URL asas)](https://broadrouter.com/ms/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): Base URL ialah alamat akar tetap sesuatu API, seperti https://api.example.com/v1, yang diletakkan oleh klien di hadapan setiap laluan endpoint, seperti /chat/completions. Dalam SDK serasi OpenAI, menukar base URL, bersama kunci API dan ID model, menghalakan kod yang sama kepada penyedia lain. - [API kelompok (batch API)](https://broadrouter.com/ms/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Batch API ialah endpoint LLM tak segerak yang menerima banyak permintaan sekali gus, biasanya sebagai fail, memprosesnya dalam tempoh yang ditetapkan dan bukan serta-merta, dan memulangkan hasilnya bersama-sama. Penyedia seperti OpenAI dan Anthropic mengenakan harga diskaun untuk permintaan kelompok berbanding API standard mereka. - [BM25](https://broadrouter.com/ms/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) ialah fungsi kedudukan untuk carian kata kunci yang memberi skor kepada dokumen berdasarkan kekerapan istilah pertanyaan muncul di dalamnya, sejauh mana istilah itu jarang dalam keseluruhan koleksi, dan panjang dokumen. Ia ialah similarity lalai dalam Lucene dan Elasticsearch. - [Byte pair encoding (BPE)](https://broadrouter.com/ms/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Byte pair encoding (BPE) ialah algoritma tokenisasi subkata yang bermula daripada aksara atau bait tunggal dan berulang kali menggabungkan pasangan bersebelahan yang paling kerap menjadi entri perbendaharaan kata baharu, sehingga mencapai saiz perbendaharaan kata sasaran. Kebanyakan tokenizer LLM semasa menggunakan BPE atau varian peringkat baitnya. - [Token input cache (cached input tokens)](https://broadrouter.com/ms/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): Token input cache ialah token prompt yang baru sahaja diproses oleh penyedia LLM dan boleh diguna semula daripada cache tanpa dikira semula, biasanya awalan yang berulang seperti prompt sistem atau dokumen. Penyedia mengebilnya pada kadar lebih rendah daripada token input baharu. - [Chain-of-thought (CoT, rantaian pemikiran)](https://broadrouter.com/ms/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Chain-of-thought (CoT) ialah teknik prompting di mana model bahasa ditunjukkan atau diminta menulis langkah penaakulan perantaraan sebelum jawapan akhirnya. Ia meningkatkan ketepatan bagi masalah berbilang langkah seperti aritmetik dan logik, dan model penaakulan kini menjananya secara automatik. - [Kadar ralat aksara (character error rate, CER)](https://broadrouter.com/ms/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Kadar ralat aksara (CER) mengukur ketepatan transkripsi atau OCR pada peringkat aksara: penggantian, pemadaman dan penyisipan yang diperlukan untuk menukar output kepada teks rujukan, dibahagikan dengan bilangan aksara dalam rujukan. Lebih rendah lebih baik, dan 0 ialah padanan sempurna. - [Chat Completions API](https://broadrouter.com/ms/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Chat Completions API ialah endpoint HTTP OpenAI, POST /v1/chat/completions, yang menerima nama model dan senarai mesej bertanda peranan, lalu memulangkan mesej seterusnya daripada model. Bentuk permintaan dan responsnya menjadi format biasa yang digunakan semula oleh penyedia dan SDK yang serasi OpenAI. - [Templat chat (chat template)](https://broadrouter.com/ms/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): Templat chat ialah peraturan pemformatan, biasanya ditulis dalam Jinja dan disertakan bersama tokenizer model, yang menukar senarai mesej chat berserta peranannya kepada satu jujukan token yang dijangka oleh model, termasuk token kawalan khasnya. Templat yang salah boleh merosotkan jawapan model dengan teruk. - [Chunking (pemecahan dokumen untuk RAG)](https://broadrouter.com/ms/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Chunking ialah langkah dalam pipeline retrieval yang memecahkan dokumen kepada cebisan lebih kecil, dipanggil chunk, sebelum ia di-embed dan diindeks. Saiz chunk menentukan berapa banyak teks dalam setiap cebisan, dan overlap chunk mengulang sebahagian teks antara chunk bersebelahan supaya idea yang terpotong di sempadan tidak hilang. - [Penukaran kod (code-switching)](https://broadrouter.com/ms/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Penukaran kod ialah penggunaan dua atau lebih bahasa, atau varian satu bahasa, dalam perbualan atau ayat yang sama, sama ada dalam pertuturan atau penulisan. Contohnya penutur Arab yang mencampurkan perkataan Inggeris ke dalam dialek, atau bertukar antara Bahasa Arab Standard Moden dan dialek mereka. - [Ejen pengekodan AI (AI coding agent)](https://broadrouter.com/ms/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): Ejen pengekodan AI ialah alat perisian di mana model bahasa bekerja pada codebase dalam satu gelung: ia membaca fail, menyuntingnya, menjalankan arahan seperti ujian, membaca output dan menentukan langkah seterusnya sehingga tugas selesai atau had dicapai. - [Keserentakan (concurrency)](https://broadrouter.com/ms/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): Keserentakan ialah bilangan permintaan yang sedang berjalan serentak daripada satu klien. API LLM sering mengehadkannya bagi setiap kunci atau pelan, berasingan daripada permintaan seminit, kerana setiap permintaan yang terbuka memegang kapasiti GPU sehingga responsnya selesai. - [Pemampatan konteks (context compaction)](https://broadrouter.com/ms/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): Pemampatan konteks ialah amalan meringkaskan perbualan yang hampir mencapai had tetingkap konteks dan meneruskan sesi daripada ringkasan itu, supaya ejen mengekalkan keputusan dan keadaan penting sambil membuang sejarah mentah seperti output alatan lama. - [Kejuruteraan konteks (context engineering)](https://broadrouter.com/ms/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): Kejuruteraan konteks ialah kerja memilih dan menyelenggara set token yang dilihat oleh model bahasa pada setiap permintaan, termasuk prompt sistem, definisi alat, sejarah, dokumen yang diambil dan nota, supaya model mendapat set maklumat bernilai tinggi yang paling kecil yang diperlukan untuk melaksanakan tugas dengan baik. - [Context rot (kemerosotan konteks)](https://broadrouter.com/ms/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): Context rot ialah kehilangan ketepatan dan kebolehpercayaan secara beransur-ansur yang ditunjukkan oleh model bahasa apabila input menjadi lebih panjang, walaupun tugasnya kekal sama dan input masih jauh dalam had tetingkap konteks. Istilah ini dipopularkan oleh laporan teknikal Chroma pada 2025 yang menguji 18 model. - [Tetingkap konteks (context window)](https://broadrouter.com/ms/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): Tetingkap konteks ialah bilangan maksimum token yang boleh dikendalikan oleh model bahasa dalam satu permintaan. Ia meliputi prompt, sejarah perbualan, definisi alatan, imej dan output yang dijana, termasuk token penaakulan. Apa-apa yang tidak muat mesti dipotong, diringkaskan atau dipecahkan kepada beberapa permintaan. - [Residensi data (data residency)](https://broadrouter.com/ms/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): Residensi data ialah keperluan atau komitmen bahawa data disimpan dan diproses di negara atau rantau tertentu. Bagi API LLM, ia meliputi lokasi prompt diproses semasa inferens dan lokasi sebarang kandungan yang disimpan, dan selalunya tidak termasuk data akaun dan pengebilan. - [Dialek Arab (dialectal Arabic)](https://broadrouter.com/ms/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): Dialek Arab ialah bahasa Arab pertuturan harian di sesuatu rantau atau bandar, seperti bahasa Arab Mesir, Teluk, Levant atau Maghribi. Dialek berbeza daripada Bahasa Arab Standard Moden dari segi bunyi, perkataan dan tatabahasa, tiada ejaan standard, dan semakin banyak ditulis dalam talian, menjadikannya lebih sukar bagi model bahasa. - [Diarisasi penutur (speaker diarization)](https://broadrouter.com/ms/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): Diarisasi penutur ialah tugas membahagikan rakaman audio kepada segmen mengikut penutur, menjawab soalan "siapa bercakap dan bila". Ia melabelkan setiap bahagian pertuturan dengan ID penutur tanpa nama seperti SPEAKER_00, dan biasanya digabungkan dengan transkrip untuk menunjukkan siapa yang menyebut setiap baris. - [Penyulingan model (knowledge distillation)](https://broadrouter.com/ms/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): Penyulingan, atau penyulingan pengetahuan (knowledge distillation), ialah kaedah latihan di mana model pelajar yang lebih kecil belajar menghasilkan semula output model guru yang lebih besar, seperti taburan kebarangkalian atau jawapan yang dijananya, supaya pelajar mengekalkan sebahagian besar kualiti guru dengan kos menjalankan yang lebih rendah. - [Perjanjian pemprosesan data (DPA)](https://broadrouter.com/ms/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): Perjanjian pemprosesan data (DPA) ialah kontrak antara pengawal dan pemproses yang mengawal cara pemproses mengendalikan data peribadi. GDPR Perkara 28(3) mewajibkannya dan menyenaraikan kandungannya: arahan yang didokumenkan, kerahsiaan, keselamatan, peraturan subpemproses, pemadaman atau pemulangan data, dan audit. - [Embedding (vektor teks)](https://broadrouter.com/ms/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): Embedding ialah senarai nombor titik terapung (vektor) yang dihasilkan oleh model untuk mewakili sekeping teks, imej atau input lain. Input yang maknanya serupa mendapat vektor yang berdekatan, jadi jarak antara vektor boleh digunakan untuk menyusun hasil carian, mengelompokkan dokumen atau mencari item berkaitan. - [Eval (penilaian LLM)](https://broadrouter.com/ms/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Eval (singkatan bagi evaluation, penilaian) ialah ujian berulang bagi model bahasa atau prompt: satu set input tetap, penerangan tentang rupa output yang baik, dan penggred (kod, manusia atau model lain) yang memberi skor kepada setiap output supaya versi boleh dibandingkan. - [Backoff eksponen (exponential backoff)](https://broadrouter.com/ms/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): Backoff eksponen ialah strategi cuba semula di mana klien menunggu lebih lama selepas setiap percubaan yang gagal, biasanya menggandakan kelewatan sehingga had tertentu, sebelum mencuba lagi. Menambah jitter rawak pada setiap penantian menyebarkan cubaan semula daripada banyak klien supaya mereka tidak mengenai pelayan pada saat yang sama. - [Few-shot prompting (prompt dengan beberapa contoh)](https://broadrouter.com/ms/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): Few-shot prompting ialah teknik di mana prompt mengandungi sebilangan kecil contoh pasangan input-output sebelum input sebenar, supaya model bahasa dapat menyimpulkan tugas dan format daripadanya. Pemberat model tidak berubah; zero-shot tidak menggunakan contoh dan one-shot menggunakan satu. - [Fine-tuning (penalaan halus)](https://broadrouter.com/ms/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): Fine-tuning ialah proses meneruskan latihan model bahasa yang telah dilatih dengan set data yang lebih kecil dan khusus tugas, seperti contoh perbualan, supaya pemberatnya berubah dan ia melaksanakan tugas atau gaya itu dengan lebih konsisten tanpa memerlukan arahan atau contoh panjang dalam setiap prompt. - [finish_reason](https://broadrouter.com/ms/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason ialah medan pada setiap choice dalam respons chat completion yang menyatakan sebab model berhenti menjana. Dalam format OpenAI, nilainya ialah stop, length, tool_calls, content_filter dan function_call yang sudah lapuk. Messages API Anthropic melaporkan idea yang sama sebagai stop_reason. - [Pelan LLM harga tetap (flat-rate LLM plan)](https://broadrouter.com/ms/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): Pelan LLM harga tetap mengenakan harga tetap bagi satu tempoh, seperti sebulan, untuk akses kepada API model bahasa dan bukannya mengebil setiap token. Pelan sebegini biasanya mengehadkan perkara lain, seperti keserentakan, kelajuan, model atau tetingkap penggunaan, jadi hadnya sama penting dengan harganya. - [FP8](https://broadrouter.com/ms/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 ialah format nombor titik apung 8-bit yang digunakan dalam pembelajaran mendalam, ditakrifkan dalam dua pengekodan: E4M3, dengan 4 bit eksponen dan 3 bit mantisa, dan E5M2, dengan 5 bit eksponen dan 2 bit mantisa. Ia menjimatkan separuh memori berbanding format 16-bit dan berjalan secara asli pada GPU terkini. - [Penalti kekerapan dan kehadiran](https://broadrouter.com/ms/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): Penalti kekerapan dan kehadiran (frequency dan presence penalty) ialah tetapan pensampelan yang merendahkan skor token yang sudah ada dalam output. Penalti kehadiran dikenakan sekali sebaik token muncul; penalti kekerapan meningkat dengan setiap ulangan. Dalam API OpenAI, julat kedua-duanya -2.0 hingga 2.0, dengan nilai lalai 0. - [GDPR](https://broadrouter.com/ms/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): GDPR (General Data Protection Regulation, Regulation (EU) 2016/679) ialah undang-undang EU tentang pemprosesan data peribadi. Ia terpakai sejak 25 Mei 2018 dan meliputi organisasi yang bertapak di EU serta sesetengah organisasi di luar EU. Bagi API LLM, ia mengawal kontrak pemproses, pemindahan antarabangsa dan peminimuman data. - [GGUF](https://broadrouter.com/ms/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF ialah format fail binari untuk menyimpan model bahasa bagi inferens dengan runtime berasaskan GGML seperti llama.cpp. Satu fail GGUF memegang pemberat, selalunya dikuantisasi, serta metadata yang diperlukan untuk memuatkan model, termasuk butiran tokenizer dan seni bina, dan ia boleh dipetakan ke memori (memory-mapped) untuk pemuatan pantas. - [Jam GPU (GPU-hour)](https://broadrouter.com/ms/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): Jam GPU ialah satu pemproses grafik yang digunakan selama satu jam. Penyedia awan mengecaj instans GPU yang disewa mengikut jam atau saat, dan penyelidik menggunakan jam GPU untuk menyatakan jumlah pengiraan yang digunakan sesuatu tugas. Anda membayar masa mesin itu, sama ada ia sibuk atau tidak. - [Grounding (penambatan pada sumber)](https://broadrouter.com/ms/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): Grounding ialah amalan menghubungkan output model bahasa dengan sumber maklumat yang khusus dan boleh disahkan, seperti dokumen yang dibekalkan, hasil carian atau rekod pangkalan data, dan mengarahkannya menjawab berdasarkan sumber tersebut. Jawapan yang di-ground boleh dijejak kembali ke sumber dan kurang berkemungkinan mengandungi fakta rekaan. - [Guardrails (pagar keselamatan LLM)](https://broadrouter.com/ms/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): Guardrails ialah semakan yang dijalankan aplikasi di sekeliling panggilan model bahasa untuk mengekalkannya dalam had yang ditetapkan. Guardrails input menapis apa yang masuk ke model, guardrails output menyemak apa yang keluar, dan guardrails alat mengesahkan tindakan sebelum dijalankan. Semakan yang gagal akan menyekat, menulis semula atau membenderakan permintaan. - [Halusinasi AI (AI hallucination)](https://broadrouter.com/ms/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): Halusinasi ialah output model bahasa yang kedengaran munasabah tetapi salah, direka-reka atau tidak disokong oleh input yang diberikan, seperti petikan rujukan rekaan, angka yang salah atau API yang tidak wujud. Ia berpunca daripada model meramalkan teks yang mungkin, bukan menyemak fakta. - [Hamzah (hamza)](https://broadrouter.com/ms/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): Hamzah (ء) ialah huruf bahasa Arab bagi hentian glotis. Ia ditulis sendirian atau di atas kerusi: alif (أ إ), wau (ؤ) atau ya (ئ), manakala آ menggabungkan alif dengan madd. Oleh sebab teks yang ditaip tidak konsisten, enjin carian dan pemarkahan pertuturan sering menyatukan bentuk alif kepada ا. - [Carian hibrid (hybrid search)](https://broadrouter.com/ms/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): Carian hibrid ialah kaedah dapatan semula yang menjalankan carian kata kunci, seperti BM25, dan carian vektor ke atas embedding bagi pertanyaan yang sama, kemudian menggabungkan kedua-dua senarai berkedudukan menjadi satu, selalunya dengan reciprocal rank fusion (RRF) atau jumlah berwajaran skor yang dinormalkan. - [Kunci idempotensi (idempotency key)](https://broadrouter.com/ms/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Kunci idempotensi ialah nilai unik, biasanya ID rawak seperti UUID, yang dilampirkan oleh klien pada permintaan yang mengubah keadaan, seperti pembayaran atau pembelian. Jika permintaan yang sama dihantar semula dengan kunci yang sama, pelayan memulangkan hasil asal dan tidak melakukan tindakan itu dua kali. - [Token imej (image tokens)](https://broadrouter.com/ms/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): Token imej ialah token input yang digunakan oleh model bahasa vision untuk mewakili imej dalam prompt. Penyedia mengiranya daripada saiz imej, biasanya dengan memecahkannya kepada patch atau jubin bersaiz tetap selepas saiz diubah, dan mengebilkannya seperti token input teks, jadi imej yang lebih besar lebih mahal. - [Penyedia inferens (inference provider)](https://broadrouter.com/ms/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): Penyedia inferens ialah syarikat yang menjalankan model pembelajaran mesin, selalunya LLM open-weight, pada GPU yang dimiliki, dipajak atau disewanya, dan menjual akses kepadanya melalui API, biasanya berharga setiap token, setiap permintaan atau setiap unit masa. Ia mengendalikan perkhidmatan, penskalaan dan perkakasan. - [Token input dan output (input and output tokens)](https://broadrouter.com/ms/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): Token input ialah token yang dibaca oleh model bahasa dalam satu permintaan, termasuk prompt sistem, sejarah perbualan, definisi alatan dan imej. Token output ialah token yang dijananya sebagai balasan, termasuk sebarang penaakulan. API LLM biasanya meletakkan harga berasingan, dan output biasanya lebih mahal. - [INT4](https://broadrouter.com/ms/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 ialah format nombor integer 4-bit yang boleh mewakili 16 tahap. Dalam inferens LLM, ia biasanya bermaksud kuantisasi pemberat sahaja 4-bit (W4A16): pemberat disimpan sebagai INT4 bersama faktor skala dan dikembangkan kepada 16-bit semasa pengiraan, menggunakan kira-kira suku daripada memori pemberat 16-bit. - [Senarai IP dibenarkan (IP allow list)](https://broadrouter.com/ms/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): Senarai IP dibenarkan ialah kawalan keselamatan yang hanya menerima permintaan daripada alamat IP atau julat CIDR yang disenaraikan dan menolak yang lain. Pada API, ia mengehadkan kerosakan akibat kunci yang bocor, kerana kunci itu tidak berfungsi dari rangkaian yang tiada dalam senarai. - [Jailbreak (LLM)](https://broadrouter.com/ms/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): Jailbreak ialah input yang direka untuk membuat model bahasa mengabaikan latihan keselamatannya atau peraturan yang ditetapkan oleh aplikasi yang menggunakannya, supaya ia menghasilkan kandungan atau mengambil tindakan yang biasanya ditolaknya. OWASP menganggap jailbreaking sebagai satu bentuk prompt injection. - [Mod JSON (JSON mode)](https://broadrouter.com/ms/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): Mod JSON ialah tetapan API LLM, response_format {"type": "json_object"} dalam API gaya OpenAI, yang menjadikan jawapan model JSON yang sah dan boleh dihurai. Tidak seperti structured outputs dengan JSON Schema, ia tidak menjamin kunci, jenis atau nilai yang terkandung dalam objek itu. - [JSON Schema (skema JSON)](https://broadrouter.com/ms/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema ialah perbendaharaan kata, ditulis dalam JSON, untuk menerangkan dan mengesahkan struktur data JSON: jenis, medan wajib, nilai yang dibenarkan dan penyarangan. Versi semasa ialah 2020-12. API LLM menggunakannya untuk mentakrifkan parameter alat dan bentuk tepat yang mesti diikuti oleh output berstruktur. - [KV cache](https://broadrouter.com/ms/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): KV cache ialah memori tempat model bahasa transformer menyimpan key dan value attention yang dikira bagi setiap token yang telah diproses, supaya setiap token baharu menggunakannya semula dan tidak mengira semula keseluruhan jujukan. Ia membesar mengikut panjang konteks dan selalunya mengehadkan bilangan permintaan yang boleh dilayan oleh GPU. - [Kependaman (latency)](https://broadrouter.com/ms/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): Kependaman, bagi API model bahasa, ialah masa yang diambil oleh sesuatu permintaan dari saat dihantar sehingga token output terakhir diterima. Ia ialah jumlah masa ke token pertama dan masa penjanaan, dan biasanya dilaporkan sebagai persentil seperti p50 dan p95, bukan purata. - [LLM-as-a-judge (LLM sebagai hakim)](https://broadrouter.com/ms/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge ialah kaedah penilaian di mana model bahasa menilai output model lain, sama ada memberi skor kepada satu jawapan berdasarkan rubrik atau memilih yang lebih baik antara dua jawapan. Ia menghampiri pertimbangan manusia pada kos jauh lebih rendah tetapi mempunyai bias yang diketahui, seperti memihak kepada jawapan pertama atau lebih panjang. - [Penanda aras LLM (LLM benchmark)](https://broadrouter.com/ms/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): Penanda aras LLM ialah set tugasan tetap, biasanya awam, dengan kaedah pemarkahan, yang digunakan untuk mengukur dan membandingkan model bahasa pada satu kemahiran, seperti soalan pengetahuan, pengekodan atau pengecaman pertuturan. Skor penanda aras menunjukkan keupayaan relatif pada ujian itu, bukan prestasi model pada beban kerja anda sendiri. - [Penghala LLM (LLM router)](https://broadrouter.com/ms/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): Penghala LLM ialah perisian atau perkhidmatan yang terletak di antara aplikasi dan satu atau lebih API model bahasa, dan memutuskan, bagi setiap permintaan, model atau penyedia mana yang mengendalikannya, menggunakan peraturan seperti kos, kependaman, ketersediaan atau jenis tugas. - [Logprobs (kebarangkalian log)](https://broadrouter.com/ms/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Logprobs ialah kebarangkalian log yang diberikan oleh model bahasa kepada token yang dijananya. API yang memulangkannya menunjukkan, bagi setiap token output, sejauh mana model menganggap token itu mungkin dan, secara pilihan, token alternatif yang paling mungkin pada kedudukan tersebut. - [LoRA (low-rank adaptation)](https://broadrouter.com/ms/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation) ialah kaedah fine-tuning cekap parameter yang membekukan pemberat model pra-latih dan melatih sepasang matriks kecil berpangkat rendah untuk lapisan terpilih. Kemas kini yang dipelajari boleh disimpan sebagai adapter kecil atau digabungkan ke dalam pemberat, tanpa menambah kependaman inferens. - [Lost in the middle (hilang di tengah)](https://broadrouter.com/ms/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle ialah kecenderungan model bahasa yang diperhatikan untuk menggunakan maklumat relevan dengan paling baik apabila ia muncul berhampiran awal atau akhir input panjang, dan paling teruk apabila ia berada di tengah. Ia dinamakan dalam kajian 2023 oleh Liu et al. - [max_tokens](https://broadrouter.com/ms/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens ialah parameter API LLM yang menetapkan bilangan maksimum token yang boleh dijana oleh model dalam satu respons. Apabila had itu dicapai, penjanaan berhenti dan respons ditandakan sebagai terpotong, contohnya dengan finish_reason ditetapkan kepada "length". - [MCP (Model Context Protocol)](https://broadrouter.com/ms/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol) ialah protokol terbuka berasaskan JSON-RPC 2.0 yang menyeragamkan cara aplikasi LLM bersambung kepada alatan dan data luaran. Pelayan MCP mendedahkan tools, resources dan prompts; klien MCP di dalam aplikasi hos menemuinya dan menghantar panggilan alatan daripada model kepada pelayan. - [Mixture of experts (MoE, campuran pakar)](https://broadrouter.com/ms/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Mixture of experts (MoE) ialah reka bentuk rangkaian neural di mana satu lapisan mengandungi banyak subrangkaian pakar dan penghala yang dipelajari menghantar setiap token kepada hanya beberapa daripadanya. Model bahasa MoE mempunyai jumlah parameter yang besar tetapi menggunakan subset aktif yang kecil bagi setiap token. - [Penamatan model (model deprecation)](https://broadrouter.com/ms/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): Penamatan model (deprecation) berlaku apabila penyedia API mengumumkan bahawa sesuatu model tidak akan ditawarkan lagi, biasanya dengan menyatakan tarikh persaraan atau penutupan dan model pengganti yang disyorkan. Selepas tarikh itu, permintaan kepada model tersebut gagal, jadi aplikasi mesti berpindah ke ID model lain lebih awal. - [Fallback model (model sandaran)](https://broadrouter.com/ms/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): Fallback model ialah corak kebolehpercayaan di mana permintaan LLM yang gagal dicuba semula pada model atau penyedia lain daripada senarai tersusun. Pencetusnya boleh jadi ralat, had kadar, ralat panjang konteks atau penolakan dasar kandungan. Ia berjalan dalam klien anda atau dalam gateway di hadapan API. - [Lesen model (model license)](https://broadrouter.com/ms/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): Lesen model ialah terma undang-undang yang menentukan cara pemberat, kod dan output sesuatu model boleh digunakan, diubah dan dikongsi. Model open-weight menggunakan lesen perisian standard seperti Apache 2.0 atau MIT, atau lesen model khusus yang boleh menambah syarat, seperti had ke atas perkhidmatan komersial yang besar. - [Bahasa Arab Standard Moden (MSA)](https://broadrouter.com/ms/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): Bahasa Arab Standard Moden (MSA) ialah bentuk bahasa Arab yang diseragamkan untuk penulisan, berita, pendidikan dan pertuturan rasmi di seluruh dunia Arab. Ia bukan bahasa harian asli sesiapa: orang bertutur dalam dialek serantau seperti bahasa Arab Mesir atau Teluk dan beralih kepada MSA dalam situasi rasmi. - [Sistem berbilang ejen (multi-agent system)](https://broadrouter.com/ms/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): Sistem berbilang ejen ialah aplikasi LLM di mana beberapa ejen, masing-masing dengan prompt, alatan dan konteks sendiri, mengerjakan bahagian berlainan bagi satu tugas. Reka bentuk yang lazim ialah orchestrator-worker: ejen ketua merancang, memulakan ejen pekerja secara selari dan menggabungkan hasil mereka. - [AI multimodal (multimodal AI)](https://broadrouter.com/ms/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): AI multimodal ialah model atau sistem yang berfungsi dengan lebih daripada satu jenis data, seperti teks, imej, video dan audio. Model multimodal mungkin membaca foto dan soalan bersama-sama, atau menerima pertuturan dan menjawab dalam teks. Model bahasa-penglihatan (vision-language model) ialah jenis yang paling lazim. - [Pengecaman entiti bernama (NER)](https://broadrouter.com/ms/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): Pengecaman entiti bernama (NER) ialah tugas mencari nama dalam teks dan melabel setiap satu dengan jenisnya, seperti orang, organisasi atau lokasi. Sistem klasik menanda setiap token; LLM pula sering diminta memulangkan senarai entiti sebagai JSON, yang kemudian disemak dengan teks sumber. - [Pengecaman aksara optik (OCR)](https://broadrouter.com/ms/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): Pengecaman aksara optik (OCR) ialah teknologi yang mengesan dan menukar teks bercetak atau tulisan tangan dalam imej, imbasan atau PDF kepada aksara yang boleh dibaca mesin. OCR klasik mengeluarkan teks dan kedudukannya pada halaman; model bahasa-penglihatan pula boleh membaca teks dan menjawab soalan tentangnya secara terus. - [Ogg Opus](https://broadrouter.com/ms/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus ialah audio yang dikodkan dengan codec Opus (RFC 6716) dan disimpan dalam bekas Ogg (RFC 7845). Ia dibina untuk pertuturan dan audio interaktif pada kadar bit rendah, dan merupakan format yang digunakan WhatsApp untuk nota suara, biasanya sebagai fail .ogg atau .opus. - [Model open-weight (model pemberat terbuka)](https://broadrouter.com/ms/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): Model open-weight ialah model pembelajaran mesin yang pemberat terlatihnya diterbitkan untuk dimuat turun dan dijalankan oleh sesiapa sahaja, biasanya di bawah lesen bersyarat. Ia berbeza daripada AI sumber terbuka sepenuhnya, yang turut mengeluarkan kod latihan dan maklumat terperinci tentang data latihan supaya orang lain boleh mengkaji dan menghasilkan semula model itu. - [API serasi OpenAI (OpenAI-compatible API)](https://broadrouter.com/ms/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): API serasi OpenAI ialah API HTTP yang menerima format permintaan dan respons yang sama seperti OpenAI API, contohnya POST /v1/chat/completions, jadi SDK dan alat OpenAI boleh menggunakannya dengan hanya menukar base URL, kunci API dan nama model. - [Panggilan alatan selari (parallel tool calls)](https://broadrouter.com/ms/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Panggilan alatan selari berlaku apabila model memulangkan lebih daripada satu panggilan alatan dalam satu respons, contohnya tiga bacaan fail serentak. Dalam format OpenAI, ia tiba sebagai beberapa entri dalam tool_calls, dan parallel_tool_calls yang ditetapkan kepada false mengehadkan model kepada sifar atau satu panggilan setiap giliran. - [Bayar ikut guna (pay-as-you-go)](https://broadrouter.com/ms/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): Bayar ikut guna ialah model harga di mana anda hanya membayar apa yang anda guna, dimeter mengikut unit seperti token, permintaan atau minit audio, tanpa langganan atau komitmen minimum. Kebanyakan API LLM mengebil dengan cara ini, daripada baki prabayar atau invois berkala. - [PDPL (undang-undang perlindungan data Saudi dan UAE)](https://broadrouter.com/ms/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) ialah nama umum bagi undang-undang data peribadi Arab Saudi (Dekri Diraja M/19, dikuatkuasakan oleh SDAIA) dan Federal Decree-Law No. 45 of 2021 UAE. Kedua-duanya mengawal pemprosesan data peribadi dan menetapkan syarat pemindahan ke luar negara, termasuk prompt yang dihantar ke API LLM asing. - [PII (maklumat pengenalan peribadi)](https://broadrouter.com/ms/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): PII (personally identifiable information) ialah maklumat yang boleh digunakan untuk membezakan atau menjejak identiti seseorang, seperti nama, nombor pengenalan atau rekod biometrik, serta maklumat lain yang dikaitkan atau boleh dikaitkan dengan orang itu. Istilah GDPR yang lebih luas ialah data peribadi. Dalam kerja LLM, penyuntingan PII membuangnya daripada prompt sebelum ia keluar dari sistem anda. - [Prefill dan decode](https://broadrouter.com/ms/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill dan decode ialah dua fasa inferens LLM. Dalam prefill, model memproses seluruh prompt secara selari dan mengisi KV cache; dalam decode, ia menjana jawapan satu token pada satu masa, menggunakan semula cache itu. Prefill kebanyakannya menentukan masa ke token pertama; decode menentukan kelajuan output. - [Kredit prabayar (prepaid credit)](https://broadrouter.com/ms/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): Kredit prabayar ialah wang yang ditambah ke akaun API sebelum digunakan. Setiap permintaan menolak kosnya daripada baki, dan apabila baki habis, permintaan bermeter berhenti sehingga anda menambah nilai. Ia mengehadkan perbelanjaan kepada jumlah yang telah anda bayar. - [Kejuruteraan prompt (prompt engineering)](https://broadrouter.com/ms/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): Kejuruteraan prompt ialah proses menulis, menyusun dan menguji arahan yang diberikan kepada model bahasa supaya ia sentiasa menghasilkan output yang memenuhi keperluan anda. Ia merangkumi mesej peranan, arahan yang jelas, contoh, format, dan pengukuran hasil berbanding kes ujian. - [Prompt injection (suntikan prompt)](https://broadrouter.com/ms/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): Prompt injection ialah serangan di mana teks yang diberikan kepada model bahasa, sama ada oleh pengguna atau tersembunyi dalam laman web, e-mel, fail atau hasil alatan, mengubah tingkah laku model bertentangan dengan niat pembangun. OWASP meletakkannya di tempat pertama (LLM01:2025) dalam Top 10 for LLM Applications. - [Kuantisasi (quantization)](https://broadrouter.com/ms/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): Kuantisasi ialah proses menyimpan nombor rangkaian neural, terutamanya pemberatnya, dalam format berketepatan lebih rendah seperti 8-bit atau 4-bit dan bukan 16-bit, supaya model menggunakan kurang memori dan selalunya berjalan lebih pantas. Kaedah yang baik memastikan kehilangan ketepatan kekal kecil. - [RAG (retrieval-augmented generation)](https://broadrouter.com/ms/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG (retrieval-augmented generation) ialah corak di mana aplikasi terlebih dahulu mendapatkan petikan yang berkaitan dengan soalan daripada stor luaran, seperti indeks carian atau vektor, kemudian menambahkannya pada prompt supaya model bahasa boleh menjawab berdasarkan teks itu dan bukan daripada ingatannya sahaja. - [Had kadar (rate limit)](https://broadrouter.com/ms/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): Had kadar ialah had yang ditetapkan oleh API ke atas jumlah penggunaan klien dalam satu tempoh masa, seperti permintaan seminit atau token seminit. Permintaan yang melebihi had ditolak, biasanya dengan HTTP 429 Too Many Requests, sehingga penggunaan turun semula di bawahnya. - [ReAct (reason and act)](https://broadrouter.com/ms/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (reason and act) ialah corak prompt yang diperkenalkan oleh Yao et al. pada 2022, di mana model bahasa menyelang-seli jejak penaakulan dengan tindakan seperti carian atau panggilan alatan. Ia membaca setiap pemerhatian, mengemas kini rancangannya, dan mengulang sehingga dapat memberi jawapan akhir. - [Token penaakulan (reasoning tokens)](https://broadrouter.com/ms/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): Token penaakulan ialah token yang dijana oleh model penaakulan untuk menyelesaikan masalah sebelum ia menulis jawapan yang kelihatan. Token ini biasanya dibilkan sebagai token output, dikira dalam had output dan tetingkap konteks, dan mungkin disembunyikan atau diringkaskan oleh API. - [Tamat masa permintaan (request timeout)](https://broadrouter.com/ms/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): Tamat masa permintaan (request timeout) ialah tempoh paling lama klien HTTP menunggu pelayan sebelum meninggalkan permintaan. Klien biasanya membahagikannya kepada connect timeout untuk membuka sambungan dan read timeout untuk menunggu data. Panggilan LLM memerlukan read timeout yang longgar kerana jawapan panjang mengambil masa. - [Reranking (penyusunan semula kedudukan)](https://broadrouter.com/ms/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Reranking ialah peringkat dapatan semula kedua di mana model yang lebih tepat tetapi lebih perlahan, biasanya cross-encoder, menskor setiap pasangan pertanyaan dan dokumen calon daripada carian pertama dan menyusunnya semula, supaya hanya beberapa yang paling relevan dihantar kepada model bahasa. - [Teks kanan ke kiri (right-to-left, RTL)](https://broadrouter.com/ms/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): Teks kanan ke kiri (RTL) ialah tulisan, seperti bahasa Arab atau Ibrani, yang dipaparkan dari kanan ke kiri. Komputer menyimpannya dalam susunan bacaan logik, dan Unicode Bidirectional Algorithm menentukan susunan paparan apabila teks RTL bercampur dengan teks kiri ke kanan, digit, nama model atau kod. - [Kadar sampel (sample rate)](https://broadrouter.com/ms/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): Kadar sampel ialah bilangan sampel audio yang diambil setiap saat, diukur dalam hertz (Hz). Model pengecaman pertuturan lazimnya menjangkakan audio mono 16,000 Hz (16 kHz). Bersama kedalaman bit dan bilangan saluran, kadar sampel menentukan saiz fail audio tanpa mampatan bagi setiap minit. - [Parameter seed](https://broadrouter.com/ms/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): Parameter seed ialah integer yang anda hantar bersama permintaan supaya API cuba memilih token yang sama apabila permintaan yang sama diulang. OpenAI menyifatkannya sebagai usaha terbaik: determinisme tidak dijamin, dan medan system_fingerprint menunjukkan apabila konfigurasi backend telah berubah. - [Inferens tanpa pelayan (serverless inference)](https://broadrouter.com/ms/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): Inferens tanpa pelayan (serverless inference) ialah cara menjalankan model pembelajaran mesin di mana platform memulakan, menskala dan menghentikan pengkomputeran di sebalik endpoint secara automatik, selalunya hingga sifar apabila tiada trafik, dan mengebil penggunaan dan bukannya mesin yang ditempah. Kelemahannya ialah kemungkinan kelewatan cold start. - [SLA (perjanjian tahap perkhidmatan)](https://broadrouter.com/ms/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): SLA (service level agreement, perjanjian tahap perkhidmatan) ialah komitmen kontrak daripada penyedia terhadap tahap perkhidmatan yang boleh diukur, biasanya peratusan uptime bulanan, dengan remedi yang ditetapkan, biasanya kredit perkhidmatan, jika tidak dicapai. SLA mentakrifkan apa yang dikira sebagai downtime, apa yang dikecualikan dan cara membuat tuntutan. - [Speculative decoding (penyahkodan spekulatif)](https://broadrouter.com/ms/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): Speculative decoding ialah teknik inferens di mana model draf yang kecil dan pantas mencadangkan beberapa token seterusnya, dan model sasaran yang besar mengesahkannya dalam satu forward pass, lalu mengekalkan token yang diterima. Ia mempercepat penjanaan tanpa mengubah taburan output. - [Speech-to-text (pertuturan kepada teks)](https://broadrouter.com/ms/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): Speech-to-text, juga dipanggil pengecaman pertuturan automatik (ASR), ialah teknologi yang menukar audio pertuturan kepada teks bertulis. API speech-to-text menerima fail atau strim audio dan memulangkan transkrip, biasanya berharga setiap minit audio. Ketepatan diukur dengan kadar ralat perkataan dan berbeza mengikut loghat, dialek dan kualiti rakaman. - [Had perbelanjaan (spend limit)](https://broadrouter.com/ms/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): Had perbelanjaan ialah had ke atas jumlah yang boleh dibelanjakan oleh akaun, projek atau kunci API dalam satu tempoh, biasanya sebulan. Apabila penggunaan mencapainya, penyedia menolak permintaan seterusnya, selalunya dengan ralat HTTP 429 atau 400, sehingga tempoh itu ditetapkan semula atau had dinaikkan. - [Sari kata SRT (SRT subtitles)](https://broadrouter.com/ms/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) ialah format sari kata teks biasa. Setiap kapsyen terdiri daripada nombor urutan, masa mula dan tamat yang ditulis sebagai hh:mm:ss,mmm --> hh:mm:ss,mmm, satu atau lebih baris teks, dan satu baris kosong. WebVTT ialah format web serupa yang digunakan oleh video HTML. - [Server-sent events (SSE)](https://broadrouter.com/ms/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Server-sent events (SSE) ialah piawaian web untuk pelayan menolak aliran event teks kepada klien melalui satu respons HTTP yang kekal lama dengan jenis kandungan text/event-stream. Setiap event terdiri daripada satu atau lebih baris medan, seperti data:, yang ditamatkan dengan baris kosong. - [Jujukan henti (stop sequences)](https://broadrouter.com/ms/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): Jujukan henti (stop sequences) ialah rentetan yang anda hantar bersama permintaan supaya model berhenti menjana sebaik sahaja ia menghasilkan salah satu daripadanya. OpenAI menamakan medan itu stop dan menerima sehingga 4 jujukan; Anthropic menamakannya stop_sequences. Teks yang sepadan biasanya tidak dimasukkan dalam output yang dipulangkan. - [Penstriman (streaming)](https://broadrouter.com/ms/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): Penstriman ialah cara memulangkan respons model bahasa dalam cebisan kecil semasa ia masih dijana, dan bukan sebagai satu jawapan di penghujung. API serasi OpenAI menstrim melalui server-sent events apabila permintaan menetapkan stream kepada true. - [Output berstruktur (structured output)](https://broadrouter.com/ms/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): Output berstruktur ialah ciri API LLM yang mengehadkan balasan model kepada format yang ditetapkan oleh pembangun, biasanya JSON Schema, supaya kod boleh menghuraikannya dengan boleh dipercayai. Ia lebih ketat daripada JSON mode, yang menjamin JSON yang sah tetapi bukan bentuk tertentu. - [Subejen (subagent)](https://broadrouter.com/ms/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): Subejen (subagent) ialah ejen berasingan yang dimulakan oleh ejen utama untuk mengendalikan satu tugas yang khusus. Ia berjalan dengan tetingkap konteks, prompt sistem dan akses alatan sendiri, melakukan kerja itu, dan hanya memulangkan hasil atau ringkasan kepada ejen utama, supaya perbualan utama kekal kecil. - [Subpemproses (subprocessor)](https://broadrouter.com/ms/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): Subpemproses ialah pihak ketiga yang dilantik oleh pemproses untuk menjalankan pemprosesan data peribadi bagi pihak pengawal. Bagi API LLM, subpemproses boleh termasuk hos GPU, penyedia inferens rakan kongsi dan kapasiti sandaran. Di bawah GDPR Perkara 28, mereka memerlukan kebenaran pengawal dan obligasi perlindungan data yang sama. - [Prompt sistem (system prompt)](https://broadrouter.com/ms/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): Prompt sistem ialah teks arahan yang dihantar oleh aplikasi kepada model bahasa sebelum perbualan, biasanya sebagai mesej dengan peranan system atau developer. Ia menetapkan peranan, peraturan, nada dan format output model, dan model sepatutnya mengutamakannya berbanding mesej pengguna yang bercanggah. - [Ta marbutah (taa marbuta)](https://broadrouter.com/ms/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): Ta marbutah (ة, U+0629) ialah huruf Arab yang hanya muncul di hujung perkataan, selalunya menandakan kata nama dan kata sifat feminin. Rupanya seperti ha (ه, U+0647) dengan dua titik, jadi teks yang ditaip sering mencampurkan kedua-duanya, dan banyak penormal carian menukar ة kepada ه. - [Temperature (suhu pensampelan)](https://broadrouter.com/ms/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): Temperature ialah parameter pensampelan yang mengawal tahap rawak output model bahasa. Model membahagikan skor setiap token dengan temperature sebelum menukarnya kepada kebarangkalian: nilai rendah membuatkan token paling mungkin mendominasi, manakala nilai tinggi meratakan taburan supaya token yang kurang mungkin lebih kerap dipilih. - [Teks kepada pertuturan (text-to-speech, TTS)](https://broadrouter.com/ms/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): Teks kepada pertuturan (text-to-speech, TTS), atau sintesis pertuturan, menukar teks bertulis kepada audio pertuturan. API TTS menerima teks, suara dan format output seperti MP3 atau WAV, lalu memulangkan pertuturan yang dijana. Ia kebalikan speech-to-text, yang menukar audio kepada teks. - [Daya pemprosesan (throughput)](https://broadrouter.com/ms/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): Daya pemprosesan (throughput), dalam penyajian LLM, ialah jumlah kerja yang disiapkan sistem bagi setiap unit masa merentas semua permintaan serentak, biasanya diukur sebagai token output sesaat atau permintaan sesaat. Ia berbeza daripada kelajuan setiap strim, iaitu sepantas mana satu permintaan menerima tokennya. - [Masa ke token pertama (TTFT)](https://broadrouter.com/ms/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): Masa ke token pertama (TTFT) ialah masa dari penghantaran permintaan kepada model bahasa sehingga token output pertama tiba. Ia termasuk masa rangkaian, giliran menunggu di pelayan dan prefill, iaitu langkah model membaca seluruh prompt, jadi ia meningkat dengan panjang prompt dan beban. - [Cap masa (timestamps) dalam speech-to-text](https://broadrouter.com/ms/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): Cap masa dalam speech-to-text ialah masa mula dan tamat, dalam saat, yang dilampirkan pada bahagian transkrip. Cap masa segmen meliputi satu frasa atau ayat; cap masa perkataan meliputi setiap perkataan. Ia diperlukan untuk sari kata, carian dalam rakaman dan menyelaraskan teks dengan penutur. - [Token](https://broadrouter.com/ms/glossary/token/) (translation of https://broadrouter.com/glossary/token/): Token ialah unit teks yang dibaca dan ditulis oleh model bahasa: satu perkataan, sebahagian perkataan, tanda baca atau jujukan bait yang dihasilkan oleh tokenizer model itu. API LLM mengira, mengehadkan dan biasanya mengebil penggunaan dalam token. - [Tokenizer (pengtoken)](https://broadrouter.com/ms/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): Tokenizer ialah komponen yang menukar teks kepada jujukan ID token integer yang dibaca oleh model bahasa, dan menukar ID yang dijana kembali kepada teks. Setiap model mempunyai tokenizer sendiri, jadi teks yang sama boleh menghasilkan bilangan token yang berbeza pada model yang berbeza. - [Token sesaat (tokens per second)](https://broadrouter.com/ms/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): Token sesaat (TPS) ialah kadar model bahasa menjana token output, biasanya diukur bagi satu permintaan selepas token pertamanya tiba. Ia menentukan sepantas mana jawapan distrim sebaik bermula. Apabila dijumlahkan merentas semua permintaan serentak, ia menjadi daya pemprosesan sistem; masa menunggu sebelum token pertama diukur secara berasingan. - [Tool calling (panggilan alatan)](https://broadrouter.com/ms/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): Tool calling, juga dipanggil function calling, ialah ciri API LLM di mana model memulangkan permintaan berstruktur untuk menjalankan fungsi yang ditakrifkan oleh pembangun, bersama argumen JSON. Aplikasi menjalankan fungsi itu dan menghantar hasilnya kembali; model sendiri tidak pernah menjalankan kod. - [tool_choice](https://broadrouter.com/ms/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice ialah parameter permintaan yang mengawal penggunaan alatan. Dalam Chat Completions OpenAI, nilainya boleh auto (model membuat keputusan), required (model mesti memanggil satu atau lebih alat), none (tiada alat), atau fungsi bernama untuk dipaksa. Pilihan setara Anthropic ialah auto, any, tool dan none. - [Top-p (pensampelan nukleus)](https://broadrouter.com/ms/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Top-p, juga dipanggil pensampelan nukleus (nucleus sampling), ialah kaedah penyahkodan di mana model bahasa memilih token seterusnya hanya daripada set terkecil token paling mungkin yang jumlah kebarangkaliannya mencapai ambang p, seperti 0.9, dan membuang ekor yang tidak mungkin. - [Penormalan Unicode (Unicode normalization)](https://broadrouter.com/ms/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Penormalan Unicode menukar teks kepada bentuk piawai supaya rentetan yang kelihatan sama menggunakan code point yang sama. Empat bentuknya ialah NFD, NFC, NFKD dan NFKC; bentuk K juga menggabungkan aksara keserasian, seperti bentuk persembahan Arab dan ligatur, kepada huruf biasa. - [Pengesanan aktiviti suara (VAD)](https://broadrouter.com/ms/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): Pengesanan aktiviti suara (VAD) ialah langkah pemprosesan isyarat yang mengelaskan bingkai audio pendek sebagai pertuturan atau bukan pertuturan. Sistem pertuturan menggunakannya untuk melangkau senyap, memecahkan rakaman panjang pada jeda, mengesan bila penutur selesai bercakap, dan mengelak daripada mentranskripsi bunyi latar. - [Pangkalan data vektor (vector database)](https://broadrouter.com/ms/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): Pangkalan data vektor ialah stor data yang menyimpan vektor embedding, selalunya bersama teks sumber dan metadata, dan menjawab pertanyaan jiran terdekat: berikan vektor pertanyaan dan ia memulangkan vektor tersimpan yang paling serupa dengannya, biasanya mengikut jarak kosinus, jarak Euclidean atau hasil darab dalam. - [Model bahasa-penglihatan (vision-language model)](https://broadrouter.com/ms/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): Model bahasa-penglihatan (VLM) ialah model yang menerima imej dan teks bersama sebagai input dan menghasilkan teks, jadi ia boleh menerangkan gambar, menjawab soalan tentangnya dan membaca dokumen, carta atau tangkapan skrin. VLM juga dipanggil model image-text-to-text atau model multimodal. - [vLLM](https://broadrouter.com/ms/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM ialah pustaka sumber terbuka untuk inferens dan penyajian LLM, yang mula-mula dibangunkan di Sky Computing Lab, UC Berkeley. Ia mengurus memori KV cache dengan PagedAttention, mengumpulkan permintaan masuk secara berterusan, menyokong banyak format kuantisasi dan menyediakan pelayan HTTP yang serasi OpenAI. - [Saiz perbendaharaan kata (vocabulary size)](https://broadrouter.com/ms/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): Saiz perbendaharaan kata ialah bilangan token berbeza dalam tokenizer model bahasa, setiap satu dengan ID dan embeddingnya sendiri. Perbendaharaan kata yang lebih besar boleh mengekod teks biasa dalam lebih sedikit token, tetapi ia membesarkan lapisan embedding dan output, dan hanya membantu untuk teks yang digunakan untuk melatihnya. - [WAV](https://broadrouter.com/ms/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format, atau WAVE) ialah bekas audio yang dibina di atas format RIFF Microsoft dan IBM. Ia biasanya memuatkan sampel PCM linear tanpa mampatan, jadi saiznya boleh diramal: kadar sampel darab bait setiap sampel darab bilangan saluran, bagi setiap saat. - [Kadar ralat perkataan (word error rate, WER)](https://broadrouter.com/ms/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): Kadar ralat perkataan (WER) ialah ukuran ketepatan standard untuk pengecaman pertuturan: bilangan penggantian, pemadaman dan penyisipan perkataan yang diperlukan untuk menukar transkrip kepada rujukan yang betul, dibahagi dengan bilangan perkataan dalam rujukan. Lebih rendah lebih baik. - [Zero data retention (ZDR, tiada pengekalan data)](https://broadrouter.com/ms/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): Zero data retention (ZDR) ialah dasar pengendalian data di mana penyedia AI memproses prompt dan output hanya untuk melayan permintaan dan tidak menyimpan kandungan itu selepasnya, sama ada dalam log, cache atau data latihan. Rekod tanpa kandungan seperti bilangan token dan masa biasanya masih disimpan. ### Japanese (ja, 日本語): 294 translated pages - [Articles in Japanese](https://broadrouter.com/ja/blog/) - [LLM API の料金改定まとめ 2026 年 10 月:値上げ 3 件、値下げなし](https://broadrouter.com/ja/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 2026 年 10 月、LLM API 7 社中 5 社の料金ページに 23 件の変更。値上げ 3 件、値下げなし、新規価格 9 モデル、掲載終了 11 モデル。各変更のアーカイブリンク付き。 - [AI コーディングエージェント 1 回の実行コスト:ステップ数、コンテキストの増加、キャッシュ](https://broadrouter.com/ja/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): コーディングエージェントは長く動くほど 1 ステップあたりのコストが上がる理由、キャッシュによる総額の変化、定額セッションがトークン課金に勝つ条件。トークン計算の例付きです。 - [LLM でアラビア語に母音記号を付ける:プロンプト、検証、DER](https://broadrouter.com/ja/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): LLM API でアラビア語の母音記号(タシュキール)を復元する方法。文字を変えないプロンプト、改変を弾くチェック、DER・WER による評価、CATT を選ぶべき場面を解説します。 - [LLM でアラビア語の文法・スペルを校正する方法](https://broadrouter.com/ja/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): LLM API でアラビア語のスペルと文法を校正する方法。最小限の修正を求めるプロンプト、書き換えを防ぐガード、誤り種別ごとの単語単位の修正、M2 方式の評価を解説します。 - [アラビア語の会議録音から議事録を作る:API パイプライン](https://broadrouter.com/ja/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): アラビア語の会議録音を、決定事項とアクションアイテム付きの議事録に変える方法。文字起こし、整形、LLM での JSON 要約、氏名と日付の確認までを解説します。 - [LLM によるアラビア語の固有表現抽出:JSON、日付、チェック](https://broadrouter.com/ja/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): アラビア語の人名、組織、地名、ヒジュラ暦の日付を LLM で JSON に抽出する方法。プロンプト、ダイアクリティカルマークを無視するスパンチェック、日付変換、採点スクリプトを紹介します。 - [アラビア語 OCR API 比較:Google、Azure、Textract、Mistral、LLM](https://broadrouter.com/ja/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): アラビア語の印刷文字と手書き文字に公式対応している OCR API、それぞれが返すもの、1,000 ページあたりの料金を、ビジョン LLM API と比較して解説します。 - [音声認識 API でアラビア語の SRT 字幕を作るには](https://broadrouter.com/ja/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): SRT ファイルに必要なもの、アラビア語でタイムスタンプや SRT を返す音声認識 API、テキストのみの API から字幕を作る検証済みのパイプラインを解説します。 - [LLM によるアラビア語テキスト分類:感情、意図、振り分け](https://broadrouter.com/ja/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): LLM API でアラビア語のレビュー、メッセージ、サポートチケットを分類する方法。ラベル設計、方言を考慮したプロンプト、JSON 出力、macro F1 でのテスト、大量処理時のコストを解説します。 - [検索と RAG のためのアラビア語テキスト正規化:変えるもの、変えないもの](https://broadrouter.com/ja/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): タシュキール、アリフの異体、ター・マルブータ、ハムザの台、ペルシア文字、数字、Unicode。検索に効くアラビア語の正規化ルールと害になるルールを、テスト済みの Python コードとともに解説します。 - [LLM でアラビア語テキストを要約する:プロンプト、長さ、評価](https://broadrouter.com/ja/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): アラビア語の記事、レポート、方言メッセージを LLM API で要約する方法。プロンプトの型、語数の制御、数字のチェック、そして既定の ROUGE がアラビア語に 0 点を付ける理由を解説します。 - [Artificial Analysis の代替サイト:2026 年に LLM を比較できる場所](https://broadrouter.com/ja/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Artificial Analysis の代替サイトを比較:Arena、Epoch AI、Vals AI、OpenRouter ランキング、Hugging Face、価格データ。それぞれ何を測り、いつ使うべきかをまとめました。 - [BroadNet の GPU ハードウェアとサービング方式(BroadNet の説明による)](https://broadrouter.com/ja/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): BroadNet Inference を支えるハードウェアについて BroadNet が説明していること。B300 から顧客先アプライアンスまでの 3 層の GPU、GPU ごとのスループット、テストから得た教訓。 - [LLM API のレスポンスをキャッシュしてコスト削減:Python の完全一致キャッシュ](https://broadrouter.com/ja/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): テスト済みの Python 製 LLM API レスポンス完全一致キャッシュ。キーに含めるもの、キャッシュしてはいけないもの、TTL、プライバシー、作る前にヒット率を測る方法を解説します。 - [Claude Code でオープンモデルを使う:プロキシ、ネイティブエンドポイント、制限](https://broadrouter.com/ja/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code は Anthropic Messages API を使います。LiteLLM プロキシやネイティブのエンドポイント経由でオープンモデルを動かす方法と、その途中で起きる問題を解説します。 - [文書 AI のワークフローに人手のレビューを組み込む](https://broadrouter.com/ja/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): 文書 AI パイプラインの人手レビュー工程を設計します。検証ルール、人に回す基準、ランダムな抜き取り監査、監査証跡、レビュー率まで。 - [エジプトのデータ保護法(2020 年法律第 151 号)と LLM API](https://broadrouter.com/ja/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): エジプトの人々の個人データを国外の LLM や音声認識 API に送る場合に、2020 年法律第 151 号と 2025 年決定第 816 号が移転、ライセンス、漏えいについて定める内容を解説します。 - [ファインチューニングの前に:解決すべき問題を定義する](https://broadrouter.com/ja/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): プロンプト、RAG、評価で問題が解決しファインチューニングが不要な場合、必要な学習データ、そして依頼する前にファインチューニングをテストする方法。 - [LLM が返す不正な JSON の直し方:原因・修復・リトライ](https://broadrouter.com/ja/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): LLM が壊れた JSON を返す理由(コードフェンス、前後の文章、max_tokens での途切れ、末尾のカンマ)と、安全に直せるものだけを直し残りは拒否する、テスト済みの Python 関数を紹介します。 - [LLM API の料金はどのくらいの頻度で変わる?3 年分のアーカイブ調査](https://broadrouter.com/ja/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): 11 社の LLM API について、2023〜2026 年のアーカイブ済み料金ページ 308 件を調べました。80% のモデルは一度も価格が変わらず、定価変更の 75% は DeepInfra でした。 - [LLM API のコストを削減する方法:順番どおりに進める 10 ステップ](https://broadrouter.com/ja/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): LLM API の請求額を正しい順番で下げる方法。計測、プレフィックスのキャッシュ、プロンプトの削減、出力と推論の上限設定、モデルの適正化、バッチ処理、定額プランが得になる条件を解説します。 - [LLM API のフェイルオーバー:自前のコードで別プロバイダーへ切り替える](https://broadrouter.com/ja/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): OpenAI 互換の LLM プロバイダー間でフェイルオーバーする方法。切り替えるべきエラー、サーキットブレーカー、モデル ID の対応付け、ストリーミング、費用とプライバシー条件の揃え方を解説します。 - [LLM API の値下げ 2023〜2026:大幅値下げを出典付きで](https://broadrouter.com/ja/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): 2023〜2026 年の LLM API の大幅値下げを、変更前後の Internet Archive リンク付きでまとめました。o3、Codestral、Gemini 1.5、そしてあまり知られていない値上げも。 - [LLM API の料金は下がる一方?52 件の値上げと予算の立て方](https://broadrouter.com/ja/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): LLM API の料金は上がることもあります。2024〜2026 年のアーカイブ済み料金ページで、同じモデルの定価値上げを 52 件確認しました。中央値は 40%。全出典と予算ルール付き。 - [LLM API のレート制限をプロバイダー別に比較:RPM・TPM・同時実行数](https://broadrouter.com/ja/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): OpenAI、Anthropic、Gemini、Groq、DeepSeek、Fireworks など 13 社の LLM API について、エントリー階層のレート制限を公式ドキュメントから整理し、RPM と TPM の意味も解説します。 - [LLM API の SLA 比較:Azure、Bedrock、Google、OpenAI は何を約束しているか](https://broadrouter.com/ja/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): 各 LLM API プロバイダーが書面で約束している内容。稼働率、何がダウンタイムに数えられるか、サービスクレジット、レイテンシ目標を、各社の SLA ページから読み解きます。 - [Batch API なしで LLM を一括処理する:再開できるランナー](https://broadrouter.com/ja/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Batch API がなくても、chat completions API で 1 万〜100 万件のプロンプトを処理する方法。同時実行の上限、レート制御、リトライ、再開できるチェックポイント、所要時間の計算を解説します。 - [Safety Map:39 の LLM は機密データをどう扱うか](https://broadrouter.com/ja/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet は 14 のラボの LLM 39 個を、認証情報、個人情報、医療・金融データで調査しました。移転する安全性としない安全性、ツール呼び出しと推論からの漏えいを解説します。 - [LLM の機密データ安全性をテストする方法:BroadNet の手法](https://broadrouter.com/ja/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): LLM が認証情報、個人情報、医療・金融データを漏らさないか、導入前にテストするために BroadNet が公開した手法。カテゴリー別、漏えいと見逃し、3 回以上の実行。 - [OpenAI API からの移行:Python、Node、cURL のコード解説](https://broadrouter.com/ja/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): OpenAI から OpenAI 互換 API へ移行するためのコード変更。クライアント設定、Responses から Chat Completions への書き換え、ツール、ストリーミング、embeddings、スモークテストを解説します。 - [BroadNet が評価して不採用にしたモデルとその理由](https://broadrouter.com/ja/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet はテストして不採用にしたモデルを理由付きで公開しています。認証情報の安全性テストに通らなかったチャットモデルや、選ばれなかった音声・OCR・埋め込みモデルです。 - [多言語に対応した AI 体験の設計](https://broadrouter.com/ja/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): アラビア語、英語、スペイン語のユーザー向け AI プロダクトの設計:UI ロケールとモデルの言語、RTL と混在テキスト、数値と日付、言語判定、コストを解説します。 - [オープンウェイトとクローズドモデルの API コスト:価格データで比較](https://broadrouter.com/ja/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): オープンウェイトモデルは API 経由だと GPT、Claude、Gemini より安いのか。各社公式ページのトークン単価、ホスト間の価格差、ユーザーあたりコストの試算で比べます。 - [OpenAI・Anthropic・Google の API 料金推移(2023〜2026)](https://broadrouter.com/ja/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): GPT、Claude、Gemini の API 料金を、2023〜2026 年のアーカイブ済み料金ページで並べて比較。どのラインが安くなり、どれが高くなったかを、アーカイブのリンク付きで示します。 - [OpenAI 互換 API の base_url 一覧:14 社を公式ドキュメントで確認](https://broadrouter.com/ja/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): OpenAI 互換 LLM API 14 社の base_url を、各社の公式ドキュメントから確認しました。ドキュメントに記載のエンドポイントと、そのまま差し替えると失敗する違いもまとめています。 - [OpenClaw と Hermes Agent を安いモデルで動かす:設定とトークン計算](https://broadrouter.com/ja/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): OpenClaw や Hermes Agent を OpenAI 互換エンドポイントにつなぐ方法、定期チェックがトークンを消費する理由、軽いコンテキストと安いモデルで費用を下げる方法を解説します。 - [専用エンドポイントはワークロードの明確化から始まる](https://broadrouter.com/ja/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): LLM の専用エンドポイントを依頼する前に決めておくこと:ワークロード、毎秒トークン数、同時実行数、コンテキスト、リージョン、ライセンス、予算。受け取った後のテスト方法も解説します。 - [コーディングエージェントのプロンプトインジェクション:実際に効く防御](https://broadrouter.com/ja/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): プロンプトインジェクションがコーディングエージェントに届く経路と、ベンダーや研究者が示す多層防御(サンドボックス、外向き通信の制限、許可リスト、承認)を解説します。 - [RAG のチャンクサイズとオーバーラップ:検索精度とコストへの影響](https://broadrouter.com/ja/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): チャンクサイズとオーバーラップが、RAG の再現率、インデックスのサイズ、埋め込みコスト、毎回 LLM に送るトークン数をどう変えるか。公開された結果と具体的な計算で解説します。 - [推論モデルの max_tokens の決め方:予算を見積もる方法](https://broadrouter.com/ja/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): max_tokens が小さいと推論モデルが空の応答を返す理由と、実測したパーセンタイル、コンテキストウィンドウ、段階的に増やすリトライから予算を決める方法を解説します。 - [ブラウザで録音して音声認識 API に送る方法](https://broadrouter.com/ja/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): MediaRecorder でマイクを録音し、WebM や Safari の MP4 を 16 kHz WAV に変換して、API キーを持つ小さなサーバー経由で文字起こし API に送ります。 - [プロンプトのトークンを減らす:システムプロンプト、ツール、データの実測](https://broadrouter.com/ja/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): システムプロンプト、few-shot 例、ツールスキーマ、データ形式、会話履歴、アラビア語の発音記号を削ったときのトークン削減量を、実際のトークナイザーで計測しました。 - [同じモデルでもホストで違う:LLM API の価格差はどれくらいか](https://broadrouter.com/ja/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): 複数のホストが販売する 119 モデルでは、最も高い定価は最も安い定価の 1.76 倍(中央値)でした。41 プロバイダーの公式料金ページのデータ、2026 年 11 月。 - [サウジ・UAE の PDPL と LLM API:越境移転チェックリスト](https://broadrouter.com/ja/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): サウジアラビアや UAE から個人データを LLM API に送る場合に、PDPL の越境移転ルールの内容、プロバイダーへの確認事項、国内での推論が必要になる場面を解説します。 - [テキスト、音声、画像のどれを使う?まずタスクから考える](https://broadrouter.com/ja/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): テキストチャット、画像入力、音声の文字起こしをタスクで選びます。BroadNet が文書化しているモデルと制限、そして TTS など提供していない機能も紹介します。 - [25 MB を超える長い音声の文字起こし:分割・送信・結合](https://broadrouter.com/ja/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): 音声認識 API のファイルサイズ上限を超える録音を文字起こしする方法。ファイルを小さくし、ffmpeg で無音部分で分割し、チャンクを並列送信してテキストを結合します。 - [LLM でアプリの文字列をアラビア語に翻訳:プレースホルダーと複数形](https://broadrouter.com/ja/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): LLM API で i18n JSON ファイルをアラビア語に翻訳する方法。プレースホルダーの固定、アラビア語の 6 つの複数形の生成、CI での ICU 構文チェック、差し込む名前の分離を解説します。 - [アラビア語の音声を API で英語テキストに翻訳する:検証済みのパイプライン](https://broadrouter.com/ja/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): アラビア語の録音を 2 回の API 呼び出しで英語テキストにする方法。language=ar で文字起こしし、LLM で翻訳します。数字のチェック、分割、コスト、評価まで解説します。 - [提供中・要相談・研究用:モデルカタログを正しく読む](https://broadrouter.com/ja/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): カタログのモデルのうち、今日呼び出せるもの、問い合わせが必要なもの、テストされただけのものを見分ける方法を、BroadNet のカタログと /v1/models を例に解説します。 - [vLLM での Uno:BroadNet のサービング研究は何を測ったか](https://broadrouter.com/ja/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet は Uno 自己ドラフトアダプターを vLLM に組み込み、H100 で測定しました。スループットは通常の vLLM の 1 ストリームで 2.58 倍、32 で 1.74 倍。手法、DOI、限界を解説します。 - [Gemma 4 26B での Uno:BroadNet の Mixture of Experts 研究](https://broadrouter.com/ja/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet は Mixture of Experts モデル Gemma 4 26B A4B 向けに初めて Uno 自己ドラフトアダプターを学習させました。測定内容、自己ドラフトが有利な場面、DOI を紹介します。 - [エージェントのトークン予算:エージェントループのトークンはどこに消えるか](https://broadrouter.com/ja/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): システムプロンプト、ツールスキーマ、ファイルの読み込み、履歴、推論。エージェントループがトークンをどう使うか、usage データで各部分を測る方法、最初に削るべきものを解説します。 - [アラビア語音声認識の単語誤り率(WER)をわかりやすく解説](https://broadrouter.com/ja/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): 単語誤り率の意味と計算方法、そしてアラビア語の WER がダイアクリティカルマーク、表記の正規化、方言にこれほど敏感な理由。Python のコードと実際の数値付きで解説します。 - [アラビア語方言の音声認識:湾岸、エジプト、レバント、マグリブ](https://broadrouter.com/ja/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): 音声認識が湾岸、エジプト、レバント、マグリブのアラビア語を苦手とする理由と、自社の音声でプロバイダーの実力がわかる方言テストセットの作り方を解説します。 - [LLM によるアラビア語と英語の翻訳:プロンプト、用語集、評価](https://broadrouter.com/ja/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): LLM API でアラビア語と英語を相互に翻訳する方法。口調を保つプロンプト、固定用語の用語集、数字と名前の扱い、翻訳品質の評価方法を解説します。 - [湾岸・エジプト・レバントのアラビア語:LLM は方言をどう扱うか](https://broadrouter.com/ja/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): LLM は湾岸、エジプト、レバントのアラビア語にどう対応するのか。失敗しやすい点、方言向けのプロンプト、テスト方法、そして BroadNet が Qwen3.8 27B について報告している内容を解説します。 - [アラビア語 OCR と Vision LLM:アラビア語の文書と請求書を読み取る](https://broadrouter.com/ja/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): アラビア語の請求書、レシート、フォームに、従来の OCR と Vision LLM のどちらを使うべきか。精度の落とし穴、JSON 出力、ページあたりのコスト、検証パイプラインを解説します。 - [アラビア語のプロンプトエンジニアリング:文体、RTL、数字、混在テキスト](https://broadrouter.com/ja/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): LLM から質の高いアラビア語を引き出すプロンプトの書き方。文体と方言の指定、数字と日付の制御、RTL 表示の整え方、アラビア語と英語の混在への対処を解説します。 - [アラビア語 RAG:うまくいく正規化、チャンク分割、検索](https://broadrouter.com/ja/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): アラビア語文書での検索拡張生成(RAG)。alef、ya、ta marbuta の正規化、適切なチャンク分割、多言語埋め込みモデルの選び方、キーワード検索の併用を解説します。 - [アラビア語の音声認識 API:2026 年の選び方](https://broadrouter.com/ja/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): アラビア語の文字起こし API の選び方。単語誤り率、方言、ファイル制限、音声 1 時間あたりの料金、プライバシーと、自社の録音で試すテスト計画を解説します。 - [アラビア語 WhatsApp チャットボットの構成:ボイスメッセージから返信まで](https://broadrouter.com/ja/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): アラビア語 WhatsApp アシスタントの実用的な構成。Webhook、キュー、ボイスメッセージの文字起こし、LLM による返信、ポリシー上のルール、コストを、応用できる Python コード付きで解説します。 - [Arabizi と LLM:3 や 7 などの数字を使うラテン文字アラビア語の扱い方](https://broadrouter.com/ja/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): Arabizi は、ع を 3、ح を 7 のような数字で表し、アラビア語をラテン文字で書く表記です。その判定方法、LLM でアラビア文字に変換する方法、正しい返信の仕方を解説します。 - [Aider に最適な LLM:モデルの選び方と接続方法](https://broadrouter.com/ja/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Aider に最適な LLM の選び方:編集形式の遵守、コンテキストサイズと 1 時間あたりのコスト、自分のリポジトリでの 1 時間テスト、そして正確な OPENAI_API_BASE の設定。 - [コーディングエージェントに最適なオープンモデル:ツール、コンテキスト、コストで選ぶ](https://broadrouter.com/ja/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): コーディングエージェントには GLM 5.3、DeepSeek V4.1 Flash、Qwen3.8 27B、Gemma 4 のどれを使うべきか。ツール呼び出し、コンテキスト、推論、タスクあたりのコストで選びます。計算付き。 - [キャッシュ入力トークンとは:LLM API の請求額を下げる方法](https://broadrouter.com/ja/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): キャッシュ入力は通常の入力のごく一部の料金です。プロンプトキャッシュの課金方法、usage の cached_tokens の読み方、キャッシュに乗るプロンプトの組み立て方を解説します。 - [最安の画像対応 LLM API:画像 1 枚・1 ページあたりのコストを計算](https://broadrouter.com/ja/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): 画像対応 LLM の呼び出しは画像 1 枚、文書 1 ページあたりいくらかかるか。Gemma 4、Qwen3.8 27B、DeepSeek V4.1 Flash のトークン計算と、DeepInfra のほうが安いケースを解説します。 - [LLM API プロバイダーの選び方:20 項目のチェックリスト](https://broadrouter.com/ja/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): LLM API を選ぶ前に確認したい 20 の質問:料金、制限、OpenAI 互換性、プライバシー、モデル、サポートと稼働率。採点表とテスト計画付き。 - [アラビア語向け LLM API の選び方:何をテストすべきか](https://broadrouter.com/ja/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): アラビア語アプリ向けの LLM の選び方。方言の理解、文章の質、翻訳、アラビア語の文書と画像、コストに加え、1 日で作れるテストセットを紹介します。 - [Cline vs Roo Code vs Aider:オープンモデルを使うチーム向け比較(2026)](https://broadrouter.com/ja/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline、Roo Code(アーカイブ済み、Zoo Code が後継)、Aider を、オープンモデルで重要な点で比較します:ツール呼び出し、コンテキスト消費、MCP、git、承認フロー。 - [LLM リクエストを送る前にトークン数を数える方法](https://broadrouter.com/ja/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): LLM API を呼ぶ前に、言語とモデルごとにプロンプトのトークン数を見積もる方法。トークナイザーが違う理由、5 言語の実測例、安全な max_tokens の余裕を解説します。 - [コーディングエージェントの LLM コストを削減する 10 の方法(削減額の試算付き)](https://broadrouter.com/ja/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Cline、Aider、opencode や自作のコーディングエージェントのトークン費用を下げる 10 の具体策。それぞれ削減額の計算と測定方法を示します。 - [DeepSeek API の代替手段:DeepSeek モデルを使う 5 つの方法](https://broadrouter.com/ja/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): DeepSeek モデルを公式 API、アグリゲーター、クラウドのモデルカタログ、自前の GPU、BroadNet Inference のどれで呼び出すか。料金、プライバシー、制限のトレードオフを比較します。 - [DeepSeek V4.1 Flash API の料金:トークン単価、キャッシュ、定額プラン](https://broadrouter.com/ja/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): DeepSeek V4.1 Flash API の 100 万トークンあたりの入力・キャッシュ入力・出力料金、月・週・時間単位の定額プラン、そして DeepSeek 公式、DeepInfra、Fireworks との比較です。 - [機能リリース前に LLM API のコストを見積もる方法](https://broadrouter.com/ja/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): 新しい AI 機能のコストモデルを段階的に作ります。ユーザー数 × 操作数 × 呼び出し数 × トークン数に、キャッシュ、推論、リトライ、レート制限を加え、月額の試算表も示します。 - [プロバイダーを乗り換える前に LLM を評価する方法](https://broadrouter.com/ja/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): 実トラフィックから小さな評価セットを作り、コードによるチェックとバイアスを抑えた LLM ジャッジで採点し、移行前に品質・コスト・レイテンシを比較します。 - [ビジョン LLM で請求書・領収書・帳票のデータを JSON に抽出する](https://broadrouter.com/ja/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): 請求書、領収書、帳票のための本番向けパイプライン。画像の前処理、JSON スキーマのプロンプト、Pydantic による検証、金額の検算、リトライ、人手のレビューキューまで。 - [LLM API は定額とトークン課金のどちらが得か:損益分岐の計算](https://broadrouter.com/ja/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): 定額の LLM プランがトークン課金より安くなるのはいつか。DeepSeek V4.1 Flash の月額請求の試算 4 例、損益分岐となる利用量、自社の利用量の測り方を紹介します。 - [LLM API の無料クレジット:初回クレジットで実際どこまで使えるか](https://broadrouter.com/ja/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): LLM API の無料クレジットで実際に何が買えるのか。BroadNet の初回クレジットで使えるトークン、音声の分数、時間制キーを試算し、Gemini と OpenRouter の無料枠と比較します。 - [GDPR と LLM API:プロバイダーに確認すべき 15 の質問](https://broadrouter.com/ja/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): LLM API の GDPR チェックリスト。DPA、サブプロセッサー、データ保持、学習利用、域外移転、侵害通知について、根拠となる GDPR の条文とあわせて確認事項を整理します。 - [GLM 5.3 と DeepSeek V4.1 Flash の比較:どちらをいつ使うか](https://broadrouter.com/ja/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 と DeepSeek V4.1 Flash を、料金、コンテキスト、ツール、構造化出力、画像、推論で比較します。ワークロード別の費用計算と、使い分けのシンプルなルール付きです。 - [LLM API の 429 エラー対処法:バックオフ、ジッター、同時実行制限](https://broadrouter.com/ja/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 と 5xx の違い、フルジッター付き指数バックオフ、LLM API 向けの同時実行リミッターとキューを、本番環境にそのまま使える Python と Node のコード付きで解説します。 - [LLM API の時間制キーとは:1 時間定額の仕組みを解説](https://broadrouter.com/ja/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): BroadNet の時間制キーは、Gemma 4 VL 26B を 1 ストリームで 1 時間、定額で使えるキーです。含まれるもの、制限、向いている人、元が取れる条件を解説します。 - [LLM API のデータ保持を比較:OpenAI、Anthropic、Gemini、Mistral](https://broadrouter.com/ja/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): OpenAI、Anthropic、Gemini API、Mistral、BroadNet は API のプロンプトの保存と学習への利用について公式に何と言っているのか。2026 年 11 月時点の各社ページから引用して比較します。 - [LLM API キーのセキュリティ:保管、ローテーション、IP 許可リスト](https://broadrouter.com/ja/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): LLM API キーをコードやログに残さず、ワークロードごとに別のキーを使い、定期的にローテーションし、サーバーの IP に固定し、漏えい時は数秒で失効させる方法を解説します。 - [LLM API 料金の仕組み:入力・キャッシュ・出力・推論トークンを解説](https://broadrouter.com/ja/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): トークン課金の LLM API 料金はどう計算されるのか。入力、キャッシュ入力、出力、推論トークン、画像、分単位の音声を、公開料金表を使った計算例とともに解説します。 - [機密データを LLM API に送るとき:マスキング、最小化、ログ削減](https://broadrouter.com/ja/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): 顧客・医療・金融データを LLM API に安全に送る方法。項目の分類、元に戻せるプレースホルダーでのマスキング、ゼロデータ保持、ログをきれいに保つ方法を解説します。 - [LLM API の支出管理:予算、キーの制限、ループ対策](https://broadrouter.com/ja/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): LLM API の請求額の暴走を防ぐ方法。前払い残高、アプリごとのキー、IP 許可リスト、利用ログ、コード内の予算、ループの上限、定額のキーを解説します。 - [LLM API の利用ログ:ユーザー別・機能別のコストと PII のマスキング](https://broadrouter.com/ja/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): LLM API の呼び出しごとにトークン、コスト、レイテンシ、リクエスト ID を記録し、支出をユーザーと機能に割り当て、書き込み前に PII をマスキングする方法。Python コード付き。 - [ユーザーあたりの LLM コスト:SaaS の AI 機能のユニットエコノミクス](https://broadrouter.com/ja/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): AI 機能がアクティブユーザー 1 人・1 か月あたりいくらかかるかを計算し、平均よりヘビーユーザーが重要な理由と、モデル費用を賄える SaaS の料金プランの決め方を解説します。 - [推論プロバイダーはオープンモデルをどう価格設定しているか:6 社比較](https://broadrouter.com/ja/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): OpenRouter、Together AI、Fireworks、Groq、DeepInfra、BroadNet がオープンモデルをどう課金しているか。トークン単価、キャッシュ、バッチ、GPU 時間、定額プランを出典付きで比較します。 - [LLM のレイテンシ:TTFT、トークン/秒、エンドツーエンドの違い](https://broadrouter.com/ja/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): 最初のトークンまでの時間(TTFT)、1 秒あたりのトークン数、エンドツーエンドのレイテンシがそれぞれ何を測るのか、クライアント側での計測コード、どの改善策がどの数字に効くのかを解説します。 - [LLM のストリーミング解説:SSE のチャンク、delta、usage、キャンセル](https://broadrouter.com/ja/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): chat completion のストリーミングが通信上どう流れるか、delta から回答を組み立てる方法、最後のチャンクから usage を読む方法、Python・Node・curl できれいにキャンセルする方法を解説します。 - [ロングコンテキストか RAG か:100 万トークンがベクトル DB に代わるとき](https://broadrouter.com/ja/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): 100 万トークンのコンテキストウィンドウがベクトルデータベースの代わりになるのはどんなときか、RAG がまだ勝つのはどんなときかを、クエリ単価、レイテンシ、再現率、保守の手間で比べます。 - [長時間動くエージェントのコンテキスト管理:100 万トークンを活かす](https://broadrouter.com/ja/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): 100 万トークンのコンテキストで、長時間のエージェントを正確かつ低コストに保つ方法。追記型の履歴、キャッシュしやすいプロンプト、圧縮のタイミング、何から削るかを解説します。 - [長時間動くエージェント:崩れないタイムアウト、リトライ、チェックポイント](https://broadrouter.com/ja/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): LLM エージェントを何時間も動かし続けるために。呼び出しごとのタイムアウト、停止を検知するストリーミング、安全なリトライ、冪等なツール副作用、ステップごとのチェックポイント。 - [OpenAI 互換モデルで MCP サーバーを使う(プロバイダーを問わず)](https://broadrouter.com/ja/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP はどの LLM でも使えます。MCP のツールを関数呼び出しに変換するのはクライアントです。Cline、Continue、opencode、n8n、自作の Python コードでオープンモデルと MCP を使う方法を解説します。 - [100 万トークンのコンテキストウィンドウが役立つとき(役立たないとき)](https://broadrouter.com/ja/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): 1,048,576 トークンに何が入るか、DeepSeek V4.1 Flash でいくらかかるか、推論とキャッシュで計算がどう変わるか、そして RAG のほうがまだ良いのはどんなときかを解説します。 - [マルチエージェント構成の本当のコスト:トークン課金 vs 定額セッション](https://broadrouter.com/ja/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): マルチエージェントシステムは、単一エージェントの数倍のトークンを消費します。定額プランとトークン課金のセッション計算、見えない倍率、それぞれが有利になる条件を解説します。 - [多言語の音声認識 API:14 言語を 1 つのエンドポイントで](https://broadrouter.com/ja/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): 英語、スペイン語、フランス語、中国語、アラビア語など 14 言語を、1 分単位で課金される 1 つの OpenAI 互換エンドポイントで文字起こし。アラビア語専用モデルを使うべき場面も解説します。 - [n8n の AI Agent を安い LLM で動かす:設定、モデル選び、1 回あたりのコスト](https://broadrouter.com/ja/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): 低コストの OpenAI 互換モデルで n8n の AI Agent ワークフローを作る方法。ツール向けのモデル選び、トークンを減らす設定、1 万回実行あたりのコスト計算を解説します。 - [ビジネスで使うならオープンウェイトかクローズドか:コスト、管理、リスク](https://broadrouter.com/ja/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): ビジネスにはオープンウェイトとクローズドのどちらの LLM がよいか。コスト、管理、プライバシー、品質差、ロックインを比べ、オープンモデルを動かす 3 つの方法と、それぞれが有利な場面を解説します。 - [OpenAI 互換 API への移行チェックリスト](https://broadrouter.com/ja/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): base_url の変更は一番簡単な部分です。問題が起きるのはモデル ID、推論トークン、ストリーミングの使用量、エラー、レート制限、キーです。コード付きのチェックリストです。 - [AI エージェントには OpenRouter か定額プランか:安いのはどちら?](https://broadrouter.com/ja/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter はトークン課金にクレジット購入時の手数料、定額プランはセッション単位の課金です。常時稼働エージェントのコスト計算、OpenRouter の手数料と制限、向き不向きを解説します。 - [コーディングエージェントを夜間に無人で安全に動かす方法](https://broadrouter.com/ja/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): コーディングエージェントを無人で動かすための実践的な構成。サンドボックス、Git のチェックポイント、タスクファイル、ステップ上限、朝のレビュー、定額プランセッションでの夜間実行を解説します。 - [プレフィックスがキャッシュされるプロンプトの組み立て方](https://broadrouter.com/ja/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): キャッシュされるプレフィックスを同一に保つ LLM プロンプトのレイアウト。システムプロンプト、ツールスキーマ、ドキュメント、追記のみの履歴と、それを確かめるテストを紹介します。 - [推論トークン:見えている以上の出力に課金される理由](https://broadrouter.com/ja/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): 推論モデルは隠れた思考を出力トークンとして課金します。その計測方法、モデルごとのコスト、reasoning_effort・max_tokens・ルーティングによる削減方法を解説します。 - [定額の LLM API で並列コーディングエージェントを動かす方法](https://broadrouter.com/ja/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): 無制限の DeepSeek V4.1 Flash でサブエージェント、オーケストレーター、バッチワーカーを動かす方法。必要なプランセッション数、キーへの割り振り方、同時実行の重なりの抑え方を解説します。 - [LLM のセルフホストと API:本当のコストと損益分岐点](https://broadrouter.com/ja/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): 公式ページの GPU レンタル料金、自前 GPU での 100 万トークンあたりコストの計算式、API より安くなるのに必要な稼働率、見落とされがちな運用コストを解説します。 - [音声認識 API の料金比較:2026 年の 1 分あたりの価格](https://broadrouter.com/ja/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): OpenAI、Google、AWS、Deepgram、AssemblyAI、Speechmatics、Groq、BroadNet Transcribe の 1 分あたりと 1,000 時間あたりの料金を公式ページから比較。追加オプションの料金も掲載します。 - [JSON モードと json_schema:壊れない構造化出力の作り方](https://broadrouter.com/ja/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON モードは正しい JSON を、json_schema は指定した形を保証します。OpenAI 互換モデルでの指定方法、Pydantic での検証、不正な出力をループで修復する方法を解説します。 - [アラビア語と英語の 1 語あたりのトークン数:アラビア語が高くつく理由](https://broadrouter.com/ja/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): 同じ意味でもアラビア語が英語より多くの LLM トークンを使いがちな理由、1 語あたりのトークン数が誤解を招く理由、usage フィールドから自社の比率を測るスクリプトを紹介します。 - [オープンモデルのツール呼び出し:スキーマ、並列呼び出し、リトライ](https://broadrouter.com/ja/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): OpenAI 互換のオープンモデルでの Function Calling。スキーマのコツ、並列ツール呼び出し、ツール呼び出しが失敗する 5 つのパターン、検証してリトライする Python ループを解説します。 - [コールセンターの通話録音を文字起こしする:アラビア語対応のパイプライン](https://broadrouter.com/ja/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): ステレオ通話をチャンネルで分け、発話ターンに切り、アラビア語と英語を文字起こしし、個人情報をマスキングして、LLM で通話ごとに要約します。動く Python パイプラインと費用付き。 - [コーディングエージェントでオープンモデルを使う:Aider、Cline、opencode、Codex](https://broadrouter.com/ja/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Aider、Cline、Roo Code、opencode、Continue、Codex CLI を任意の OpenAI 互換エンドポイントに接続する方法。各ツールの正確なフィールド名と、つまずきやすい設定をまとめました。 - [ビジネス向けボイスメッセージの文字起こし:WhatsApp と Telegram からチケットへ](https://broadrouter.com/ja/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): WhatsApp Business と Telegram ボットに届く顧客のボイスメッセージを、文字起こしとサポートチケットに変えます。ダウンロード、文字起こし、LLM による項目抽出、振り分けまで。 - [短期プロジェクト向け LLM 週間プラン:Orbit、1 時間、1 か月の比較](https://broadrouter.com/ja/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): 1 週間無制限の DeepSeek V4.1 Flash プランが、時間制セッション、月額プラン、トークン課金より得になるのはいつか。ハッカソン、スプリント、移行作業の計算付きで解説します。 - [LLM API プランの「無制限」は本当は何を意味するのか](https://broadrouter.com/ja/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): トークンメーター、5 時間のローリングウィンドウ、週間上限、レート制限、同時実行数。「無制限」の AI プランがエージェントにとって本当に無制限かを見極める 7 つの質問です。 - [Whisper API の代替:乗り換えるべきタイミングとテスト方法](https://broadrouter.com/ja/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): OpenAI Whisper API のホスト型代替が適するのはどんなときか。自分の音声での単語誤り率、言語、1 時間あたりの料金、制限、プライバシーをテストコード付きで解説します。 - [ゼロデータ保持の LLM API:確認すべき 10 の質問チェックリスト](https://broadrouter.com/ja/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): LLM API の「ゼロデータ保持」が本来カバーすべき範囲(プロンプト、出力、ログ、学習、再委託先、監査)と、データを送る前に確認すべき 10 の質問をまとめます。 - [Guides in Japanese](https://broadrouter.com/ja/guides/) - [OpenAI Python SDK を BroadNet Inference で使う](https://broadrouter.com/ja/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): 公式の OpenAI Python SDK は、base_url を 1 か所変えるだけで BroadNet Inference に接続できます。インストール、キーの設定、最初のリクエスト、ストリーミングまでを説明します。 - [OpenAI Node.js SDK を BroadNet Inference で使う](https://broadrouter.com/ja/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): 公式の OpenAI Node.js / TypeScript SDK は、baseURL を設定するだけで BroadNet Inference に接続できます。インストール、最初のリクエスト、トークンのストリーミング、リクエスト ID の記録までを説明します。 - [BroadNet Inference をアカウントなしでブラウザから試す](https://broadrouter.com/ja/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): BroadNet のブラウザ用プレイグラウンドで、アラビア語の音声認識、書類の読み取り、Gemma 4 とのチャットを登録もカードもなしで試し、同じリクエストを API に移せます。 - [curl で BroadNet Inference に最初のリクエストを送る](https://broadrouter.com/ja/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): curl で BroadNet Inference を初めて呼び出す手順です。チャット補完、ストリーミング応答、音声の文字起こしを試し、レスポンス、usage、エラーの読み方を確認します。 - [chat completions のストリーミングと接続切断への対処](https://broadrouter.com/ja/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): BroadNet Inference のストリーミングの仕組みです。Server-Sent Events、[DONE] マーカー、最後のチャンクに入るトークン使用量、ストリームが途中で切れたときの対処を説明します。 - [推論モデルが空の回答を返す理由](https://broadrouter.com/ja/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash、GLM 5.3、Qwen3.8 27B は回答の前に推論し、その推論は max_tokens に含まれます。空の応答を防ぐ方法を説明します。 - [BroadNet Inference のエラーコードとリトライすべき場合](https://broadrouter.com/ja/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): BroadNet Inference API が返す HTTP ステータスの意味、指数バックオフでリトライすべきもの、先に修正すべきもの、サポートにリクエスト ID と一緒に送る情報をまとめました。 - [IP 許可リストで API キーを自社サーバーに限定する](https://broadrouter.com/ja/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): IP 許可リストで BroadNet Inference のキーを守り、推論用キーとポータル認証情報を分け、漏えいしたキーを数秒で失効させる方法です。curl の例付き。 - [エージェント向け無制限 LLM API プランのしくみ](https://broadrouter.com/ja/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): 無制限の DeepSeek V4.1 Flash プランにはトークンの計量も週の上限もありません。セッション、並列エージェント、キー、課金が実際にどう動くかを説明します。 - [Aider を無制限の DeepSeek V4.1 Flash で動かす](https://broadrouter.com/ja/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): コーディングエージェント Aider を OpenAI 互換の設定で BroadNet Inference に接続し、定額・無制限の DeepSeek V4.1 Flash プランで動かす手順です。 - [LangChain で BroadNet Inference を使う](https://broadrouter.com/ja/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): base_url と api_key で LangChain の ChatOpenAI を BroadNet Inference に接続する方法と、推論モデルの出力のうち LangChain が保持するもの・しないものを解説します。 - [Cline で BroadNet Inference を使う(OpenAI Compatible プロバイダー)](https://broadrouter.com/ja/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): VS Code 拡張機能 Cline を OpenAI Compatible プロバイダー経由で BroadNet Inference に接続します。入力する項目、モデル ID、コンテキストサイズ、よくある問題の対処法をまとめました。 - [Roo Code で BroadNet Inference を使う(OpenAI Compatible)](https://broadrouter.com/ja/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code は 2026 年 5 月に終了しました。既存のインストールやフォーク向けの設定と、移行先(Cline または Zoo Code フォーク)で BroadNet Inference を使う方法です。 - [Continue で BroadNet Inference を使う(config.yaml と apiBase)](https://broadrouter.com/ja/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): config.yaml で provider に openai、apiBase を指定して、BroadNet Inference のモデルを IDE 拡張機能 Continue に追加します。tool_use の capabilities とコンテキスト長も解説します。 - [opencode に BroadNet Inference をカスタムプロバイダーとして追加する](https://broadrouter.com/ja/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): opencode.jsonc で BroadNet Inference を opencode のカスタム OpenAI 互換プロバイダーとして追加し、モデルのエイリアス、機能、上限を設定する方法です。 - [Open WebUI を BroadNet Inference に接続する](https://broadrouter.com/ja/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Open WebUI に BroadNet Inference を OpenAI 互換の接続として追加し、モデル ID を許可リストに登録して、接続確認の失敗、空の応答、レート制限に対処する方法です。 - [LibreChat に BroadNet Inference をカスタムエンドポイントとして追加する](https://broadrouter.com/ja/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): librechat.yaml で BroadNet Inference を LibreChat の OpenAI 互換カスタムエンドポイントとして追加します。.env のキー、モデル一覧、タイトル用モデル、よくある問題の対処法も解説します。 - [n8n で BroadNet Inference を使う(OpenAI 認証情報と Base URL)](https://broadrouter.com/ja/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): n8n の OpenAI 認証情報の Base URL を BroadNet Inference に向け、OpenAI Chat Model ノードでモデルを ID 指定し、Responses API のトグルをオフにする手順です。 - [Vercel AI SDK で BroadNet Inference を使う](https://broadrouter.com/ja/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): @ai-sdk/openai-compatible の createOpenAICompatible を使い、Vercel AI SDK から BroadNet Inference を呼び出します。generateText、streamText、ツール、推論の強度を解説します。 - [アラビア語の音声認識 API を使う手順](https://broadrouter.com/ja/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): BroadNet Transcribe Arabic で、OpenAI 互換のエンドポイントからアラビア語音声を文字起こしします。対応形式、language フィールド、Python と curl の例、料金を説明します。 - [LiteLLM から BroadNet Inference にルーティングする(SDK とプロキシ)](https://broadrouter.com/ja/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): openai/ のモデルプレフィックスと api_base を使い、LiteLLM の Python SDK またはプロキシの config.yaml から BroadNet Inference を呼び出します。レート制限の合わせ方も解説します。 - [OpenAI Agents SDK を BroadNet Inference で動かす](https://broadrouter.com/ja/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Python 版 OpenAI Agents SDK を AsyncOpenAI と OpenAIChatCompletionsModel 経由で BroadNet Inference につなぎ、OpenAI のキーが必要なトレーシングをオフにする方法です。 - [WhatsApp のボイスメッセージを API で文字起こしする](https://broadrouter.com/ja/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): WhatsApp の .ogg ボイスメッセージやスマートフォンの .m4a 録音を ffmpeg で WAV に変換し、1 回の API 呼び出しでアラビア語ほか 13 言語に文字起こしします。 - [n8n で WhatsApp のボイスメッセージを文字起こしする](https://broadrouter.com/ja/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): WhatsApp のボイスメッセージを受け取り、Meta からダウンロードし、ffmpeg で変換して、アラビア語またはほかの 13 言語で文字起こしする n8n ワークフローです。 - [ビジョンモデルでアラビア語の請求書とレシートを読み取る](https://broadrouter.com/ja/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): 請求書やレシートの写真を BroadNet Inference のビジョンモデルに送り、アラビア語か英語で質問して、合計金額、日付、発行元を JSON で受け取る方法です。 - [Zed で BroadNet Inference を OpenAI 互換プロバイダーとして使う](https://broadrouter.com/ja/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Zed エディターの settings.json で BroadNet Inference を OpenAI 互換プロバイダーとして追加します。モデル ID、トークン上限、capabilities、キーの環境変数まで説明します。 - [Kilo Code で BroadNet Inference を使う(OpenAI Compatible プロバイダー)](https://broadrouter.com/ja/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Kilo Code を OpenAI Compatible のカスタムプロバイダーとして BroadNet Inference に接続します。VS Code のダイアログまたは kilo.json で、モデル ID、上限、ツールのフラグを設定します。 - [goose に BroadNet Inference をカスタムプロバイダーとして追加する](https://broadrouter.com/ja/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): AI エージェント goose を、OpenAI Compatible のカスタムプロバイダーとして BroadNet Inference に接続します。goose Desktop、goose configure、JSON ファイルのどれでも設定できます。 - [Pydantic AI を BroadNet Inference で使う(OpenAICompatibleProvider)](https://broadrouter.com/ja/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): OpenAIChatModel と OpenAICompatibleProvider で Pydantic AI のエージェントを BroadNet Inference 上で動かす方法です。ツール呼び出し、構造化出力、モデルプロファイルの調整も扱います。 - [CrewAI で BroadNet Inference をカスタム OpenAI エンドポイントとして使う](https://broadrouter.com/ja/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): LLM(custom_openai=True, base_url=...) で CrewAI のエージェントを BroadNet Inference に向け、ツール、長いコンテキスト、定額プランに合わせてモデルを選ぶ方法です。 - [LlamaIndex で OpenAILike を使って BroadNet Inference に接続する](https://broadrouter.com/ja/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): OpenAILike クラスで LlamaIndex を BroadNet Inference に接続し、api_base、context_window、is_chat_model、is_function_calling_model を正しく設定する方法です。 - [OpenAI-API-compatible プラグインで Dify に BroadNet Inference を追加する](https://broadrouter.com/ja/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): OpenAI-API-compatible モデルプロバイダープラグインで Dify を BroadNet Inference に接続します。各フィールドの正確な値、コンテキストサイズ、関数呼び出しの設定も解説します。 - [Jan で BroadNet Inference を OpenAI 互換のカスタムプロバイダーとして使う](https://broadrouter.com/ja/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): デスクトップアプリ Jan に BroadNet Inference を OpenAI 互換のカスタムプロバイダーとして追加し、モデル ID を手動で登録して、モデルごとにツールとビジョンをオンにします。 - [SillyTavern を BroadNet Inference に接続する(Custom OpenAI 互換)](https://broadrouter.com/ja/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): Chat Completion のソース Custom (OpenAI-compatible) で SillyTavern を BroadNet Inference に接続します。正しいモデル ID、プロンプトの後処理、トークン上限も解説します。 - [Spring AI で BroadNet Inference を使う(OpenAI base-url)](https://broadrouter.com/ja/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): spring.ai.openai.base-url、BroadNet のモデル ID、max-tokens を設定して Spring AI の OpenAI チャットスターターを BroadNet Inference に向け、ChatClient から呼び出します。 - [Go の openai-go SDK から BroadNet Inference を呼び出す](https://broadrouter.com/ja/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): 公式の openai-go v3 SDK を option.WithBaseURL で BroadNet Inference につなぎ、Go でチャット補完、使用量付きのストリーミング、エラー処理を行う方法です。 - [BroadNet Inference で LangGraph エージェントを作る](https://broadrouter.com/ja/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): ChatOpenAI、base_url、BroadNet のモデル ID で LangGraph の StateGraph や create_agent を BroadNet Inference 上で動かします。ツール、ストリーミング、メモリーも扱います。 - [Haystack のパイプラインとエージェントで BroadNet Inference を使う](https://broadrouter.com/ja/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): api_base_url、BroadNet のモデル ID、Secret のキーで Haystack の OpenAIChatGenerator を BroadNet Inference に向けます。パイプライン、ツール、ストリーミング、エージェントに対応します。 - [Semantic Kernel で BroadNet Inference を使う](https://broadrouter.com/ja/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Microsoft Semantic Kernel を BroadNet Inference に接続します。Python では AsyncOpenAI クライアント、C# ではカスタムエンドポイントを使い、チャット、プラグイン、エージェントに対応します。 - [Mastra のエージェントを BroadNet Inference で動かす](https://broadrouter.com/ja/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): id、url、apiKey を指定した model オブジェクトで Mastra のエージェントを BroadNet Inference 上で動かします。モデル ID に broadnet/ プレフィックスを二重に付ける理由も説明します。 - [BroadNet Inference で Instructor を使って構造化出力を得る](https://broadrouter.com/ja/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Instructor で BroadNet Inference から Pydantic オブジェクトを抽出します。from_openai または from_provider でカスタム base_url を指定し、適切な Mode を選びます。 - [DSPy で BroadNet Inference を使う(model_type="chat" を設定)](https://broadrouter.com/ja/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): openai/ プレフィックス、api_base、model_type="chat" で dspy.LM を BroadNet Inference 用に設定します。DSPy 3.5 は OpenAI のモデルを Responses API に送るためです。 - [promptfoo で BroadNet のモデルを評価する](https://broadrouter.com/ja/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): openai:chat プロバイダー、apiBaseUrl、apiKeyEnvar を使って、promptfoo の評価とモデルの並列比較を BroadNet Inference で実行し、採点モデルも設定する方法です。 - [OpenAIModel で smolagents を BroadNet Inference で動かす](https://broadrouter.com/ja/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): OpenAIModel と api_base で Hugging Face の smolagents を BroadNet Inference に接続します。CodeAgent はどのモデルでも、ToolCallingAgent は GLM 5.3 と Qwen3.8 27B で動きます。 - [Flowise を BroadNet Inference に接続する(OpenAI Custom Model)](https://broadrouter.com/ja/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): OpenAI Custom Model ノードで Flowise の chatflow から BroadNet Inference を使います。Base Path、OpenAI API 認証情報、手入力する BroadNet モデル名を解説します。 - [Langflow を BroadNet Inference に接続する](https://broadrouter.com/ja/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Langflow の OpenAI コンポーネントの OpenAI API Base とモデル名の手入力で BroadNet Inference を使う方法と、OpenAI Compatible プロバイダーでは一覧に出ない理由を説明します。 - [Glossary entries in Japanese](https://broadrouter.com/ja/glossary/) - [AI エージェント(AI agent)](https://broadrouter.com/ja/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): AI エージェントとは、言語モデルが目標に向けて自分で手順を決めるシステムです。どのツールを呼ぶかを決め、結果を読み、最終回答を出すか上限に達するまで繰り返します。決まった箇所でモデルを呼ぶ固定のスクリプトは、ふつうワークフローと呼ばれます。 - [API キー(API key)](https://broadrouter.com/ja/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): API キーとは、API を呼び出すクライアントを識別し、認可するための秘密の文字列です。LLM API では通常 Authorization: Bearer ヘッダーで送り、キーを持つ人は誰でも、そのキーの所有アカウントに課金されるリクエストを送れます。 - [アラビア語の発音記号(タシュキール、tashkeel)](https://broadrouter.com/ja/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): アラビア語の発音記号(タシュキール)とは、文字の上下に書く小さな記号で、短母音、子音の重複、母音がないことを示します。省略可能で、ニュースを含むほとんどのアラビア語テキストでは書かれないため、読み手もソフトウェアも文脈から母音を推測します。 - [アラビア語方言識別(Arabic dialect identification)](https://broadrouter.com/ja/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): アラビア語方言識別とは、テキストや録音がエジプト、湾岸、レバント、マグリブの各方言や現代標準アラビア語など、どのアラビア語の変種かを地域・国・都市の単位で判定するタスクです。メッセージの振り分け、プロンプト選択、評価セットの分割に使います。 - [アラビア語テキストの正規化(Arabic text normalization)](https://broadrouter.com/ja/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): アラビア語テキストの正規化とは、照合の前に同じ単語の表記ゆれを 1 つの形にそろえる処理です。発音記号とタトウィールを取り除き、アリフの形を統一し、多くの場合 ى を ي に、ة を ه に置き換えます。検索キーと評価に使い、表示するテキストには使いません。 - [アラビージー(Arabizi)](https://broadrouter.com/ja/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): アラビージー(Arabizi)とは、ラテン文字でくだけて書くアラビア語で、ラテン文字にない音は ع を 3 で表すように数字で書きます。チャットや SNS でよく使われ、ほとんどが方言で、決まった綴りがなく、英語やフランス語が混ざることもよくあります。 - [AWQ(Activation-aware Weight Quantization)](https://broadrouter.com/ja/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ(Activation-aware Weight Quantization)とは、LLM の重みを低い精度、通常は 4 ビット整数に圧縮する学習後量子化の手法です。小さなキャリブレーションセットの活性化の統計から最も重要な重みのチャネルを見つけ、精度が落ちにくいようにスケーリングします。誤差逆伝播は不要です。 - [ベース URL(base URL)](https://broadrouter.com/ja/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): ベース URL とは、https://api.example.com/v1 のような API の固定のルートアドレスで、クライアントが /chat/completions などのエンドポイントパスの前に付けます。OpenAI 互換の SDK では、API キーとモデル ID と一緒にベース URL を変えるだけで、同じコードを別のプロバイダーに向けられます。 - [バッチ API(Batch API)](https://broadrouter.com/ja/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): バッチ API とは、多数のリクエストを通常はファイルとして一度に受け付け、すぐにではなく決められた時間枠の中で処理し、結果をまとめて返す非同期の LLM エンドポイントです。OpenAI や Anthropic などの事業者は、バッチリクエストを標準 API より割り引いた価格にしています。 - [BM25](https://broadrouter.com/ja/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25(Okapi BM25)とは、キーワード検索のランキング関数で、クエリの語が文書に何回出現するか、その語がコレクション全体でどれだけ珍しいか、文書の長さはどれくらいかをもとに文書をスコアリングします。Lucene と Elasticsearch のデフォルトの類似度です。 - [バイトペアエンコーディング(BPE)](https://broadrouter.com/ja/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): バイトペアエンコーディング(BPE)とは、1 文字または 1 バイトから始め、最頻出の隣接ペアを新しい語彙項目として結合する操作を目標の語彙サイズまで繰り返すサブワード分割のアルゴリズムです。今の LLM のトークナイザーの多くは BPE かそのバイトレベル版です。 - [キャッシュ入力トークン(cached input tokens)](https://broadrouter.com/ja/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): キャッシュ入力トークンとは、LLM プロバイダーが直前に処理し、再計算せずにキャッシュから再利用できるプロンプトのトークンです。通常はシステムプロンプトやドキュメントなど繰り返し送るプレフィックスで、新しい入力トークンより安い単価で課金されます。 - [Chain-of-Thought(CoT、思考の連鎖)](https://broadrouter.com/ja/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Chain-of-Thought(CoT)とは、言語モデルに最終的な答えの前に途中の推論ステップを書かせる、あるいはその例を示すプロンプトの手法です。算数や論理のような多段階の問題で精度が上がり、今では推論モデルが自動的に生成します。 - [文字誤り率(CER、character error rate)](https://broadrouter.com/ja/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): 文字誤り率(CER)とは、書き起こしや OCR の精度を文字単位で測る指標で、出力を参照テキストに直すのに必要な置換・削除・挿入の数を、参照テキストの文字数で割ったものです。低いほど良く、0 は完全一致です。 - [Chat Completions API(チャット補完 API)](https://broadrouter.com/ja/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Chat Completions API とは、OpenAI の HTTP エンドポイント POST /v1/chat/completions で、モデル名とロール付きメッセージのリストを受け取り、モデルの次のメッセージを返します。そのリクエストと応答の形は、OpenAI 互換のプロバイダーや SDK が使う共通形式になりました。 - [チャットテンプレート(chat template)](https://broadrouter.com/ja/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): チャットテンプレートとは、ロール付きのメッセージのリストを、制御トークンも含めてモデルが想定する 1 本のトークン列に変換する書式ルールです。通常は Jinja で書かれ、トークナイザーと一緒に配布されます。誤ったテンプレートでは回答の質が大幅に落ちます。 - [チャンク分割(chunking)](https://broadrouter.com/ja/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): チャンク分割とは、検索パイプラインで埋め込みとインデックス化の前に文書をチャンクと呼ぶ小さな断片に分ける処理です。チャンクサイズは各断片の量を決め、オーバーラップは隣接チャンク間でテキストを重複させて境界で切れた内容の欠落を防ぎます。 - [コードスイッチング(code-switching)](https://broadrouter.com/ja/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): コードスイッチングとは、同じ会話や同じ文の中で、話し言葉でも書き言葉でも 2 つ以上の言語、あるいは 1 つの言語の複数の変種を使うことです。アラビア語話者が方言に英単語を混ぜたり、現代標準アラビア語と方言を行き来したりするのがその例です。 - [AI コーディングエージェント(AI coding agent)](https://broadrouter.com/ja/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): AI コーディングエージェントとは、言語モデルがループの中でコードベースに取り組むソフトウェアツールです。ファイルを読み、編集し、テストなどのコマンドを実行し、出力を読んで次の手順を決めることを、タスクが終わるか上限に達するまで繰り返します。 - [同時実行数(concurrency)](https://broadrouter.com/ja/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): 同時実行数(concurrency)とは、クライアントが同時に処理中にしているリクエストの数です。開いているリクエストは応答が終わるまで GPU の容量を占有するため、LLM API は毎分リクエスト数とは別に、キーやプランごとにこれを制限することがよくあります。 - [コンテキストの圧縮(context compaction)](https://broadrouter.com/ja/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): コンテキストの圧縮とは、コンテキストウィンドウの上限に近づいた会話を要約し、その要約からセッションを続ける手法です。エージェントは重要な決定や状態を保ちながら、古いツール出力などの生の履歴を捨てられます。 - [コンテキストエンジニアリング(context engineering)](https://broadrouter.com/ja/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): コンテキストエンジニアリングとは、システムプロンプト、ツール定義、履歴、検索した文書、メモなど、モデルが各リクエストで見るトークンの集合を選び、保守する作業です。タスクに必要な、最小限で情報価値の高い内容をモデルに渡すことを目指します。 - [コンテキストの劣化(context rot)](https://broadrouter.com/ja/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): コンテキストの劣化(context rot)とは、入力が長くなるにつれて言語モデルの精度と信頼性が徐々に下がる現象です。タスクが同じで、入力がコンテキストウィンドウに十分収まっていても起こります。18 モデルを検証した Chroma の 2025 年の技術レポートで広まった用語です。 - [コンテキストウィンドウ(context window)](https://broadrouter.com/ja/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): コンテキストウィンドウとは、言語モデルが 1 回のリクエストで扱えるトークン数の上限です。プロンプト、会話履歴、ツール定義、画像、推論トークンを含む出力がすべて入ります。収まらない部分は削るか要約するか、複数のリクエストに分けます。 - [データレジデンシー(data residency)](https://broadrouter.com/ja/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): データレジデンシーとは、データを特定の国や地域で保存・処理するという要件、またはその約束です。LLM API では、推論時にプロンプトがどこで処理されるか、保存されたコンテンツがどこにあるかが対象で、アカウントや請求のデータは対象外のことが多いです。 - [アラビア語の方言(dialectal Arabic)](https://broadrouter.com/ja/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): アラビア語の方言とは、エジプト、湾岸、レバント、マグリブなど、地域や都市で日常的に話されるアラビア語です。音、語彙、文法が現代標準アラビア語と異なり、標準的な綴りがなく、ネットで書かれることも増え、言語モデルには扱いにくい対象です。 - [話者ダイアライゼーション(speaker diarization)](https://broadrouter.com/ja/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): 話者ダイアライゼーションとは、録音を話者ごとの区間に分け、「誰がいつ話したか」に答えるタスクです。各発話区間に SPEAKER_00 のような匿名の話者 ID を付け、たいていは文字起こしと組み合わせて、各行を誰が話したかを示します。 - [知識蒸留(knowledge distillation)](https://broadrouter.com/ja/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): 蒸留(知識蒸留)とは、小さな生徒モデルが大きな教師モデルの出力(確率分布や生成した回答など)を再現するように学習する手法です。生徒モデルは、より低い実行コストで教師モデルの品質の多くを保てます。 - [データ処理契約(DPA)](https://broadrouter.com/ja/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): データ処理契約(DPA)とは、管理者と処理者の間で処理者による個人データの取り扱いを定める契約です。GDPR 第 28 条 3 項がこれを義務づけ、文書化された指示、守秘、セキュリティ、サブプロセッサーのルール、データの削除・返却、監査を内容に挙げています。 - [エンベディング(埋め込み、embeddings)](https://broadrouter.com/ja/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): エンベディング(埋め込み)とは、テキストや画像などの入力を表すためにモデルが出力する浮動小数点数のリスト(ベクトル)です。意味の近い入力ほどベクトルも近いので、その距離で検索結果の順位付け、文書の分類、関連項目の発見ができます。 - [評価(eval、LLM evaluation)](https://broadrouter.com/ja/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): eval(evaluation の略)とは、言語モデルやプロンプトを繰り返し検証できるテストです。固定した入力のセット、良い出力の条件、各出力を採点する採点者(コード、人間、別のモデル)で構成され、バージョン同士を比較できます。 - [指数バックオフ(exponential backoff)](https://broadrouter.com/ja/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): 指数バックオフ(exponential backoff)とは、失敗するたびに待ち時間を長くして再試行するリトライ戦略で、通常は上限まで遅延を 2 倍ずつ増やします。各待ち時間にランダムなジッターを加えると、多数のクライアントの再試行が分散し、同時にサーバーへ集中しなくなります。 - [Few-shot プロンプティング(few-shot prompting)](https://broadrouter.com/ja/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): Few-shot プロンプティングとは、実際の入力の前に、入力と出力の具体例を少数プロンプトに入れ、言語モデルにタスクと形式を推測させる手法です。モデルの重みは変わりません。例を使わないものを zero-shot、1 つだけ使うものを one-shot と呼びます。 - [ファインチューニング(fine-tuning)](https://broadrouter.com/ja/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): ファインチューニングとは、事前学習済みの言語モデルを、会話の例などタスク固有の小さなデータセットでさらに学習させることです。重みが変わるため、毎回のプロンプトに長い指示や例を入れなくても、そのタスクや文体をより安定してこなせます。 - [finish_reason(終了理由)](https://broadrouter.com/ja/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason とは、チャット補完の応答で各 choice に付くフィールドで、モデルが生成を止めた理由を示します。OpenAI 形式の値は stop、length、tool_calls、content_filter と、非推奨の function_call です。Anthropic の Messages API では同じ考え方を stop_reason で返します。 - [定額 LLM プラン(flat-rate LLM plan)](https://broadrouter.com/ja/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): 定額 LLM プランとは、言語モデル API の利用に対し、トークン単位ではなく月などの期間ごとに固定料金を課すプランです。通常は同時実行数、速度、使えるモデル、利用時間枠など別の何かが制限されるため、料金と同じくらい制限の内容が重要です。 - [FP8(8 ビット浮動小数点)](https://broadrouter.com/ja/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 とは、ディープラーニングで使われる 8 ビットの浮動小数点数形式です。指数 4 ビット・仮数 3 ビットの E4M3 と、指数 5 ビット・仮数 2 ビットの E5M2 の 2 種類があります。16 ビット形式に比べてメモリが半分になり、最近の GPU ではネイティブに動作します。 - [frequency penalty と presence penalty](https://broadrouter.com/ja/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): frequency penalty と presence penalty とは、出力にすでに現れたトークンのスコアを下げるサンプリング設定です。presence penalty はトークンが一度でも現れると適用され、frequency penalty は繰り返すたびに大きくなります。OpenAI の API ではどちらも -2.0〜2.0 で、デフォルトは 0 です。 - [GDPR(EU 一般データ保護規則)](https://broadrouter.com/ja/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): GDPR(一般データ保護規則、規則(EU)2016/679)とは、個人データの処理に関する EU の法律です。2018 年 5 月 25 日から適用され、EU に拠点を置く組織と一部の EU 域外の組織が対象です。LLM API では、処理者との契約、国際移転、データ最小化を規律します。 - [GGUF](https://broadrouter.com/ja/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF とは、llama.cpp など GGML ベースのランタイムで推論するための言語モデルのバイナリファイル形式です。1 ファイルに重み(多くは量子化済み)と、トークナイザーやアーキテクチャなど読み込みに必要なメタデータが入り、メモリマップで高速に読み込めます。 - [GPU 時間(GPU-hour)](https://broadrouter.com/ja/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): GPU 時間とは、GPU 1 基を 1 時間使うことです。クラウド事業者はレンタルした GPU インスタンスを時間単位や秒単位で課金し、研究者はジョブに使った計算量を GPU 時間で表します。稼働しているかどうかに関係なく、マシンの時間に料金がかかります。 - [グラウンディング(grounding)](https://broadrouter.com/ja/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): グラウンディングとは、言語モデルの出力を、提供した文書、検索結果、データベースのレコードなど、具体的で検証可能な情報源に結びつけ、それに基づいて答えるよう指示することです。根拠のある回答は情報源までたどれ、事実の捏造も起きにくくなります。 - [ガードレール(guardrails)](https://broadrouter.com/ja/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): ガードレールとは、言語モデルの呼び出しを決めた範囲に収めるため、アプリがその前後で実行するチェックです。入力、出力、ツール操作をそれぞれ検証し、チェックに通らなければリクエストをブロック、書き換え、またはフラグ付けします。 - [ハルシネーション(hallucination)](https://broadrouter.com/ja/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): ハルシネーションとは、言語モデルが出力する、もっともらしいが誤っている、作り話である、または入力に根拠がない内容です。架空の引用、誤った数字、存在しない API などがあり、事実を確認せずにありそうなテキストを予測することから生じます。 - [ハムザ(hamza)](https://broadrouter.com/ja/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): ハムザ(ء)とは、アラビア語で声門閉鎖音を表す文字です。単独か、alef(أ إ)、waw(ؤ)、yeh(ئ)を台にして書き、آ は alef と madda の組み合わせです。表記が揺れるため、検索や音声の採点では alef 系の字形を ا にまとめることがよくあります。 - [ハイブリッド検索(hybrid search)](https://broadrouter.com/ja/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): ハイブリッド検索とは、同じクエリに対して BM25 などのキーワード検索と、エンベディングによるベクトル検索の両方を実行し、2 つのランキングを 1 つに統合する検索手法です。統合には Reciprocal Rank Fusion(RRF)や正規化スコアの重み付き和がよく使われます。 - [冪等キー(idempotency key)](https://broadrouter.com/ja/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): 冪等キーとは、決済や購入など状態を変えるリクエストにクライアントが付ける一意の値で、通常は UUID などのランダムな ID です。同じキーで同じリクエストが再送されると、サーバーは処理を 2 回実行せず、最初の結果を返します。 - [画像トークン(image tokens)](https://broadrouter.com/ja/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): 画像トークンとは、ビジョン言語モデルがプロンプト内の画像を表す入力トークンです。数は画像サイズから計算され、通常はリサイズ後に固定サイズのパッチやタイルに分割します。テキストの入力トークンと同様に課金され、大きい画像ほど高くなります。 - [推論プロバイダー(inference provider)](https://broadrouter.com/ja/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): 推論プロバイダーとは、機械学習モデル(多くはオープンウェイトの LLM)を、所有、リース、レンタルした GPU で動かし、API 経由で利用を販売する企業です。料金は通常、トークン、リクエスト、時間単位です。サービングやスケーリング、ハードウェアも担います。 - [入力トークンと出力トークン(input and output tokens)](https://broadrouter.com/ja/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): 入力トークンとは、モデルがリクエストで読むトークンで、システムプロンプト、会話履歴、ツール定義、画像を含みます。出力トークンは、推論を含めモデルが返答として生成するトークンです。LLM API は通常この 2 つを別の単価で課金し、出力のほうが高額です。 - [INT4(4 ビット量子化)](https://broadrouter.com/ja/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 とは、16 段階を表せる 4 ビットの整数形式です。LLM の推論では通常、4 ビットの重みのみ量子化(W4A16)を指します。重みをスケール係数とともに INT4 で保存し、計算時に 16 ビットに戻すため、メモリは 16 ビットの重みの約 4 分の 1 で済みます。 - [IP 許可リスト(IP allow list)](https://broadrouter.com/ja/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): IP 許可リストとは、登録した IP アドレスや CIDR 範囲からのリクエストだけを受け付け、それ以外をすべて拒否するセキュリティ対策です。API に使うと、リストにないネットワークからはキーが使えないため、キー漏えいの被害を抑えられます。 - [ジェイルブレイク(LLM jailbreak)](https://broadrouter.com/ja/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): ジェイルブレイクとは、言語モデルに安全性の学習や導入先アプリのルールを無視させ、通常なら拒否する内容の生成や操作をさせるために作られた入力です。OWASP はジェイルブレイクをプロンプトインジェクションの一形態として扱っています。 - [JSON モード(JSON mode)](https://broadrouter.com/ja/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): JSON モード(JSON mode)とは、OpenAI 形式の API で response_format {"type": "json_object"} と指定する設定で、モデルの応答を有効でパース可能な JSON にします。JSON Schema を使う構造化出力と違い、どのキー、型、値が入るかは保証しません。 - [JSON Schema(JSON スキーマ)](https://broadrouter.com/ja/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema とは、JSON データの構造(型、必須フィールド、許可される値、入れ子)を記述・検証するための、JSON で書かれた語彙です。現行バージョンは 2020-12 です。LLM API はツールのパラメーターや、構造化出力が従うべき正確な形の定義に使います。 - [KV キャッシュ(KV cache)](https://broadrouter.com/ja/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): KV キャッシュとは、Transformer 言語モデルが処理済みの各トークンの attention のキーと値を保存するメモリです。新しいトークンはそれを再利用し、系列全体を再計算せずに済みます。コンテキスト長とともに大きくなり、GPU が同時に処理できるリクエスト数を制約しがちです。 - [レイテンシ(latency)](https://broadrouter.com/ja/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): LLM API のレイテンシとは、リクエストを送ってから最後の出力トークンを受け取るまでの時間です。最初のトークンまでの時間(TTFT)と生成時間の合計で、平均ではなく p50 や p95 などのパーセンタイルで示すのが一般的です。 - [LLM-as-a-Judge(LLM による評価)](https://broadrouter.com/ja/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-Judge とは、言語モデルが他のモデルの出力を採点する評価手法です。1 つの回答を評価基準で採点するか、2 つの回答から良いほうを選びます。人手の判断をはるかに安く近似できますが、先に出た回答や長い回答を好むなどの既知のバイアスがあります。 - [LLM ベンチマーク(LLM benchmark)](https://broadrouter.com/ja/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): LLM ベンチマークとは、知識問題、コーディング、音声認識など 1 つの能力で言語モデルを測り比べるための、固定された(通常は公開の)タスクと採点方法です。スコアはそのテストでの相対的な能力を示し、自社のワークロードでの実力は示しません。 - [LLM ルーター(LLM router)](https://broadrouter.com/ja/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): LLM ルーターとは、アプリと 1 つ以上の言語モデル API の間に入り、コスト、レイテンシ、可用性、タスクの種類などのルールに基づいて、リクエストごとにどのモデルやプロバイダーが処理するかを決めるソフトウェアまたはサービスです。 - [logprobs(対数確率)](https://broadrouter.com/ja/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): logprobs(対数確率)とは、言語モデルが生成したトークンに割り当てた確率の対数です。logprobs を返す API では、出力トークンごとにモデルがどれだけありそうだと判断したかと、必要に応じてその位置の有力な代替トークンがわかります。 - [LoRA(Low-Rank Adaptation)](https://broadrouter.com/ja/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA(低ランク適応)とは、事前学習済みモデルの重みを凍結し、選んだ層の小さな低ランク行列のペアだけを学習する、効率的なファインチューニング手法です。更新は小さなアダプターで保存でき、重みにマージすれば推論レイテンシも増えません。 - [Lost in the Middle(中間の情報の見落とし)](https://broadrouter.com/ja/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the Middle とは、言語モデルが長い入力の中で、関連情報が冒頭や末尾にあるときに最もよく活用し、中間にあるときに最も活用できないという傾向です。Liu らの 2023 年の研究で名付けられました。 - [max_tokens(最大出力トークン数)](https://broadrouter.com/ja/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens とは、モデルが 1 回の応答で生成できるトークン数の上限を決める LLM API のパラメーターです。生成が上限に達すると止まり、応答は途中で打ち切られたものとして、たとえば finish_reason が "length" になります。 - [MCP(Model Context Protocol)](https://broadrouter.com/ja/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP(Model Context Protocol)とは、LLM アプリと外部のツールやデータの接続方法を標準化した、JSON-RPC 2.0 ベースのオープンなプロトコルです。サーバーがツール、リソース、プロンプトを公開し、ホストアプリ内のクライアントがそれらを見つけて、モデルのツール呼び出しを渡します。 - [混合エキスパート(MoE)](https://broadrouter.com/ja/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): 混合エキスパート(MoE)とは、層の中に多数のエキスパート(サブネットワーク)を持ち、学習済みのルーターが各トークンを少数のエキスパートだけに送る設計です。MoE の言語モデルは総パラメータ数が大きくても、トークンごとに使うのは一部だけです。 - [モデルの非推奨化(model deprecation)](https://broadrouter.com/ja/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): モデルの非推奨化とは、API プロバイダーがモデルの提供終了を告知することで、通常は廃止(retirement や shutdown)の日付と推奨される後継モデルが示されます。その日を過ぎるとリクエストは失敗するため、アプリは事前に別のモデル ID へ移行する必要があります。 - [モデルフォールバック(model fallback)](https://broadrouter.com/ja/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): モデルフォールバックとは、失敗した LLM リクエストを順序付きリストの別のモデルやプロバイダーで再試行する手法です。きっかけはエラー、レート制限、コンテキスト長エラー、ポリシーによる拒否などで、クライアントか前段のゲートウェイで実行します。 - [モデルライセンス(model license)](https://broadrouter.com/ja/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): モデルライセンスとは、モデルの重み、コード、出力の利用・改変・共有の条件を定めた法的な取り決めです。オープンウェイトモデルは Apache 2.0 や MIT などの標準ライセンス、または大規模な商用サービスへの制限などの条件を加えうる独自ライセンスで公開されます。 - [現代標準アラビア語(MSA)](https://broadrouter.com/ja/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): 現代標準アラビア語(MSA)とは、アラブ世界全体で書き言葉、ニュース、教育、公式のスピーチに使われる標準化されたアラビア語です。誰の日常の母語でもなく、人々はエジプト方言や湾岸方言などの地域方言を話し、改まった場面で MSA に切り替えます。 - [マルチエージェントシステム(multi-agent system)](https://broadrouter.com/ja/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): マルチエージェントシステムとは、それぞれプロンプト、ツール、コンテキストを持つ複数のエージェントが 1 つのタスクを分担する LLM アプリケーションです。代表的なオーケストレーター・ワーカー型では、リードが計画し、並列のワーカーの結果を統合します。 - [マルチモーダル AI(multimodal AI)](https://broadrouter.com/ja/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): マルチモーダル AI とは、テキスト、画像、動画、音声など複数の種類のデータを扱うモデルやシステムです。写真と質問を一緒に読んだり、音声を受け取ってテキストで答えたりします。最も一般的なのは視覚言語モデルです。 - [固有表現抽出(NER、named entity recognition)](https://broadrouter.com/ja/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): 固有表現抽出(NER)とは、テキスト中の名前を見つけ、人物、組織、場所などの種類をそれぞれにラベル付けするタスクです。従来のシステムはすべてのトークンにタグを付けますが、LLM ではエンティティの一覧を JSON で返させ、原文と照合するのが一般的です。 - [OCR(光学文字認識)](https://broadrouter.com/ja/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): OCR(光学文字認識)とは、画像、スキャン、PDF 内の印刷文字や手書き文字を検出し、機械で読める文字に変換する技術です。従来の OCR は文字とページ上の位置を出力し、視覚言語モデルは文字を読んでその内容についての質問に直接答えられます。 - [Ogg Opus](https://broadrouter.com/ja/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus とは、Opus コーデック(RFC 6716)で符号化し、Ogg コンテナ(RFC 7845)に格納した音声です。低ビットレートでの音声や双方向の音声向けに設計されており、WhatsApp がボイスメッセージに使う形式で、通常は .ogg または .opus ファイルです。 - [オープンウェイトモデル(open-weight model)](https://broadrouter.com/ja/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): オープンウェイトモデルとは、学習済みの重みが公開され、誰でもダウンロードして実行できるモデルで、通常は条件付きのライセンスで提供されます。学習コードや学習データの詳細も公開し、他者が検証・再現できる完全なオープンソース AI とは別物です。 - [OpenAI 互換 API(OpenAI-compatible API)](https://broadrouter.com/ja/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): OpenAI 互換 API とは、POST /v1/chat/completions など OpenAI API と同じリクエスト形式とレスポンス形式を受け付ける HTTP API です。OpenAI の SDK やツールは、ベース URL、API キー、モデル名を変えるだけで利用できます。 - [並列ツール呼び出し(parallel tool calls)](https://broadrouter.com/ja/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): 並列ツール呼び出しとは、モデルが 1 回の応答で複数のツール呼び出し(たとえば 3 つのファイル読み込み)を返すことです。OpenAI 形式では tool_calls 内の複数のエントリとして届き、parallel_tool_calls を false にするとモデルは 1 ターンに 0 個か 1 個しか呼び出しません。 - [従量課金(pay-as-you-go)](https://broadrouter.com/ja/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): 従量課金とは、使った分だけを支払う料金体系です。トークン、リクエスト、音声の分数などの単位で計測され、サブスクリプションや最低利用額はありません。多くの LLM API はこの方式で、前払いの残高か定期的な請求書から支払います。 - [PDPL(サウジアラビアと UAE の個人データ保護法)](https://broadrouter.com/ja/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL(個人データ保護法)とは、勅令 M/19 で公布され SDAIA が執行するサウジアラビアの個人データ法と、UAE の 2021 年連邦政令法第 45 号の通称です。どちらも個人データの処理を規制し、外国の LLM API に送るプロンプトを含む国外移転に条件を定めています。 - [PII(個人を特定できる情報)](https://broadrouter.com/ja/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): PII(個人を特定できる情報)とは、氏名、ID 番号、生体情報など個人の身元を識別・追跡できる情報と、その人に結び付いた、または結び付けうる情報です。GDPR ではより広い「個人データ」を使います。LLM では、プロンプトが社外に出る前に PII をマスキングします。 - [プリフィルとデコード(prefill and decode)](https://broadrouter.com/ja/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): プリフィルとデコードは LLM 推論の 2 つのフェーズです。プリフィルではプロンプト全体を並列に処理して KV キャッシュを埋め、デコードではそのキャッシュを使って 1 トークンずつ生成します。前者は主に初回トークンまでの時間、後者は出力速度を決めます。 - [プリペイドクレジット(prepaid credit)](https://broadrouter.com/ja/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): プリペイドクレジットとは、使用前に API アカウントに入金しておくお金です。リクエストごとに費用が残高から差し引かれ、残高がなくなると、チャージするまで従量課金のリクエストは止まります。支出は支払った額までに抑えられます。 - [プロンプトエンジニアリング(prompt engineering)](https://broadrouter.com/ja/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): プロンプトエンジニアリングとは、言語モデルに与える指示を書き、構成し、テストして、要件を満たす出力を安定して得るためのプロセスです。ロールのメッセージ、明確な指示、例、書式、テストケースによる結果の測定を含みます。 - [プロンプトインジェクション(prompt injection)](https://broadrouter.com/ja/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): プロンプトインジェクションとは、ユーザーの入力や、Web ページ、メール、ファイル、ツールの結果に隠されたテキストが、開発者の意図に反して言語モデルの振る舞いを変える攻撃です。OWASP は LLM アプリケーションの Top 10 で 1 位(LLM01:2025)としています。 - [量子化(quantization)](https://broadrouter.com/ja/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): 量子化とは、ニューラルネットワークの数値(主に重み)を 16 ビットではなく 8 ビットや 4 ビットなどの低い精度で保存する処理です。モデルのメモリ使用量が減り、多くの場合は速く動きます。優れた手法では精度の低下を小さく抑えられます。 - [RAG(検索拡張生成)](https://broadrouter.com/ja/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG(検索拡張生成)とは、アプリケーションがまず検索エンジンやベクトルインデックスなどの外部ストアから質問に関連する文章を取り出し、それをプロンプトに加えて、言語モデルが記憶だけでなくそのテキストをもとに回答できるようにする手法です。 - [レート制限(rate limit)](https://broadrouter.com/ja/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): レート制限とは、1 分あたりのリクエスト数やトークン数など、一定時間内にクライアントが API を使える量に API 側が設ける上限です。上限を超えたリクエストは、使用量が下回るまで通常 HTTP 429 Too Many Requests で拒否されます。 - [ReAct(Reasoning and Acting)](https://broadrouter.com/ja/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct(Reasoning and Acting)とは、Yao らが 2022 年に提案したプロンプトのパターンで、言語モデルが推論の過程と、検索やツール呼び出しなどの行動を交互に行います。各観察結果を読んで計画を更新し、最終回答を出せるまで繰り返します。 - [推論トークン(reasoning tokens)](https://broadrouter.com/ja/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): 推論トークンとは、推論モデルが目に見える回答を書く前に、問題を考え抜くために生成するトークンです。通常は出力トークンとして課金され、出力の上限とコンテキストウィンドウに算入されます。API によっては非表示になったり、要約されたりします。 - [リクエストタイムアウト(request timeout)](https://broadrouter.com/ja/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): リクエストタイムアウト(request timeout)とは、HTTP クライアントがサーバーを待つ最長時間で、超えるとリクエストを諦めます。通常は接続タイムアウトと読み取りタイムアウトに分かれます。長い回答には時間がかかるため、LLM では読み取りタイムアウトを長めにします。 - [リランキング(reranking)](https://broadrouter.com/ja/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): リランキングとは検索の第 2 段階で、より正確だが遅いモデル(通常はクロスエンコーダー)が、一次検索で得たクエリと候補文書のペアをそれぞれ採点して並べ替え、最も関連性の高い少数だけを言語モデルに渡します。 - [RTL テキスト(右から左に書く文字、right-to-left)](https://broadrouter.com/ja/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): RTL(right-to-left)テキストとは、アラビア語やヘブライ語のように右から左へ表示される文字のことです。コンピューターは論理的な読み順で保存し、左から右の文字、数字、モデル名、コードと混在する場合は Unicode 双方向アルゴリズムが表示順を決めます。 - [サンプリングレート(sample rate)](https://broadrouter.com/ja/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): サンプリングレートとは、1 秒間に取り込む音声サンプルの数で、単位はヘルツ(Hz)です。音声認識モデルは一般に 16,000 Hz(16 kHz)のモノラル音声を前提とします。ビット深度、チャンネル数とあわせて、非圧縮音声の 1 分あたりのサイズを決めます。 - [seed パラメーター(シード値)](https://broadrouter.com/ja/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): seed パラメーターとは、リクエストに付ける整数で、同じリクエストの繰り返し時に API ができるだけ同じトークンを選ぶようにします。OpenAI はベストエフォートとしており、決定性は保証されません。バックエンド構成の変更は system_fingerprint でわかります。 - [サーバーレス推論(serverless inference)](https://broadrouter.com/ja/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): サーバーレス推論とは、エンドポイント背後の計算資源をプラットフォームが自動で起動・拡縮・停止し、アイドル時はゼロまで縮小することも多く、予約マシンでなく使用量で課金するモデルの実行方式です。コールドスタートの遅延が起こりうるのが難点です。 - [SLA(サービスレベルアグリーメント)](https://broadrouter.com/ja/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): SLA(サービスレベルアグリーメント)とは、プロバイダーが測定可能なサービス水準、通常は月間稼働率を契約で約束し、未達時の補償、通常はサービスクレジットを定めたものです。何をダウンタイムとみなすか、何が対象外か、どう請求するかも定めます。 - [投機的デコーディング(speculative decoding)](https://broadrouter.com/ja/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): 投機的デコーディングとは、小さく高速なドラフトモデルが次のトークンを複数提案し、大きなターゲットモデルが 1 回のフォワードパスでそれを検証して受け入れた分を残す推論手法です。出力の分布を変えずに生成を速くします。 - [音声認識(Speech-to-Text)](https://broadrouter.com/ja/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): 音声認識(Speech-to-Text)とは、話し言葉の音声を文字に変換する技術で、自動音声認識(ASR)とも呼ばれます。API は音声ファイルやストリームから文字起こしを返し、通常は音声 1 分単位で課金されます。精度は単語誤り率で測り、訛り、方言、録音品質で変わります。 - [支出上限(spend limit)](https://broadrouter.com/ja/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): 支出上限とは、API のアカウント、プロジェクト、キーが一定期間(通常は 1 か月)に使える金額の上限です。利用額が達すると、プロバイダーは期間がリセットされるか上限が引き上げられるまで、以降のリクエストを HTTP 429 や 400 などのエラーで拒否します。 - [SRT 字幕(SRT subtitles)](https://broadrouter.com/ja/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT(SubRip)とは、プレーンテキストの字幕形式です。各字幕は連番、hh:mm:ss,mmm --> hh:mm:ss,mmm と書く開始・終了時刻、1 行以上のテキスト、空行で構成されます。WebVTT は HTML の動画で使われる、よく似た Web 向けの形式です。 - [SSE(Server-Sent Events)](https://broadrouter.com/ja/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): SSE(Server-Sent Events)とは、content type が text/event-stream の長時間続く 1 つの HTTP 応答を通じて、サーバーがクライアントにテキストイベントのストリームを送るための Web 標準です。各イベントは data: などのフィールド行 1 行以上と、最後の空行からなります。 - [停止シーケンス(stop sequences)](https://broadrouter.com/ja/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): 停止シーケンス(stop sequences)とは、リクエストと一緒に送る文字列で、モデルがそのいずれかを生成した時点で生成を止めます。OpenAI ではフィールド名が stop で最大 4 つまで、Anthropic では stop_sequences です。一致した文字列は通常、返される出力に含まれません。 - [ストリーミング(streaming)](https://broadrouter.com/ja/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): ストリーミングとは、言語モデルの応答を最後にまとめて返すのではなく、生成中に小さなチャンクに分けて返す方式です。OpenAI 互換 API では、リクエストで stream を true にすると Server-Sent Events(SSE)でストリーミングされます。 - [構造化出力(Structured Outputs)](https://broadrouter.com/ja/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): 構造化出力とは、モデルの回答を開発者が指定した形式(通常は JSON Schema)に制約し、コードで確実に解析できるようにする LLM API の機能です。有効な JSON であることだけを保証し、特定の形までは保証しない JSON モードより厳格です。 - [サブエージェント(subagent)](https://broadrouter.com/ja/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): サブエージェントとは、メインのエージェントが 1 つの限定的なタスクのために起動する別のエージェントです。独自のコンテキストウィンドウ、システムプロンプト、ツール権限で動き、結果や要約だけをメインに返すので、メインの会話を小さく保てます。 - [サブプロセッサー(再委託先、subprocessor)](https://broadrouter.com/ja/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): サブプロセッサーとは、処理者が管理者に代わって個人データを処理するために委託する第三者です。LLM API では、GPU ホスト、提携する推論プロバイダー、バックアップ容量が含まれえます。GDPR 第 28 条により、管理者の承認と同等のデータ保護義務が必要です。 - [システムプロンプト(system prompt)](https://broadrouter.com/ja/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): システムプロンプトとは、アプリケーションが会話の前に言語モデルへ送る指示文で、通常は system または developer ロールのメッセージです。モデルの役割、ルール、口調、出力形式を決め、ユーザーのメッセージと矛盾する場合もモデルはこちらに従うことになっています。 - [ター・マルブータ(taa marbuta、ة)](https://broadrouter.com/ja/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): ター・マルブータ(ة、U+0629)とは、語末にだけ現れるアラビア文字で、女性名詞や形容詞を示すことが多い文字です。2 つの点が付いたハー(ه、U+0647)のように見えるため、入力されたテキストでは両者が混同され、多くの検索用正規化は ة を ه に変換します。 - [temperature(温度)](https://broadrouter.com/ja/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): temperature(温度)とは、言語モデルの出力のランダム性を決めるサンプリングのパラメーターです。トークンのスコアをこの値で割ってから確率にするため、低いと最有力のトークンが支配的になり、高いと分布が平らになって確率の低いトークンも選ばれやすくなります。 - [音声合成(TTS、text-to-speech)](https://broadrouter.com/ja/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): 音声合成(TTS、text-to-speech)とは、書かれたテキストを話し言葉の音声に変換する技術です。TTS API はテキスト、声、MP3 や WAV などの出力形式を受け取り、生成した音声を返します。音声をテキストにする音声認識とは逆方向の処理です。 - [スループット(throughput)](https://broadrouter.com/ja/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): スループット(throughput)とは、LLM の推論提供で、すべての同時リクエストを合わせてシステムが単位時間に完了する仕事の総量です。通常は出力トークン毎秒やリクエスト毎秒で測り、1 リクエストがトークンを受け取る速さ(ストリーム単位の速度)とは異なります。 - [TTFT(最初のトークンまでの時間)](https://broadrouter.com/ja/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): TTFT(最初のトークンまでの時間)とは、言語モデルにリクエストを送ってから最初の出力トークンが届くまでの時間です。ネットワーク、サーバーでの待ち行列、モデルがプロンプト全体を読むプリフィルを含むため、プロンプトの長さと負荷に応じて伸びます。 - [タイムスタンプ(音声認識、timestamps)](https://broadrouter.com/ja/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): 音声認識のタイムスタンプとは、文字起こしの各部分に付く開始・終了時刻(秒)です。セグメント単位は句や文を、単語単位は各単語を対象にします。字幕、録音内の検索、テキストと話者の対応付けに必要です。 - [トークン(token)](https://broadrouter.com/ja/glossary/token/) (translation of https://broadrouter.com/glossary/token/): トークンとは、言語モデルが読み書きするテキストの単位です。モデルのトークナイザーが作る単語、単語の一部、句読点、バイト列のいずれかにあたります。LLM API は利用量をトークンで数え、上限を設け、通常はトークン単位で課金します。 - [トークナイザー(tokenizer)](https://broadrouter.com/ja/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): トークナイザーとは、テキストを言語モデルが読む整数のトークン ID の列に変換し、生成された ID をテキストに戻す部品です。モデルごとに独自のトークナイザーがあるため、同じテキストでもモデルによってトークン数が変わります。 - [トークン毎秒(TPS、tokens per second)](https://broadrouter.com/ja/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): トークン毎秒(TPS)とは、言語モデルが出力トークンを生成する速さで、通常は 1 リクエストの最初のトークン以降で測ります。応答が流れる速さを決め、全同時リクエストの合計がシステムのスループットです。最初のトークンまでの待ち時間は別に測ります。 - [ツール呼び出し(Function Calling、tool calling)](https://broadrouter.com/ja/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): ツール呼び出し(Function Calling とも呼ばれます)とは、開発者が定義した関数の実行を求める構造化されたリクエストを、モデルが JSON の引数付きで返す LLM API の機能です。関数を実行して結果を返すのはアプリケーションで、モデル自身がコードを実行することはありません。 - [tool_choice(ツール選択パラメーター)](https://broadrouter.com/ja/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice とは、ツールの使用を制御するリクエストパラメーターです。OpenAI の Chat Completions では auto(モデルが判断)、required(1 つ以上のツールを必ず呼ぶ)、none(ツールなし)、または強制する関数名を指定します。Anthropic では auto、any、tool、none が対応します。 - [top-p(Nucleus サンプリング)](https://broadrouter.com/ja/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): top-p(Nucleus サンプリング)とは、確率の高い順に合計が閾値 p(0.9 など)に達する最小のトークン集合からだけ次のトークンを選び、可能性の低い裾を切り捨てるデコード手法です。 - [Unicode 正規化(Unicode normalization)](https://broadrouter.com/ja/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Unicode 正規化とは、見た目が同じ文字列が同じコードポイントを使うように、テキストを標準形に変換することです。NFD、NFC、NFKD、NFKC の 4 形式があり、K の付く形式はアラビア文字の表示形や合字などの互換文字も通常の文字にまとめます。 - [音声区間検出(VAD)](https://broadrouter.com/ja/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): 音声区間検出(VAD)とは、音声を短いフレームに分け、各フレームが発話か非発話かを判定する信号処理のステップです。音声システムは無音のスキップ、長い録音の間での分割、話し終わりの検出、背景雑音の文字起こし防止に使います。 - [ベクトルデータベース(vector database)](https://broadrouter.com/ja/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): ベクトルデータベースとは、エンベディングのベクトルを多くの場合元のテキストやメタデータとともに保存し、最近傍検索に答えるデータストアです。クエリベクトルに最も似た保存済みベクトルを、主にコサイン距離、ユークリッド距離、内積で返します。 - [視覚言語モデル(VLM、vision-language model)](https://broadrouter.com/ja/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): 視覚言語モデル(VLM)とは、画像とテキストを一緒に入力として受け取り、テキストを出力するモデルです。画像の説明、画像についての質問への回答、文書・グラフ・スクリーンショットの読み取りができます。image-text-to-text モデルやマルチモーダルモデルとも呼ばれます。 - [vLLM](https://broadrouter.com/ja/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM とは、LLM の推論とサービングのためのオープンソースライブラリで、UC バークレーの Sky Computing Lab で開発が始まりました。PagedAttention で KV キャッシュのメモリを管理し、リクエストを連続的にバッチ処理し、多くの量子化形式に対応し、OpenAI 互換の HTTP サーバーを提供します。 - [語彙サイズ(vocabulary size)](https://broadrouter.com/ja/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): 語彙サイズとは、トークナイザーにある異なるトークンの数で、各トークンが固有の ID とエンベディングを持ちます。大きいほど一般的なテキストを少ないトークンで表せますが、埋め込み層と出力層が大きくなり、効果は語彙の学習に使ったテキストに限られます。 - [WAV](https://broadrouter.com/ja/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV(Waveform Audio File Format、WAVE とも)とは、Microsoft と IBM の RIFF 形式をベースにした音声コンテナです。通常は非圧縮のリニア PCM サンプルを格納するため、サイズは予測できます。1 秒あたりのバイト数は、サンプリングレート × サンプルあたりのバイト数 × チャンネル数です。 - [単語誤り率(WER、word error rate)](https://broadrouter.com/ja/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): 単語誤り率(WER)とは、音声認識の標準的な精度指標です。書き起こしを正しい参照テキストに直すのに必要な単語の置換・削除・挿入の数を、参照テキストの単語数で割って求めます。値が低いほど良好です。 - [ゼロデータ保持(ZDR)](https://broadrouter.com/ja/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): ゼロデータ保持(ZDR)とは、AI プロバイダーがプロンプトと出力をリクエストの処理のためだけに扱い、その後はログ、キャッシュ、学習データのいずれにも保存しないデータ取り扱い方針です。トークン数や処理時間など、内容を含まない記録は通常残ります。 ### Korean (ko, 한국어): 294 translated pages - [Articles in Korean](https://broadrouter.com/ko/blog/) - [2026년 10월 LLM API 가격 변동: 인상 3건, 인하 없음](https://broadrouter.com/ko/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 2026년 10월 LLM API 제공업체 7곳 중 5곳의 가격 페이지에서 변경 23건: 인상 3건, 인하 없음, 신규 가격 9개 모델, 삭제 11개 모델. 건마다 아카이브 링크를 제공합니다. - [AI 코딩 에이전트 1회 실행 비용: 단계 수, 컨텍스트 증가, 캐싱](https://broadrouter.com/ko/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): 코딩 에이전트는 실행이 길어질수록 단계당 비용이 커지는 이유, 캐싱이 총액을 바꾸는 방식, 정액 세션이 토큰당 과금보다 유리해지는 시점을 토큰 계산으로 설명합니다. - [LLM으로 아랍어 발음 부호 넣기: 프롬프트, 검증, DER](https://broadrouter.com/ko/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): LLM API로 아랍어 발음 부호(타쉬킬)를 복원하는 방법입니다. 글자를 그대로 두는 프롬프트, 수정을 걸러 내는 검증, DER·WER 채점, CATT를 쓸 때를 정리했습니다. - [LLM으로 아랍어 문법·맞춤법 교정하기: 실전 가이드](https://broadrouter.com/ko/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): LLM API로 아랍어 맞춤법과 문법을 교정하는 방법입니다. 최소 수정 프롬프트, 다시 쓰기 방지, 오류 유형별 단어 단위 수정, M2 방식 채점까지 다룹니다. - [아랍어 회의 녹음으로 회의록 만들기: API 파이프라인](https://broadrouter.com/ko/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): 아랍어 회의 녹음을 결정 사항과 실행 항목이 담긴 회의록으로: 전사하고, 정리하고, LLM으로 JSON 요약을 만든 뒤 이름과 날짜를 확인합니다. - [LLM으로 아랍어 개체명 추출하기: JSON, 날짜, 검증](https://broadrouter.com/ko/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): LLM으로 아랍어 인명, 기관, 지명, 히즈라력 날짜를 JSON으로 추출하기: 프롬프트, 발음 부호를 무시하는 스팬 검사, 날짜 변환, 채점 스크립트를 담았습니다. - [아랍어 OCR API 비교: Google, Azure, Textract, Mistral, LLM](https://broadrouter.com/ko/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): 아랍어 인쇄체와 손글씨를 공식 지원하는 OCR API는 어디인지, 각각 무엇을 돌려주고 1,000페이지당 얼마인지, 비전 LLM API와 비교했습니다. - [음성 인식 API로 아랍어 자막(SRT) 만들기: 필요한 것](https://broadrouter.com/ko/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): SRT 파일에 필요한 것, 아랍어에 타임스탬프나 SRT를 반환하는 음성 인식 API, 그리고 텍스트 전용 API로 자막을 만드는 검증된 파이프라인을 정리했습니다. - [LLM으로 아랍어 텍스트 분류하기: 감성 분석, 의도, 티켓 분류](https://broadrouter.com/ko/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): LLM API로 아랍어 리뷰, 메시지, 지원 티켓 분류하기: 라벨 설계, 방언을 고려한 프롬프트, JSON 출력, 매크로 F1 테스트, 대량 처리 비용까지 정리했습니다. - [검색과 RAG를 위한 아랍어 텍스트 정규화: 바꿀 것과 둘 것](https://broadrouter.com/ko/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): 타쉬킬, 알리프 형태, 타 마르부타, 함자 받침, 페르시아 문자, 숫자, 유니코드까지. 검색에 도움이 되는 아랍어 정규화 규칙과 해가 되는 규칙, 테스트한 파이썬 코드를 정리했습니다. - [LLM으로 아랍어 텍스트 요약하기: 프롬프트, 길이 조절, 평가](https://broadrouter.com/ko/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): LLM API로 아랍어 기사, 보고서, 방언 메시지 요약하기: 프롬프트 패턴, 단어 수 조절, 숫자 검증, 그리고 기본 ROUGE가 아랍어에 0점을 주는 이유까지 정리했습니다. - [Artificial Analysis 대안: 2026년 LLM 비교 사이트 정리](https://broadrouter.com/ko/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Artificial Analysis 대안을 비교합니다. Arena, Epoch AI, Vals AI, OpenRouter 순위, Hugging Face, 가격 데이터를 각각 무엇을 측정하고 언제 쓰는지 기준으로 정리했습니다. - [BroadNet의 GPU 하드웨어와 서빙 방식: BroadNet이 밝힌 내용](https://broadrouter.com/ko/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): BroadNet Inference를 받치는 하드웨어에 대해 BroadNet이 밝힌 내용입니다. B300부터 고객사 어플라이언스까지 3개 GPU 계층, GPU별 처리량, 테스트에서 얻은 교훈을 정리했습니다. - [LLM API 응답 캐싱으로 비용 줄이기: Python 정확 일치 캐시](https://broadrouter.com/ko/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): 테스트를 거친 Python용 LLM API 응답 정확 일치 캐시. 키에 넣을 것, 절대 캐시하면 안 되는 것, TTL, 개인정보, 구축 전 적중률 측정법을 다룹니다. - [오픈 모델로 Claude Code 쓰기: 프록시, 네이티브 엔드포인트, 한계](https://broadrouter.com/ko/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code는 Anthropic Messages API를 씁니다. LiteLLM 프록시나 네이티브 엔드포인트로 오픈 모델에서 돌리는 방법과, 그 과정에서 깨지는 부분을 정리합니다. - [문서 AI 워크플로에 사람의 검토 단계 넣기](https://broadrouter.com/ko/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): 문서 AI 파이프라인의 사람 검토 단계를 설계하는 법: 검증 규칙, 사람에게 보낼 문서, 무작위 감사 표본, 감사 추적, 검토율까지 정리했습니다. - [이집트 개인정보보호법(2020년 법률 제151호)과 LLM API](https://broadrouter.com/ko/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): 이집트인의 개인정보를 해외 LLM이나 음성 인식 API로 보낼 때, 2020년 법률 제151호와 2025년 결정 제816호가 이전, 허가, 유출에 대해 정한 내용을 정리했습니다. - [파인튜닝 전에: 해결할 문제부터 정의하세요](https://broadrouter.com/ko/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): 프롬프트, RAG, 평가로 문제가 해결되어 파인튜닝이 필요 없는 경우, 필요한 학습 데이터, 그리고 파인튜닝을 요청하기 전에 시험하는 방법을 정리했습니다. - [LLM이 잘못된 JSON을 반환할 때 해결법: 원인, 복구, 재시도](https://broadrouter.com/ko/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): LLM이 깨진 JSON(코드 펜스, 설명 문장, max_tokens 잘림, 후행 쉼표)을 반환하는 이유와, 안전한 것만 복구하고 나머지는 거부하는 검증된 Python 함수를 소개합니다. - [LLM API 가격은 얼마나 자주 바뀔까? 3년치 가격 페이지 아카이브 분석](https://broadrouter.com/ko/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): 2023~2026년 LLM API 제공업체 11곳의 보관된 가격 페이지 308개를 읽었습니다. 모델의 80%는 가격이 한 번도 바뀌지 않았고, 정가 변경의 75%는 DeepInfra였습니다. - [LLM API 비용 줄이는 방법: 순서대로 따라 하는 10단계 체크리스트](https://broadrouter.com/ko/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): LLM API 요금은 순서가 중요합니다. 측정, 프롬프트 접두부 캐싱, 프롬프트 다듬기, 출력·추론 상한, 모델 선택, 배치, 그리고 정액 플랜이 유리해지는 시점까지 정리했습니다. - [LLM API 페일오버: 직접 작성한 코드로 두 번째 제공업체로 전환하기](https://broadrouter.com/ko/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): OpenAI 호환 LLM 제공업체 간 페일오버 방법. 어떤 오류에서 전환할지, 서킷 브레이커, 모델 ID 매핑, 스트리밍, 비용과 개인정보 조건 맞추기까지 정리했습니다. - [LLM API 가격 인하 2023~2026: 가장 큰 인하 사례와 출처](https://broadrouter.com/ko/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): 2023~2026년 가장 큰 LLM API 가격 인하를 Internet Archive 변경 전·후 링크와 함께 정리했습니다. o3, Codestral, Gemini 1.5, 그리고 잘 알려지지 않은 인상까지. - [LLM API 가격은 내리기만 할까? 가격 인상 52건과 예산 세우는 법](https://broadrouter.com/ko/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): LLM API 가격도 오릅니다. 2024~2026년 보관된 가격 페이지에서 같은 모델의 정가 인상 52건(중앙값 40%)을 찾았습니다. 모든 출처와 예산 규칙을 정리합니다. - [제공업체별 LLM API 속도 제한 비교: RPM, TPM, 동시 요청](https://broadrouter.com/ko/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks 등 13곳의 첫 단계 LLM API 속도 제한을 각사 문서에서 확인해 비교하고, RPM과 TPM의 뜻도 설명합니다. - [LLM API SLA 비교: Azure, Bedrock, Google, OpenAI는 무엇을 보장하나](https://broadrouter.com/ko/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): LLM API 제공업체가 서면으로 약정하는 내용을 SLA 페이지에서 직접 확인했습니다. 가동률, 다운타임 정의, 서비스 크레딧, 지연 시간 목표를 비교합니다. - [Batch API 없이 LLM 일괄 처리하기: 재개 가능한 실행기](https://broadrouter.com/ko/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Batch API 없이 채팅 완성 API로 프롬프트 1만~100만 건을 처리하는 법: 동시 요청 상한, 속도 조절, 재시도, 재개 가능한 체크포인트, 소요 시간 계산. - [Safety Map: LLM 39개는 민감한 데이터를 어떻게 다루나](https://broadrouter.com/ko/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet이 14개 연구소의 LLM 39개를 자격 증명, 개인정보, 건강·금융 데이터로 조사했습니다. 전이된 것과 안 된 것, 도구 호출과 추론을 통한 유출을 정리했습니다. - [LLM의 민감 데이터 안전성 테스트 방법: BroadNet의 방법론](https://broadrouter.com/ko/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): 배포 전에 LLM이 자격 증명, 개인정보, 건강·금융 데이터를 유출하는지 테스트하는 BroadNet의 공개 방법론입니다. 범주, 유출과 누락의 구분, 3회 이상 실행을 다룹니다. - [OpenAI API에서 옮겨 가기: Python, Node, cURL 코드로 보는 전환 과정](https://broadrouter.com/ko/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): OpenAI에서 OpenAI 호환 API로 옮길 때 바꿔야 할 코드: 클라이언트 설정, Responses에서 Chat Completions로, 도구, 스트리밍, 임베딩, 스모크 테스트까지 정리했습니다. - [BroadNet이 평가 후 탈락시킨 모델과 그 이유](https://broadrouter.com/ko/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet은 테스트 후 탈락시킨 모델과 이유를 공개합니다. 자격 증명 안전성 검사에서 떨어진 채팅 모델, 경쟁에서 밀린 음성 인식, OCR, 임베딩 모델을 정리했습니다. - [다국어 AI 서비스 설계: 아랍어·영어·스페인어 UX 가이드](https://broadrouter.com/ko/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): 아랍어, 영어, 스페인어 사용자를 위한 AI 제품 설계: UI 로캘과 모델 언어 분리, RTL과 혼합 방향 텍스트, 숫자·날짜 형식, 언어 감지, 언어별 비용. - [오픈 웨이트 vs 클로즈드 모델 API 비용: 가격 데이터로 본 차이](https://broadrouter.com/ko/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): 오픈 웨이트 모델을 API로 쓰면 GPT, Claude, Gemini보다 쌀까요? 각 제공업체 페이지의 토큰당 가격, 호스트별 가격 차이, 사용자 1명당 비용 예시로 비교합니다. - [OpenAI vs Anthropic vs Google API 가격 변천사, 2023~2026](https://broadrouter.com/ko/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): 보관된 가격 페이지로 2023~2026년 GPT, Claude, Gemini API 가격을 나란히 비교합니다. 어느 라인이 싸지고 어느 라인이 비싸졌는지, 아카이브 링크와 함께 정리했습니다. - [OpenAI 호환 API base_url 목록: 제공업체 14곳 직접 확인](https://broadrouter.com/ko/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): OpenAI 호환 LLM API 14곳의 base_url을 각 제공업체 공식 문서에서 직접 확인했습니다. 지원 엔드포인트와, 그대로 교체할 때 문제가 되는 차이점도 정리했습니다. - [저렴한 모델로 OpenClaw와 Hermes Agent 쓰기: 설정과 토큰 계산](https://broadrouter.com/ko/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): OpenClaw나 Hermes Agent를 OpenAI 호환 엔드포인트에 연결하고, 예약 점검이 토큰을 태우는 이유와 가벼운 컨텍스트, 저렴한 모델로 비용을 줄이는 법을 알아봅니다. - [전용 엔드포인트는 분명한 워크로드 정의에서 시작합니다](https://broadrouter.com/ko/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): LLM 전용 엔드포인트를 문의하기 전에 정해야 할 것: 워크로드, 초당 토큰, 동시성, 컨텍스트, 리전, 라이선스, 예산, 그리고 받은 엔드포인트를 테스트하는 방법. - [코딩 에이전트의 프롬프트 인젝션: 실제로 효과 있는 방어책](https://broadrouter.com/ko/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): 프롬프트 인젝션이 코딩 에이전트에 들어오는 경로와, 업체와 연구자가 문서화한 다층 방어(샌드박스, 외부 통신 제한, 허용 목록, 사람의 승인)를 정리했습니다. - [RAG 청크 크기와 오버랩: 검색 품질과 비용에 미치는 영향](https://broadrouter.com/ko/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): 청크 크기와 오버랩이 RAG 재현율, 인덱스 크기, 임베딩 비용, 그리고 매 쿼리가 LLM에 보내는 토큰 수를 어떻게 바꾸는지, 공개된 실험 결과와 계산 예시로 설명합니다. - [추론 모델의 max_tokens 설정: 토큰 예산을 정하는 방법](https://broadrouter.com/ko/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): 추론 모델이 낮은 max_tokens에서 빈 응답을 반환하는 이유와, 측정한 백분위수, 컨텍스트 윈도우, 예산을 늘려 가는 재시도로 적정 예산을 정하는 방법을 설명합니다. - [브라우저에서 녹음한 오디오를 음성 인식(STT) API로 보내는 방법](https://broadrouter.com/ko/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): MediaRecorder로 마이크를 녹음하고, WebM이나 Safari의 MP4를 16kHz WAV로 변환한 뒤, API 키를 보관하는 작은 서버를 거쳐 전사 API로 보내는 방법입니다. - [프롬프트 토큰 줄이기: 시스템 프롬프트, 도구, 데이터 절감 실측](https://broadrouter.com/ko/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): 시스템 프롬프트, 퓨샷 예시, 도구 스키마, 데이터 형식, 대화 이력, 아랍어 발음 부호를 줄였을 때 토큰이 얼마나 줄어드는지 실제 토크나이저로 측정했습니다. - [같은 모델, 다른 호스트: LLM API 가격은 얼마나 다를까](https://broadrouter.com/ko/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): 여러 호스트가 판매하는 모델 119개에서 가장 비싼 정가는 가장 싼 정가의 중앙값 1.76배입니다. 2026년 11월, 제공업체 41곳의 자체 가격 페이지 데이터입니다. - [사우디·UAE 개인정보보호법(PDPL)과 LLM API: 국외 이전 체크리스트](https://broadrouter.com/ko/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): 사우디아라비아나 UAE에서 개인정보를 LLM API로 보낼 때 PDPL 국외 이전 규정의 내용, 제공업체에 물어볼 질문, 국내 추론이 필요한 경우를 정리했습니다. - [텍스트, 음성, 비전 중 무엇을 쓸까요? 작업부터 정하세요](https://broadrouter.com/ko/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): 텍스트 채팅, 이미지 입력, 오디오 전사를 작업에 따라 고르는 법을 BroadNet이 문서화한 모델과 한도, 그리고 TTS처럼 제공하지 않는 기능과 함께 정리했습니다. - [25MB 넘는 긴 오디오 전사하기: 분할, 전송, 이어 붙이기](https://broadrouter.com/ko/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): 음성 인식 API의 파일 크기 제한을 넘는 녹음을 전사하는 방법: 파일을 줄이고, ffmpeg로 무음 구간에서 분할하고, 청크를 병렬로 보낸 뒤 텍스트를 이어 붙입니다. - [LLM으로 앱 문자열 아랍어 번역하기: 플레이스홀더와 복수형](https://broadrouter.com/ko/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): LLM API로 i18n JSON 파일을 아랍어로 번역하는 방법입니다. 플레이스홀더 고정, 아랍어 복수형 6개 생성, CI에서 ICU 문법 검증, 삽입된 이름 격리까지 다룹니다. - [API로 아랍어 음성을 영어 텍스트로 번역하기: 테스트한 파이프라인](https://broadrouter.com/ko/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): API 호출 두 번으로 아랍어 녹음을 영어 텍스트로: language=ar로 전사하고 LLM으로 번역합니다. 숫자 검사, 분할, 비용, 채점까지 함께 다룹니다. - [라이브, 문의 시 제공, 연구용: 모델 카탈로그 제대로 읽는 법](https://broadrouter.com/ko/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): BroadNet 카탈로그와 /v1/models를 예로, 지금 바로 호출할 수 있는 모델, 문의가 필요한 모델, 테스트만 한 모델을 구분하는 방법을 알아봅니다. - [vLLM의 Uno: BroadNet 서빙 연구가 측정한 것](https://broadrouter.com/ko/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet이 Uno 자기 초안 어댑터를 vLLM에 통합하고 H100에서 측정했습니다. 스트림 1개에서 일반 처리량의 2.58배, 32개에서 1.74배. 방법, DOI, 한계를 정리했습니다. - [Gemma 4 26B에 적용한 Uno: BroadNet의 전문가 혼합(MoE) 연구](https://broadrouter.com/ko/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet이 전문가 혼합 모델 Gemma 4 26B A4B용 Uno 자기 초안 어댑터를 처음으로 학습했습니다. 측정 내용, 자기 초안이 유리한 경우, DOI를 정리했습니다. - [에이전트 토큰 예산: 에이전트 루프의 토큰은 어디로 가는가](https://broadrouter.com/ko/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): 시스템 프롬프트, 도구 스키마, 파일 읽기, 대화 기록, 추론. 에이전트 루프가 토큰을 쓰는 방식, 사용량 데이터로 부분별 측정하는 법, 먼저 줄일 곳을 정리합니다. - [아랍어 음성 인식 단어 오류율(WER) 쉽게 이해하기](https://broadrouter.com/ko/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): 단어 오류율의 의미와 계산법, 그리고 아랍어 WER이 발음 부호, 철자 정규화, 방언에 특히 민감한 이유를 파이썬 코드와 실제 수치로 설명합니다. - [아랍어 방언 음성 인식: 걸프, 이집트, 레반트, 마그레브](https://broadrouter.com/ko/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): 걸프, 이집트, 레반트, 마그레브 아랍어에서 음성 인식이 약한 이유와, 내 오디오에서 제공업체 성능을 보여 주는 방언별 테스트 세트 만드는 법을 정리했습니다. - [LLM으로 아랍어–영어 번역하기: 프롬프트, 용어집, 평가](https://broadrouter.com/ko/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): LLM API로 아랍어를 영어로, 영어를 아랍어로 번역하는 방법입니다. 어조를 지키는 프롬프트, 고정 용어집, 숫자와 이름 처리, 품질 평가 방법을 정리했습니다. - [걸프, 이집트, 레반트 아랍어: LLM은 방언을 어떻게 처리할까](https://broadrouter.com/ko/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): LLM이 걸프, 이집트, 레반트 아랍어를 어떻게 처리하는지 정리했습니다. 실패하는 지점, 방언 프롬프트 작성법, 테스트 방법, 그리고 BroadNet이 Qwen3.8 27B에 대해 밝힌 결과까지. - [아랍어 OCR vs 비전 LLM: 아랍어 문서와 청구서 읽기](https://broadrouter.com/ko/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): 아랍어 청구서, 영수증, 양식에 기존 OCR과 비전 LLM을 언제 쓸지 정리했습니다. 정확도 함정, JSON 출력, 페이지당 비용, 검증 파이프라인까지 다룹니다. - [아랍어 프롬프트 엔지니어링: 문체, RTL, 숫자, 혼용 텍스트](https://broadrouter.com/ko/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): LLM에서 좋은 아랍어를 얻는 프롬프트 작성법: 문체와 방언 지정, 숫자와 날짜 통제, 깔끔한 RTL 서식, 아랍어와 영어 혼용 처리까지 정리했습니다. - [아랍어 RAG: 제대로 작동하는 정규화, 청킹, 검색](https://broadrouter.com/ko/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): 아랍어 문서용 검색 증강 생성(RAG): 알리프, 야, 타 마르부타 정규화, 적절한 청킹, 다국어 임베딩 선택, 키워드 검색 결합까지 정리했습니다. - [아랍어 음성 인식(STT) API: 2026년 선택 가이드](https://broadrouter.com/ko/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): 아랍어 전사 API 고르는 법: 단어 오류율, 방언, 파일 한도, 오디오 시간당 가격, 개인정보 보호, 그리고 내 녹음으로 하는 테스트 계획까지 정리했습니다. - [아랍어 WhatsApp 챗봇 아키텍처: 음성 메시지에서 답장까지](https://broadrouter.com/ko/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): 실제로 동작하는 아랍어 WhatsApp 어시스턴트 아키텍처: 웹훅, 큐, 음성 메시지 전사, LLM 답장, 정책 규칙과 비용까지, 바로 응용할 수 있는 파이썬 코드와 함께 정리했습니다. - [아라비지와 LLM: 3, 7 같은 숫자를 쓰는 프랑코 아랍어 처리하기](https://broadrouter.com/ko/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): 아라비지는 ع를 3으로, ح를 7로 쓰는 등 숫자를 섞어 아랍어를 라틴 문자로 표기합니다. 감지하는 법, LLM으로 아랍 문자로 바꾸는 법, 올바르게 답하는 법을 정리했습니다. - [Aider에 가장 좋은 LLM: 모델 고르는 법과 연결 방법](https://broadrouter.com/ko/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Aider에 맞는 LLM을 고르는 법. 편집 형식 준수, 컨텍스트 크기, 시간당 비용, 내 저장소로 1시간 테스트하기, 그리고 정확한 OPENAI_API_BASE 설정을 정리합니다. - [코딩 에이전트에 가장 좋은 오픈 모델: 도구, 컨텍스트, 비용으로 고르기](https://broadrouter.com/ko/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): 코딩 에이전트에는 GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B, Gemma 4 중 무엇이 맞을까요? 도구 호출, 컨텍스트, 추론, 작업당 비용으로 계산과 함께 고릅니다. - [캐시된 입력 토큰: 프롬프트 캐싱으로 LLM API 요금 줄이기](https://broadrouter.com/ko/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): 캐시된 입력은 일반 입력 가격의 일부만 과금됩니다. 프롬프트 캐싱의 과금 방식, usage의 cached_tokens 읽는 법, 캐시가 적중하도록 프롬프트를 구성하는 방법을 정리했습니다. - [가장 저렴한 비전 LLM API: 이미지 1장, 문서 1페이지당 비용 계산](https://broadrouter.com/ko/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): 비전 LLM 호출 비용을 이미지와 문서 페이지 단위로 계산합니다. Gemma 4, Qwen3.8 27B, DeepSeek V4.1 Flash의 토큰 계산과 DeepInfra가 더 싼 경우를 정리했습니다. - [LLM API 제공업체 선택 가이드: 20가지 체크리스트](https://broadrouter.com/ko/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): LLM API를 고르기 전에 물어야 할 20가지 질문입니다. 가격, 한도, OpenAI 호환성, 개인정보, 모델, 지원과 가동률, 그리고 점수표와 테스트 계획까지 정리했습니다. - [아랍어용 LLM API 선택: 무엇을 테스트해야 하나](https://broadrouter.com/ko/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): 아랍어 앱용 LLM 고르는 법: 방언 이해, 글쓰기 품질, 번역, 아랍어 문서와 이미지, 비용, 그리고 하루 만에 만들 수 있는 테스트 세트까지 정리했습니다. - [오픈 모델을 쓰는 팀을 위한 Cline vs Roo Code vs Aider 비교(2026)](https://broadrouter.com/ko/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code(아카이브 후 Zoo Code로 이어짐), Aider를 오픈 모델에서 중요한 기준으로 비교합니다. 도구 호출, 컨텍스트 사용량, MCP, git, 승인 방식. - [LLM 요청 전에 토큰 수 계산하는 방법](https://broadrouter.com/ko/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): LLM API를 호출하기 전에 언어와 모델별로 프롬프트 토큰 수를 추정하는 방법. 토크나이저마다 결과가 다른 이유, 5개 언어 실측 예시, 안전한 max_tokens 여유분을 정리했습니다. - [코딩 에이전트 LLM 비용 줄이는 10가지 방법과 예상 절감액](https://broadrouter.com/ko/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Cline, Aider, opencode 또는 직접 만든 코딩 에이전트의 토큰 비용을 줄이는 구체적인 방법 10가지. 절감액 계산식과 측정 방법을 함께 정리합니다. - [DeepSeek API 대안: DeepSeek 모델을 호출하는 5가지 방법](https://broadrouter.com/ko/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): DeepSeek 모델을 공식 API, 애그리게이터, 클라우드 카탈로그, 자체 GPU, BroadNet Inference로 호출할 때의 가격, 개인정보, 한도를 비교합니다. - [DeepSeek V4.1 Flash API 가격: 토큰당, 캐시된 입력, 정액 플랜](https://broadrouter.com/ko/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): DeepSeek V4.1 Flash API의 100만 토큰당 입력·캐시된 입력·출력 가격, 월간·주간·시간제 정액 플랜, DeepSeek·DeepInfra·Fireworks와의 비교를 정리했습니다. - [기능 출시 전에 LLM API 비용을 예상하는 방법](https://broadrouter.com/ko/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): 새 AI 기능의 비용 모델을 단계별로 만듭니다. 사용자 × 작업 × 호출 × 토큰에 캐시, 추론, 재시도, 속도 제한을 반영하고 월 비용 계산 예시 표로 정리합니다. - [제공업체를 바꾸기 전에 LLM을 평가하는 방법](https://broadrouter.com/ko/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): 실제 트래픽으로 작은 평가 세트를 만들고, 코드 검사와 편향을 줄인 LLM 심사로 점수를 매긴 뒤, 마이그레이션 전에 품질, 비용, 지연 시간을 비교하는 방법입니다. - [비전 LLM으로 송장·영수증·양식 데이터를 JSON으로 추출하기](https://broadrouter.com/ko/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): 송장, 영수증, 양식을 위한 프로덕션 파이프라인입니다. 이미지 준비, JSON 스키마 프롬프트, Pydantic 검증, 합계 검사, 재시도, 사람 검토 큐를 다룹니다. - [LLM API 정액제 vs 토큰당 과금: 손익분기점 계산](https://broadrouter.com/ko/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): LLM 정액 플랜이 토큰당 과금보다 싸지는 시점은 언제일까요? DeepSeek V4.1 Flash 월 청구액 예시 4가지, 손익분기 사용량, 직접 측정하는 방법을 정리합니다. - [무료 LLM API 크레딧: 초기 크레딧으로 실제로 얼마나 쓸 수 있나](https://broadrouter.com/ko/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): 무료 LLM API 크레딧으로 실제로 무엇을 할 수 있을까요? BroadNet 초기 크레딧의 토큰, 오디오 분, 시간제 키 환산과 Gemini, OpenRouter 무료 티어 비교. - [GDPR과 LLM API: 제공업체에 물어야 할 15가지 질문](https://broadrouter.com/ko/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): LLM API를 위한 GDPR 체크리스트입니다. DPA, 하위 프로세서, 보관, 학습 사용, 국외 이전, 침해 통지를 질문별 근거가 되는 GDPR 조항과 함께 정리했습니다. - [GLM 5.3 vs DeepSeek V4.1 Flash: 언제 어떤 모델을 써야 할까](https://broadrouter.com/ko/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3과 DeepSeek V4.1 Flash를 가격, 컨텍스트, 도구 호출, 구조화된 출력, 이미지, 추론 기준으로 비교하고 작업별 비용 예시와 간단한 선택 기준을 정리했습니다. - [LLM API 429 오류 처리 방법: 백오프, 지터, 동시 처리 제한](https://broadrouter.com/ko/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429와 5xx의 차이, 전체 지터를 적용한 지수 백오프, LLM API용 동시 처리 제한과 큐를 프로덕션에 바로 쓸 수 있는 Python·Node 코드와 함께 설명합니다. - [시간제 LLM API 키: 시간당 정액 요금 완벽 정리](https://broadrouter.com/ko/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): BroadNet 시간제 키는 정해진 가격에 Gemma 4 VL 26B를 스트림 하나로 1시간 쓰는 상품입니다. 포함 내용, 한도, 적합한 용도, 이득이 되는 시점을 정리했습니다. - [LLM API 데이터 보관 정책 비교: OpenAI, Anthropic, Gemini, Mistral](https://broadrouter.com/ko/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): OpenAI, Anthropic, Gemini API, Mistral, BroadNet이 API 프롬프트 보관과 학습 사용에 대해 공식적으로 밝힌 내용을 2026년 11월 각사 페이지에서 그대로 인용해 비교합니다. - [LLM API 키 보안: 보관, 교체, IP 허용 목록](https://broadrouter.com/ko/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): LLM API 키를 코드와 로그에서 떼어 놓고, 워크로드마다 키를 따로 쓰고, 주기적으로 교체하고, 서버 IP로 잠그고, 유출된 키를 몇 초 안에 폐기하는 방법을 설명합니다. - [LLM API 가격 구조: 입력, 캐시, 출력, 추론 토큰 요금 정리](https://broadrouter.com/ko/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): LLM API 토큰당 가격은 어떻게 매겨질까요? 입력, 캐시된 입력, 출력, 추론 토큰, 이미지, 분당 오디오 요금을 공개 요금표 기준의 계산 예시로 설명합니다. - [LLM API에 민감한 데이터 보내기: 가리고, 줄이고, 덜 기록하기](https://broadrouter.com/ko/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): 고객, 건강, 금융 데이터를 LLM API에 안전하게 보내는 법: 필드 분류, 되돌릴 수 있는 플레이스홀더로 마스킹, 데이터 비보관, 깨끗한 로그. - [LLM API 비용 통제: 예산, 키 한도, 무한 루프 방지](https://broadrouter.com/ko/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): LLM API 청구액 폭주를 막는 방법. 선불 잔액, 앱별 키, IP 허용 목록, 사용량 기록, 코드 내 예산, 루프 상한, 정액 키까지 계층별로 정리했습니다. - [LLM API 사용량 로깅: 사용자·기능별 비용 추적과 개인정보 마스킹](https://broadrouter.com/ko/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): 모든 LLM API 호출을 토큰, 비용, 지연 시간, 요청 ID와 함께 기록하고, 사용자와 기능별로 비용을 배분하며, 기록 전에 개인정보를 마스킹하는 방법을 Python 코드로 설명합니다. - [사용자당 LLM 비용: SaaS AI 기능의 단위 경제성](https://broadrouter.com/ko/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): AI 기능이 활성 사용자 1명당 월 얼마인지 계산하고, 평균보다 헤비 유저가 더 중요한 이유와 모델 비용을 감당하는 SaaS 요금제 설계법을 정리합니다. - [오픈 모델 API 가격 비교: 추론 제공업체 6곳의 과금 방식](https://broadrouter.com/ko/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): OpenRouter, Together AI, Fireworks, Groq, DeepInfra, BroadNet이 오픈 모델에 매기는 가격을 비교합니다. 토큰당, 캐시, 배치, GPU 시간, 정액 플랜을 출처와 함께 정리합니다. - [LLM 지연 시간: TTFT, 초당 토큰 수, 엔드투엔드 시간의 차이](https://broadrouter.com/ko/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): 첫 토큰까지의 시간(TTFT), 초당 토큰 수, 엔드투엔드 지연 시간이 각각 무엇을 재는지, 클라이언트에서 코드로 측정하는 방법, 어떤 개선이 어떤 수치를 움직이는지 설명합니다. - [LLM 스트리밍 원리: SSE 청크, 델타, usage, 취소](https://broadrouter.com/ko/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): 채팅 완성 스트리밍이 실제로 어떻게 전송되는지, 델타로 답변을 재구성하고 마지막 청크에서 usage를 읽고 깔끔하게 취소하는 방법을 Python, Node, curl로 설명합니다. - [롱 컨텍스트 vs RAG: 100만 토큰이 벡터 DB를 대체할 때](https://broadrouter.com/ko/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): 100만 토큰 컨텍스트 윈도우가 벡터 데이터베이스를 대체할 수 있는 경우와 여전히 RAG가 나은 경우를 쿼리당 비용, 지연 시간, 재현율, 유지 관리 측면에서 비교합니다. - [1M 컨텍스트 창으로 장시간 실행 에이전트의 컨텍스트 관리하기](https://broadrouter.com/ko/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): 100만 토큰 창으로 긴 에이전트 세션을 정확하고 저렴하게 유지하는 법: 추가 전용 기록, 캐시 친화적 프롬프트, 압축 시점, 먼저 잘라낼 것. - [장시간 실행 에이전트: 버티는 타임아웃, 재시도, 체크포인트](https://broadrouter.com/ko/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): LLM 에이전트를 몇 시간씩 안정적으로 돌리는 법: 호출별 타임아웃, 멈춤을 감지하는 스트리밍, 안전한 재시도, 멱등적인 도구 부수 효과, 단계별 체크포인트. - [OpenAI 호환 모델에서 MCP 서버 사용하기(모든 제공업체)](https://broadrouter.com/ko/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP는 어떤 LLM과도 동작합니다. 클라이언트가 MCP 도구를 함수 호출로 바꿔 주기 때문입니다. Cline, Continue, opencode, n8n, 직접 작성한 Python 코드에서 오픈 모델로 MCP를 쓰는 방법을 정리했습니다. - [100만 토큰 컨텍스트 윈도우, 언제 쓸 만하고 언제 아닌가](https://broadrouter.com/ko/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): 1,048,576토큰에 무엇이 들어가는지, DeepSeek V4.1 Flash에서 비용이 얼마인지, 추론과 캐싱이 계산을 어떻게 바꾸는지, 그리고 언제 여전히 RAG가 나은지 정리했습니다. - [멀티 에이전트 구성의 실제 비용: 토큰당 과금 vs 정액 세션](https://broadrouter.com/ko/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): 멀티 에이전트 시스템은 에이전트 하나보다 몇 배의 토큰을 씁니다. 정액 플랜과 토큰당 과금의 세션 계산, 숨은 배수 요인, 각각이 유리한 경우를 정리합니다. - [다국어 음성 인식(STT) API: 14개 언어를 엔드포인트 하나로](https://broadrouter.com/ko/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): 영어, 스페인어, 프랑스어, 중국어, 한국어, 아랍어 등 14개 언어를 분당 요금의 OpenAI 호환 엔드포인트 하나로 전사하는 방법과 아랍어 전용 모델을 쓸 때를 정리했습니다. - [저렴한 LLM으로 n8n AI Agent 쓰기: 설정, 모델 선택, 실행당 비용](https://broadrouter.com/ko/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): 저렴한 OpenAI 호환 모델로 n8n AI Agent 워크플로를 만드는 법. 도구용 모델 선택, 토큰을 줄이는 설정, 실행 10,000회당 비용 계산을 정리합니다. - [기업용 오픈 웨이트 vs 비공개 AI 모델: 비용, 통제, 위험](https://broadrouter.com/ko/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): 기업에는 오픈 웨이트 LLM과 비공개 LLM 중 무엇이 맞을까요? 비용, 통제, 개인정보, 품질 격차, 종속성을 비교하고 오픈 모델을 실행하는 세 가지 방법을 정리했습니다. - [OpenAI 호환 API로 전환하기: 마이그레이션 체크리스트](https://broadrouter.com/ko/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): base_url 변경은 쉬운 부분입니다. 문제는 모델 ID, 추론 토큰, 스트리밍 사용량, 오류, 속도 제한, 키에서 생깁니다. 코드와 함께 보는 체크리스트입니다. - [AI 에이전트에는 OpenRouter와 정액 플랜 중 무엇이 더 저렴할까?](https://broadrouter.com/ko/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter는 토큰당 과금에 크레딧 구매 수수료가 붙고, 정액 플랜은 세션 단위로 과금합니다. 상시 동작 에이전트의 비용 계산, OpenRouter 수수료와 한도, 어느 쪽이 맞는지 정리했습니다. - [코딩 에이전트를 밤새 무인으로 안전하게 돌리는 방법](https://broadrouter.com/ko/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): 코딩 에이전트를 무인으로 돌리기 위한 실전 구성: 샌드박스, git 체크포인트, 작업 파일, 단계 제한, 아침 리뷰, 밤새 쓰는 정액 플랜 세션까지 정리했습니다. - [접두부가 캐시되도록 프롬프트를 구성하는 방법](https://broadrouter.com/ko/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): 캐시되는 접두부가 매번 똑같도록 LLM 프롬프트를 배치하는 방법. 시스템 프롬프트, 도구 스키마, 문서, 덧붙이기만 하는 이력, 그리고 이를 증명하는 테스트까지 다룹니다. - [추론 토큰: 보이는 것보다 많은 출력 요금을 내는 이유](https://broadrouter.com/ko/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): 추론 모델은 숨겨진 생각을 출력 토큰으로 과금합니다. 측정 방법, 모델별 비용, reasoning_effort, max_tokens, 라우팅으로 줄이는 방법을 설명합니다. - [정액제 LLM API로 코딩 에이전트를 병렬 실행하는 방법](https://broadrouter.com/ko/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): 무제한 DeepSeek V4.1 Flash에서 서브에이전트, 오케스트레이터, 배치 워커 돌리기. 필요한 플랜 세션 수, 키별로 나누는 법, 동시 요청을 제한하는 법을 정리합니다. - [LLM 자체 호스팅 vs API: 실제 비용과 손익분기점](https://broadrouter.com/ko/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): 공식 페이지 기준 GPU 임대 가격, 내 GPU에서 100만 토큰당 비용을 구하는 공식, API보다 싸지려면 필요한 가동률, 흔히 빠뜨리는 운영 비용을 정리합니다. - [음성 인식(STT) API 가격 비교: 2026년 분당 요금 총정리](https://broadrouter.com/ko/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq, BroadNet Transcribe의 분당·1,000시간당 STT 가격을 공식 페이지 기준으로 비교하고 부가 기능 요금도 정리했습니다. - [JSON 모드 vs json_schema: 믿고 쓸 수 있는 구조화된 출력](https://broadrouter.com/ko/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON 모드는 유효한 JSON을, json_schema는 원하는 구조를 보장합니다. OpenAI 호환 모델에서 각각 요청하는 법, Pydantic 검증, 잘못된 출력을 루프로 고치는 법을 정리했습니다. - [아랍어와 영어의 단어당 토큰 수: 아랍어가 더 비싼 이유](https://broadrouter.com/ko/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): 같은 뜻이라도 아랍어가 영어보다 LLM 토큰을 더 쓰는 이유, 단어당 토큰 수가 오해를 부르는 이유, usage 필드로 내 비율을 측정하는 스크립트를 정리했습니다. - [오픈 모델 도구 호출: 스키마, 병렬 호출, 재시도](https://broadrouter.com/ko/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): OpenAI 호환 오픈 모델의 함수 호출: 스키마 작성 팁, 병렬 도구 호출, 도구 호출이 실패하는 다섯 가지 경우, 검증하고 재시도하는 Python 루프를 정리했습니다. - [콜센터 녹음 전사하는 법: 아랍어까지 지원하는 파이프라인](https://broadrouter.com/ko/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): 스테레오 통화를 채널별로 나누고, 발화 구간을 자르고, 아랍어와 영어를 전사하고, 개인정보를 가리고, LLM으로 통화를 요약합니다. 비용까지 담은 Python 파이프라인입니다. - [코딩 에이전트에서 오픈 모델 쓰기: Aider, Cline, opencode, Codex](https://broadrouter.com/ko/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Aider, Cline, Roo Code, opencode, Continue, Codex CLI를 OpenAI 호환 엔드포인트에 연결하는 법. 도구별 정확한 필드 이름과 자주 문제가 되는 설정을 정리합니다. - [비즈니스용 음성 메시지 텍스트 변환: WhatsApp·Telegram에서 티켓까지](https://broadrouter.com/ko/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): WhatsApp Business와 Telegram 봇으로 받은 고객 음성 메시지를 전사 텍스트와 지원 티켓으로 바꾸는 방법입니다. 다운로드, 전사, LLM 필드 추출, 라우팅을 다룹니다. - [단기 프로젝트용 주간 LLM 플랜: Orbit, 시간제, 월간 비교](https://broadrouter.com/ko/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): 1주일 무제한 DeepSeek V4.1 Flash 플랜이 시간제 세션, 월간 플랜, 토큰당 과금보다 유리한 경우를 해커톤, 스프린트, 마이그레이션 사례의 계산과 함께 정리했습니다. - [LLM API 플랜의 "무제한"은 실제로 무엇을 뜻할까](https://broadrouter.com/ko/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): 토큰 계량, 5시간 롤링 윈도우, 주간 한도, 속도 제한, 동시 처리. "무제한" AI 플랜이 내 에이전트에게도 정말 무제한인지 가려내는 7가지 질문을 정리했습니다. - [Whisper API 대안: 언제 바꿔야 하고 어떻게 테스트할까](https://broadrouter.com/ko/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): OpenAI Whisper API 대신 다른 호스팅 API를 쓸 때를 정리했습니다. 내 오디오의 단어 오류율, 지원 언어, 시간당 가격, 한도, 개인정보 보호를 테스트 코드와 함께 확인하세요. - [데이터 비보관(ZDR) LLM API: 확인해야 할 질문 10가지](https://broadrouter.com/ko/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): LLM API의 "데이터 비보관"이 프롬프트, 출력, 로그, 학습, 하위 처리자, 감사에서 무엇을 뜻해야 하는지 정리하고, 데이터를 보내기 전에 물어볼 질문 10가지를 소개합니다. - [Guides in Korean](https://broadrouter.com/ko/guides/) - [BroadNet Inference에서 OpenAI Python SDK 사용하기](https://broadrouter.com/ko/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): 공식 OpenAI Python SDK에서 base_url 하나만 바꿔 BroadNet Inference에 연결합니다. 설치, 키 설정, 첫 요청, 응답 스트리밍까지 순서대로 안내합니다. - [BroadNet Inference에서 OpenAI Node.js SDK 사용하기](https://broadrouter.com/ko/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): 공식 OpenAI Node.js·TypeScript SDK의 baseURL을 BroadNet Inference로 지정합니다. 설치, 첫 요청, 토큰 스트리밍, 요청 ID 로깅까지 안내합니다. - [계정 없이 브라우저에서 BroadNet Inference 체험하기](https://broadrouter.com/ko/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): BroadNet 브라우저 플레이그라운드에서 가입이나 카드 없이 아랍어 음성 인식, 문서 읽기, Gemma 4 채팅을 테스트하고, 같은 요청을 API로 옮기는 방법입니다. - [curl로 BroadNet Inference에 첫 요청 보내기](https://broadrouter.com/ko/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): curl로 BroadNet Inference를 처음 호출해 봅니다. 채팅 완성, 스트리밍 응답, 오디오 전사를 실행하고 응답, usage, 오류를 읽는 법까지 정리했습니다. - [채팅 완성 스트리밍과 끊긴 연결 처리하기](https://broadrouter.com/ko/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): BroadNet Inference의 스트리밍 응답을 정리했습니다. Server-Sent Events 형식, [DONE] 마커, 마지막 청크의 토큰 사용량, 스트림이 중간에 끊겼을 때의 대처법을 다룹니다. - [추론 모델이 빈 답변을 반환하는 이유](https://broadrouter.com/ko/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B는 답하기 전에 추론하며, 그 추론도 max_tokens에 포함됩니다. 빈 응답이 오지 않게 하는 방법을 정리했습니다. - [BroadNet Inference 오류 코드와 재시도해야 하는 경우](https://broadrouter.com/ko/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): BroadNet Inference API의 HTTP 상태 코드별 의미, 백오프로 재시도할 오류와 먼저 고쳐야 할 오류, 요청 ID와 함께 지원팀에 보낼 정보를 정리했습니다. - [IP 허용 목록으로 API 키를 내 서버에만 묶기](https://broadrouter.com/ko/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): IP 허용 목록으로 BroadNet Inference 키를 보호하고, 추론용 키와 포털 자격 증명을 분리하고, 유출된 키를 몇 초 만에 폐기하는 방법을 curl 예제와 함께 설명합니다. - [에이전트용 무제한 LLM API 플랜은 어떻게 작동하나요?](https://broadrouter.com/ko/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): 무제한 DeepSeek V4.1 Flash 플랜에는 토큰 계량도 주간 한도도 없습니다. 세션, 병렬 에이전트, 키, 결제가 실제로 어떻게 작동하는지 정리했습니다. - [Aider를 무제한 DeepSeek V4.1 Flash로 실행하기](https://broadrouter.com/ko/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Aider 코딩 에이전트를 OpenAI 호환 설정으로 BroadNet Inference에 연결하고, 정액 요금의 무제한 DeepSeek V4.1 Flash 플랜으로 실행하는 방법을 설명합니다. - [BroadNet Inference에서 LangChain 사용하기](https://broadrouter.com/ko/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): LangChain의 ChatOpenAI를 base_url과 api_key로 BroadNet Inference에 연결하고, 추론 모델의 응답 중 LangChain이 유지하는 것과 버리는 것을 알아봅니다. - [BroadNet Inference에서 Cline 사용하기(OpenAI Compatible 제공자)](https://broadrouter.com/ko/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): VS Code 확장 Cline을 OpenAI Compatible 제공자로 BroadNet Inference에 연결합니다. 입력할 필드, 모델 ID, 컨텍스트 크기, 자주 겪는 문제의 해결법을 정리했습니다. - [BroadNet Inference에서 Roo Code 사용하기(OpenAI Compatible)](https://broadrouter.com/ko/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code는 2026년 5월에 서비스가 종료되었습니다. 기존 설치와 포크용 설정, 그리고 BroadNet Inference와 함께 옮겨 갈 곳(Cline 또는 Zoo Code 포크)을 정리했습니다. - [Continue에서 BroadNet Inference 쓰기(config.yaml과 apiBase)](https://broadrouter.com/ko/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): Continue IDE 확장의 config.yaml에 provider openai와 apiBase로 BroadNet Inference 모델을 추가합니다. tool_use capabilities와 컨텍스트 길이 설정도 다룹니다. - [opencode에 BroadNet Inference를 커스텀 제공자로 추가하기](https://broadrouter.com/ko/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): opencode.jsonc에서 BroadNet Inference를 opencode 코딩 에이전트의 커스텀 OpenAI 호환 제공자로 추가하고, 모델 별칭과 capabilities, limit을 설정합니다. - [Open WebUI를 BroadNet Inference에 연결하기](https://broadrouter.com/ko/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): BroadNet Inference를 Open WebUI에 OpenAI 호환 연결로 추가하고 모델 ID를 허용 목록에 넣는 방법, 연결 확인 실패·빈 응답·속도 제한 문제의 해결법을 정리했습니다. - [LibreChat에 BroadNet Inference를 커스텀 엔드포인트로 추가하기](https://broadrouter.com/ko/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): librechat.yaml에 BroadNet Inference를 OpenAI 호환 커스텀 엔드포인트로 추가하는 방법입니다. .env의 키, 모델 목록, 제목 모델, 문제 해결까지 다룹니다. - [n8n에서 BroadNet Inference 사용하기(OpenAI 자격 증명, Base URL)](https://broadrouter.com/ko/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): n8n의 OpenAI 자격 증명에서 Base URL을 BroadNet Inference로 바꾸고, OpenAI Chat Model 노드에서 모델을 ID로 지정한 뒤 Responses API 토글을 끄는 방법입니다. - [BroadNet Inference에서 Vercel AI SDK 사용하기](https://broadrouter.com/ko/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): @ai-sdk/openai-compatible의 createOpenAICompatible로 Vercel AI SDK에서 BroadNet Inference를 호출합니다. generateText, streamText, 도구, reasoning effort까지 다룹니다. - [아랍어 음성 인식(STT) API 사용법: 단계별 가이드](https://broadrouter.com/ko/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): OpenAI 호환 엔드포인트에서 BroadNet Transcribe Arabic으로 아랍어 오디오를 전사하는 방법입니다. 지원 형식, language 필드, Python과 curl 예제, 가격을 정리했습니다. - [LiteLLM을 BroadNet Inference에 연결하기(SDK와 프록시)](https://broadrouter.com/ko/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): LiteLLM에서 openai/ 모델 접두사와 api_base로 BroadNet Inference를 호출합니다. Python SDK와 프록시 config.yaml 설정, 키에 맞춘 속도 제한까지 다룹니다. - [BroadNet Inference에서 OpenAI Agents SDK 실행하기](https://broadrouter.com/ko/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): AsyncOpenAI와 OpenAIChatCompletionsModel로 OpenAI Agents SDK(Python)를 BroadNet Inference에 연결하고, OpenAI 키가 필요한 트레이싱을 끄는 방법을 설명합니다. - [API로 WhatsApp 음성 메시지를 텍스트로 전사하기](https://broadrouter.com/ko/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): WhatsApp .ogg 음성 메시지나 휴대폰 .m4a 녹음을 ffmpeg로 WAV로 변환하고, API 호출 한 번으로 아랍어 또는 다른 13개 언어로 전사하는 방법입니다. - [n8n에서 WhatsApp 음성 메시지 전사하기](https://broadrouter.com/ko/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): WhatsApp 음성 메시지를 받아 Meta에서 내려받고, ffmpeg로 변환한 뒤 아랍어나 다른 13개 언어로 전사하는 n8n 워크플로입니다. - [비전 모델로 아랍어 인보이스와 영수증 읽기](https://broadrouter.com/ko/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): 인보이스나 영수증 사진을 BroadNet Inference의 비전 모델에 보내 합계, 날짜, 공급업체를 JSON으로 받습니다. 질문은 아랍어나 영어로 할 수 있습니다. - [Zed에서 BroadNet Inference를 OpenAI 호환 제공자로 쓰기](https://broadrouter.com/ko/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Zed 에디터의 settings.json에 BroadNet Inference를 OpenAI 호환 제공자로 추가하는 방법입니다. 모델 ID, 토큰 한도, capabilities, 키 환경 변수를 정리했습니다. - [Kilo Code에서 BroadNet Inference 쓰기(OpenAI Compatible 제공자)](https://broadrouter.com/ko/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): VS Code 대화 상자나 kilo.json으로 Kilo Code를 BroadNet Inference에 OpenAI Compatible 커스텀 제공자로 연결하고, 모델 ID, 한도, 도구 플래그를 설정합니다. - [goose에 BroadNet Inference를 커스텀 제공자로 추가하기](https://broadrouter.com/ko/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): goose AI 에이전트를 BroadNet Inference에 OpenAI Compatible 커스텀 제공자로 연결합니다. goose Desktop, goose configure, JSON 파일 세 가지 방법을 다룹니다. - [BroadNet Inference에서 Pydantic AI 사용하기(OpenAICompatibleProvider)](https://broadrouter.com/ko/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): OpenAIChatModel과 OpenAICompatibleProvider로 Pydantic AI 에이전트를 BroadNet Inference에서 실행하고, 도구 호출, 구조화된 출력, 모델 프로필 설정까지 다룹니다. - [CrewAI에서 BroadNet Inference를 커스텀 OpenAI 엔드포인트로 쓰기](https://broadrouter.com/ko/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): LLM(custom_openai=True, base_url=...)으로 CrewAI 에이전트를 BroadNet Inference에 연결하고, 도구·긴 컨텍스트·정액 플랜에 맞는 모델을 고르는 방법을 설명합니다. - [OpenAILike로 LlamaIndex와 BroadNet Inference 연결하기](https://broadrouter.com/ko/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): OpenAILike 클래스로 LlamaIndex를 BroadNet Inference에 연결하고 api_base, context_window, is_chat_model, is_function_calling_model을 올바르게 설정하는 방법입니다. - [OpenAI-API-compatible 플러그인으로 Dify에 BroadNet Inference 추가하기](https://broadrouter.com/ko/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): OpenAI-API-compatible 모델 제공자 플러그인으로 Dify를 BroadNet Inference에 연결합니다. 필드별 입력값, 컨텍스트 크기, 함수 호출 설정을 정리했습니다. - [Jan에서 BroadNet Inference를 OpenAI 호환 커스텀 제공자로 쓰기](https://broadrouter.com/ko/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Jan 데스크톱 앱에 BroadNet Inference를 OpenAI 호환 커스텀 제공자로 추가하고, 모델 ID를 직접 입력한 뒤 모델별로 도구와 비전을 켜는 방법입니다. - [SillyTavern을 BroadNet Inference에 연결하기(Custom OpenAI 호환)](https://broadrouter.com/ko/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): SillyTavern의 Custom (OpenAI-compatible) Chat Completion 소스로 BroadNet Inference에 연결합니다. 올바른 모델 ID, 프롬프트 후처리, 토큰 한도까지 정리했습니다. - [BroadNet Inference에서 Spring AI 사용하기(OpenAI base-url)](https://broadrouter.com/ko/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Spring AI의 OpenAI 채팅 스타터를 spring.ai.openai.base-url, BroadNet 모델 ID, max-tokens로 BroadNet Inference에 연결하고 ChatClient로 호출합니다. - [openai-go SDK로 Go에서 BroadNet Inference 호출하기](https://broadrouter.com/ko/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): 공식 openai-go v3 SDK에서 option.WithBaseURL로 BroadNet Inference에 연결합니다. Go에서 chat completions, 사용량 포함 스트리밍, 오류 처리까지 다룹니다. - [BroadNet Inference로 LangGraph 에이전트 만들기](https://broadrouter.com/ko/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): ChatOpenAI, base_url, BroadNet 모델 ID로 LangGraph StateGraph나 create_agent를 BroadNet Inference에서 실행하는 방법과 도구, 스트리밍, 메모리 설정을 알아봅니다. - [Haystack 파이프라인과 에이전트를 BroadNet Inference로 실행하기](https://broadrouter.com/ko/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): api_base_url, BroadNet 모델 ID, Secret 키로 Haystack의 OpenAIChatGenerator를 BroadNet Inference에 연결해 파이프라인, 도구, 스트리밍, 에이전트에 씁니다. - [BroadNet Inference에서 Semantic Kernel 사용하기](https://broadrouter.com/ko/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Microsoft Semantic Kernel을 Python에서는 AsyncOpenAI 클라이언트로, C#에서는 커스텀 엔드포인트로 BroadNet Inference에 연결해 채팅, 플러그인, 에이전트에 씁니다. - [BroadNet Inference에서 Mastra 에이전트 사용하기](https://broadrouter.com/ko/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): id, url, apiKey를 지정한 model 객체로 Mastra 에이전트를 BroadNet Inference에서 실행하는 방법과, 모델 ID에 broadnet/ 접두사를 두 번 써야 하는 이유를 설명합니다. - [BroadNet Inference에서 Instructor로 구조화된 출력 받기](https://broadrouter.com/ko/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): 커스텀 base_url과 from_openai 또는 from_provider로 Instructor를 BroadNet Inference에 연결해 Pydantic 객체를 추출하고, 알맞은 Mode를 고르는 방법입니다. - [DSPy에서 BroadNet Inference 쓰기(model_type="chat" 설정)](https://broadrouter.com/ko/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): DSPy 3.5는 OpenAI 모델을 Responses API로 보내므로, openai/ 접두사와 api_base, model_type="chat"으로 dspy.LM을 BroadNet Inference에 맞게 설정합니다. - [promptfoo로 BroadNet 모델 평가하기](https://broadrouter.com/ko/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): openai:chat 제공자와 apiBaseUrl, apiKeyEnvar로 BroadNet Inference에서 promptfoo 평가와 모델 비교 테스트를 실행하고, 채점 모델까지 설정합니다. - [OpenAIModel로 BroadNet Inference에서 smolagents 실행하기](https://broadrouter.com/ko/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Hugging Face smolagents를 OpenAIModel과 api_base로 BroadNet Inference에 연결합니다. CodeAgent는 어떤 모델에서든, ToolCallingAgent는 GLM 5.3과 Qwen3.8 27B에서 씁니다. - [Flowise에 BroadNet Inference 연결하기(OpenAI Custom Model)](https://broadrouter.com/ko/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Flowise 챗플로우에서 OpenAI Custom Model 노드, Base Path, OpenAI API 자격 증명, 직접 입력한 BroadNet 모델 이름으로 BroadNet Inference를 쓰는 방법입니다. - [Langflow에 BroadNet Inference 연결하기](https://broadrouter.com/ko/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Langflow의 OpenAI 컴포넌트에서 OpenAI API Base 필드와 직접 입력한 모델 이름으로 BroadNet Inference를 쓰는 방법과, OpenAI Compatible 제공자 목록에 나타나지 않는 이유를 설명합니다. - [Glossary entries in Korean](https://broadrouter.com/ko/glossary/) - [AI 에이전트(AI agent)](https://broadrouter.com/ko/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): AI 에이전트는 언어 모델이 목표를 향해 스스로 단계를 정하는 시스템입니다. 어떤 도구를 호출할지 결정하고, 결과를 읽고, 최종 답변을 내거나 한도에 닿을 때까지 이를 반복합니다. 정해진 지점에서 모델을 호출하는 고정된 스크립트는 보통 워크플로라고 부릅니다. - [API 키(API key)](https://broadrouter.com/ko/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): API 키는 API를 호출하는 클라이언트를 식별하고 승인하는 비밀 문자열입니다. LLM API는 보통 이 키를 Authorization: Bearer 헤더로 받으며, 키를 가진 사람은 누구나 키 소유 계정으로 과금되는 요청을 보낼 수 있습니다. - [아랍어 발음 구별 기호(타슈킬, tashkeel)](https://broadrouter.com/ko/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): 아랍어 발음 구별 기호, 즉 타슈킬(tashkeel)은 글자 위아래에 붙여 짧은 모음, 겹자음, 모음 없음을 나타내는 작은 기호입니다. 필수가 아니어서 뉴스를 포함한 대부분의 아랍어 텍스트는 이를 생략하며, 독자와 소프트웨어는 문맥으로 모음을 추론합니다. - [아랍어 방언 식별(Arabic dialect identification)](https://broadrouter.com/ko/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): 아랍어 방언 식별은 텍스트나 녹음이 이집트, 걸프, 레반트, 마그레브 방언이나 현대 표준 아랍어 중 어떤 변종인지를 지역, 국가, 도시 단위로 판별하는 작업입니다. 메시지 라우팅, 프롬프트 선택, 평가 세트 분할에 쓰입니다. - [아랍어 텍스트 정규화(Arabic text normalization)](https://broadrouter.com/ko/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): 아랍어 텍스트 정규화는 같은 단어의 여러 표기를 매칭 전에 하나의 형태로 바꾸는 단계입니다. 발음 구별 기호와 타트윌을 지우고, 알리프 형태를 통일하며, ى를 ي로, ة를 ه로 바꾸는 경우가 많습니다. 화면에 보여 줄 텍스트가 아니라 검색 키와 평가에 적용합니다. - [아라비지(Arabizi)](https://broadrouter.com/ko/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): 아라비지(Arabizi)는 아랍어를 라틴 문자로 비공식적으로 적는 방식으로, ع를 3으로 쓰는 것처럼 라틴 문자에 없는 아랍어 소리를 숫자로 나타냅니다. 채팅과 소셜 미디어에서 흔하고, 대부분 방언이며, 표준 철자가 없고, 영어나 프랑스어와 섞이는 경우가 많습니다. - [AWQ(활성화 인식 가중치 양자화)](https://broadrouter.com/ko/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ(activation-aware weight quantization, 활성화 인식 가중치 양자화)는 LLM의 가중치를 주로 4비트 정수 같은 낮은 정밀도로 압축하는 학습 후 양자화 방법입니다. 작은 보정(calibration) 세트의 활성값 통계로 가장 중요한 가중치 채널을 찾아 스케일을 조정해 정확도 손실을 줄이며, 역전파가 필요 없습니다. - [base URL(베이스 URL)](https://broadrouter.com/ko/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): base URL은 https://api.example.com/v1처럼 API의 고정된 루트 주소로, 클라이언트가 /chat/completions 같은 모든 엔드포인트 경로 앞에 붙입니다. OpenAI 호환 SDK에서는 base URL을 API 키, 모델 ID와 함께 바꾸면 같은 코드로 다른 제공업체를 호출할 수 있습니다. - [배치 API(Batch API)](https://broadrouter.com/ko/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): 배치 API는 많은 요청을 보통 파일 형태로 한 번에 받아, 즉시가 아니라 정해진 시간 창 안에 처리한 뒤 결과를 함께 돌려주는 비동기 LLM 엔드포인트입니다. OpenAI와 Anthropic 같은 제공업체는 배치 요청을 표준 API보다 할인된 가격으로 제공합니다. - [BM25](https://broadrouter.com/ko/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25(Okapi BM25)는 키워드 검색용 랭킹 함수로, 쿼리 단어가 문서에 얼마나 자주 나오는지, 그 단어가 전체 컬렉션에서 얼마나 드문지, 문서가 얼마나 긴지를 바탕으로 문서 점수를 매깁니다. Lucene과 Elasticsearch의 기본 유사도 함수입니다. - [바이트 페어 인코딩(BPE)](https://broadrouter.com/ko/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): 바이트 페어 인코딩(BPE)은 개별 문자나 바이트에서 시작해, 가장 자주 나란히 나오는 쌍을 새 어휘 항목으로 반복해서 병합하다가 목표 어휘 크기에 이르면 멈추는 서브워드 토큰화 알고리즘입니다. 현재 대부분의 LLM 토크나이저는 BPE나 그 바이트 수준 변형을 씁니다. - [캐시된 입력 토큰(cached input tokens)](https://broadrouter.com/ko/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): 캐시된 입력 토큰은 LLM 제공업체가 최근에 처리해 두었다가 다시 계산하지 않고 캐시에서 재사용할 수 있는 프롬프트 토큰입니다. 보통 시스템 프롬프트나 문서처럼 반복되는 접두부(prefix)이며, 제공업체는 이를 새 입력 토큰보다 낮은 단가로 과금합니다. - [생각의 사슬(CoT, chain-of-thought)](https://broadrouter.com/ko/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): 생각의 사슬(chain-of-thought, CoT)은 언어 모델이 최종 답 전에 중간 추론 단계를 쓰도록 예시를 보여 주거나 요청하는 프롬프팅 기법입니다. 산술과 논리처럼 여러 단계가 필요한 문제에서 정확도를 높이며, 지금은 추론 모델이 이를 자동으로 생성합니다. - [문자 오류율(CER)](https://broadrouter.com/ko/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): 문자 오류율(CER)은 전사나 OCR의 정확도를 문자 단위로 측정하는 지표로, 출력을 정답 텍스트로 바꾸는 데 필요한 치환, 삭제, 삽입 횟수를 정답 텍스트의 문자 수로 나눈 값입니다. 낮을수록 좋고, 0이면 완전히 일치합니다. - [Chat Completions API](https://broadrouter.com/ko/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Chat Completions API는 OpenAI의 HTTP 엔드포인트인 POST /v1/chat/completions로, 모델 이름과 역할이 붙은 메시지 목록을 받아 모델의 다음 메시지를 반환합니다. 이 요청과 응답 구조는 OpenAI 호환 제공업체와 SDK가 재사용하는 공통 형식이 되었습니다. - [채팅 템플릿(chat template)](https://broadrouter.com/ko/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): 채팅 템플릿은 역할이 붙은 채팅 메시지 목록을 모델이 기대하는 단일 토큰 시퀀스로 바꾸는 서식 규칙으로, 모델 고유의 특수 제어 토큰까지 포함합니다. 보통 Jinja로 작성되어 모델의 토크나이저와 함께 배포됩니다. 잘못된 템플릿을 쓰면 모델의 답변 품질이 크게 떨어질 수 있습니다. - [청킹(chunking, RAG)](https://broadrouter.com/ko/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): 청킹(chunking)은 검색 파이프라인에서 문서를 임베딩하고 인덱싱하기 전에 청크라고 부르는 작은 조각으로 나누는 단계입니다. 청크 크기는 조각 하나에 담기는 텍스트 양을 정하고, 청크 오버랩은 이웃한 청크 사이에 일부 텍스트를 반복해 경계에서 잘린 내용이 사라지지 않게 합니다. - [코드 스위칭(code-switching)](https://broadrouter.com/ko/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): 코드 스위칭(code-switching)은 말이나 글에서 같은 대화나 문장 안에 둘 이상의 언어, 또는 한 언어의 여러 변종을 섞어 쓰는 것입니다. 아랍어 사용자가 방언에 영어 단어를 섞거나, 현대 표준 아랍어와 자기 방언을 오가는 것이 그 예입니다. - [AI 코딩 에이전트(AI coding agent)](https://broadrouter.com/ko/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): AI 코딩 에이전트는 언어 모델이 코드베이스를 대상으로 루프를 돌며 일하는 소프트웨어 도구입니다. 파일을 읽고, 수정하고, 테스트 같은 명령을 실행하고, 출력을 읽은 뒤 다음 단계를 정하는 과정을 작업이 끝나거나 한도에 이를 때까지 반복합니다. - [동시 처리(concurrency)](https://broadrouter.com/ko/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): 동시 처리(concurrency)는 클라이언트가 같은 시점에 진행 중인 요청 수입니다. LLM API는 분당 요청 수와 별도로 키나 플랜마다 이를 제한하는 경우가 많습니다. 열려 있는 요청은 응답이 끝날 때까지 GPU 용량을 차지하기 때문입니다. - [컨텍스트 압축(context compaction)](https://broadrouter.com/ko/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): 컨텍스트 압축(context compaction)은 컨텍스트 윈도우 한도에 가까워진 대화를 요약하고 그 요약에서 세션을 이어 가는 방식입니다. 에이전트는 핵심 결정과 상태는 유지하고, 오래된 도구 출력 같은 원시 기록은 버립니다. - [컨텍스트 엔지니어링(context engineering)](https://broadrouter.com/ko/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): 컨텍스트 엔지니어링(context engineering)은 언어 모델이 요청마다 보는 토큰 집합, 즉 시스템 프롬프트, 도구 정의, 대화 기록, 검색한 문서, 메모를 고르고 관리하는 작업입니다. 목표는 모델이 작업을 잘 해내는 데 필요한, 신호가 강한 정보의 가장 작은 집합을 건네는 것입니다. - [컨텍스트 로트(context rot)](https://broadrouter.com/ko/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): 컨텍스트 로트(context rot)는 작업이 그대로이고 입력이 컨텍스트 윈도우 안에 충분히 들어가더라도, 입력이 길어질수록 언어 모델의 정확도와 신뢰성이 점점 떨어지는 현상입니다. 18개 모델을 시험한 2025년 Chroma 기술 보고서로 널리 알려진 용어입니다. - [컨텍스트 윈도우(context window)](https://broadrouter.com/ko/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): 컨텍스트 윈도우는 언어 모델이 요청 한 번에 다룰 수 있는 최대 토큰 수입니다. 프롬프트, 대화 기록, 도구 정의, 이미지, 그리고 추론 토큰을 포함한 생성 출력이 모두 여기에 들어갑니다. 들어가지 않는 내용은 잘라 내거나 요약하거나 여러 요청으로 나눠야 합니다. - [데이터 레지던시(data residency)](https://broadrouter.com/ko/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): 데이터 레지던시(data residency)는 데이터를 특정 국가나 지역 안에서 저장하고 처리해야 한다는 요건 또는 약속입니다. LLM API에서는 추론 중 프롬프트가 처리되는 위치와 저장된 콘텐츠가 있는 위치를 다루며, 계정과 결제 데이터는 제외되는 경우가 많습니다. - [아랍어 방언(dialectal Arabic)](https://broadrouter.com/ko/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): 아랍어 방언(dialectal Arabic)은 이집트, 걸프, 레반트, 마그레브 아랍어처럼 지역이나 도시에서 일상적으로 쓰는 구어 아랍어입니다. 방언은 소리, 어휘, 문법에서 현대 표준 아랍어와 다르고, 표준 철자가 없으며, 온라인에서 점점 더 많이 글로 쓰이기 때문에 언어 모델이 다루기 더 어렵습니다. - [화자 분리(speaker diarization)](https://broadrouter.com/ko/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): 화자 분리(speaker diarization)는 오디오 녹음을 화자별 구간으로 나눠 "누가 언제 말했는가"에 답하는 작업입니다. 각 발화 구간에 SPEAKER_00 같은 익명 화자 ID를 붙이며, 보통 전사본과 결합해 각 줄을 누가 말했는지 보여 줍니다. - [지식 증류(knowledge distillation)](https://broadrouter.com/ko/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): 증류(distillation), 또는 지식 증류는 더 작은 학생 모델이 더 큰 교사 모델의 출력, 예컨대 확률 분포나 생성한 답을 재현하도록 학습하는 방법입니다. 그 결과 학생 모델은 더 낮은 실행 비용으로 교사 모델의 품질을 상당 부분 유지합니다. - [DPA(데이터 처리 계약)](https://broadrouter.com/ko/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): DPA(데이터 처리 계약)는 처리자가 개인정보를 어떻게 다룰지 정하는 컨트롤러와 처리자 사이의 계약입니다. GDPR 제28조 제3항이 이를 요구하며 내용도 정합니다. 문서화된 지시, 비밀 유지, 보안, 하위 처리자 규칙, 데이터 삭제 또는 반환, 감사가 들어갑니다. - [임베딩(embedding)](https://broadrouter.com/ko/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): 임베딩은 모델이 텍스트, 이미지 등의 입력을 표현하기 위해 만드는 부동소수점 숫자 목록(벡터)입니다. 의미가 비슷한 입력은 서로 가까운 벡터를 얻으므로, 벡터 사이의 거리로 검색 결과의 순위를 매기거나 문서를 군집화하거나 관련 항목을 찾을 수 있습니다. - [LLM 평가(eval)](https://broadrouter.com/ko/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): eval(evaluation의 줄임말, 평가)은 언어 모델이나 프롬프트를 반복해서 시험하는 테스트입니다. 고정된 입력 세트, 좋은 출력의 기준, 그리고 각 출력에 점수를 매기는 채점기(코드, 사람 또는 다른 모델)로 이루어져 버전끼리 비교할 수 있게 해 줍니다. - [지수 백오프(exponential backoff)](https://broadrouter.com/ko/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): 지수 백오프(exponential backoff)는 클라이언트가 시도에 실패할 때마다 더 오래 기다렸다가 다시 시도하는 재시도 전략으로, 보통 상한까지 지연 시간을 두 배씩 늘립니다. 각 대기 시간에 무작위 지터를 더하면 여러 클라이언트의 재시도가 분산되어 같은 순간에 서버로 몰리지 않습니다. - [퓨샷 프롬프팅(few-shot prompting)](https://broadrouter.com/ko/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): 퓨샷 프롬프팅(few-shot prompting)은 실제 입력 앞에 입력-출력 예시를 몇 개 넣어, 언어 모델이 예시에서 작업과 형식을 추론하게 하는 기법입니다. 모델의 가중치는 바뀌지 않습니다. 예시가 없으면 제로샷, 하나면 원샷이라고 합니다. - [파인튜닝(fine-tuning)](https://broadrouter.com/ko/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): 파인튜닝(fine-tuning, 미세 조정)은 사전 학습된 언어 모델을 예시 대화 같은 작고 작업에 특화된 데이터셋으로 계속 학습시키는 과정입니다. 가중치가 바뀌므로, 모델은 프롬프트마다 긴 지시문이나 예시를 넣지 않아도 해당 작업이나 스타일을 더 안정적으로 수행합니다. - [finish_reason(종료 이유)](https://broadrouter.com/ko/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason은 chat completion 응답의 각 choice에 붙는 필드로, 모델이 생성을 멈춘 이유를 알려 줍니다. OpenAI 형식의 값은 stop, length, tool_calls, content_filter, 그리고 deprecated된 function_call입니다. Anthropic Messages API는 같은 개념을 stop_reason으로 보고합니다. - [LLM 정액 플랜(Flat-rate LLM plan)](https://broadrouter.com/ko/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): LLM 정액 플랜은 언어 모델 API를 토큰당 과금하지 않고, 한 달 같은 기간 단위로 고정 요금을 받는 요금제입니다. 이런 플랜은 보통 동시 실행 수, 속도, 모델, 사용 시간대 같은 다른 항목을 제한하므로 가격만큼 제한 조건도 중요합니다. - [FP8(8비트 부동소수점)](https://broadrouter.com/ko/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8은 딥러닝에 쓰이는 8비트 부동소수점 숫자 형식으로, 지수 4비트와 가수 3비트인 E4M3, 지수 5비트와 가수 2비트인 E5M2 두 가지 인코딩이 있습니다. 16비트 형식보다 메모리를 절반으로 줄이며, 최신 GPU에서 네이티브로 실행됩니다. - [빈도 페널티와 존재 페널티(frequency/presence penalty)](https://broadrouter.com/ko/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): 빈도 페널티(frequency penalty)와 존재 페널티(presence penalty)는 출력에 이미 나온 토큰의 점수를 낮추는 샘플링 설정입니다. 존재 페널티는 토큰이 한 번이라도 나오면 한 번 적용되고, 빈도 페널티는 반복될 때마다 커집니다. OpenAI API에서 둘 다 범위는 -2.0~2.0이고 기본값은 0입니다. - [GDPR(EU 일반 개인정보 보호법)](https://broadrouter.com/ko/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): GDPR(일반 개인정보 보호법, 규정(EU) 2016/679)은 개인정보 처리에 관한 EU 법입니다. 2018년 5월 25일부터 적용되며, EU에 설립된 조직과 일부 EU 밖 조직에 적용됩니다. LLM API에서는 처리자 계약, 국외 이전, 데이터 최소화를 규율합니다. - [GGUF(llama.cpp 모델 파일 형식)](https://broadrouter.com/ko/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF는 llama.cpp 같은 GGML 기반 런타임에서 추론하기 위해 언어 모델을 저장하는 바이너리 파일 형식입니다. GGUF 파일 하나에 (대개 양자화된) 가중치와 토크나이저, 아키텍처 정보 등 모델을 불러오는 데 필요한 메타데이터가 들어 있으며, 메모리 매핑으로 빠르게 불러올 수 있습니다. - [GPU 시간(GPU-hour)](https://broadrouter.com/ko/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): GPU 시간(GPU-hour)은 그래픽 프로세서 하나를 1시간 사용한 양입니다. 클라우드 제공업체는 임대한 GPU 인스턴스를 시간 또는 초 단위로 청구하고, 연구자는 작업에 든 연산량을 GPU 시간으로 나타냅니다. 기계가 일을 하든 안 하든 그 시간만큼 비용을 냅니다. - [그라운딩(grounding)](https://broadrouter.com/ko/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): 그라운딩(grounding)은 언어 모델의 출력을 제공한 문서, 검색 결과, 데이터베이스 레코드처럼 구체적이고 검증 가능한 정보 출처에 연결하고, 그 출처를 근거로 답하도록 지시하는 방법입니다. 그라운딩된 답은 출처까지 추적할 수 있고 지어낸 사실이 들어갈 가능성이 낮습니다. - [가드레일(guardrails)](https://broadrouter.com/ko/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): 가드레일(guardrails)은 언어 모델 호출이 정해진 한계 안에 머물도록 애플리케이션이 호출 앞뒤에서 실행하는 검사입니다. 입력 가드레일은 모델에 들어가는 내용을, 출력 가드레일은 나오는 내용을 검사하고, 도구 가드레일은 실행 전에 동작을 검증합니다. 검사에 실패하면 요청을 차단하거나 고치거나 표시합니다. - [할루시네이션(환각, hallucination)](https://broadrouter.com/ko/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): 할루시네이션(환각)은 언어 모델의 출력이 그럴듯하게 들리지만 사실이 아니거나, 지어냈거나, 주어진 입력으로 뒷받침되지 않는 경우를 말합니다. 지어낸 인용, 틀린 숫자, 존재하지 않는 API 등이 그 예입니다. 모델이 사실을 확인하지 않고 그럴듯한 텍스트를 예측하기 때문에 생깁니다. - [함자(hamza)](https://broadrouter.com/ko/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): 함자(hamza, ء)는 성문 파열음을 나타내는 아랍어 글자입니다. 단독으로 쓰거나 받침 글자 위에 씁니다. 받침은 알리프(أ إ), 와우(ؤ), 야(ئ)이며, آ는 알리프와 마다를 합친 것입니다. 입력된 텍스트에서 함자 표기가 일관되지 않아 검색 엔진과 음성 채점은 알리프 형태들을 ا 하나로 합치는 경우가 많습니다. - [하이브리드 검색(hybrid search)](https://broadrouter.com/ko/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): 하이브리드 검색(hybrid search)은 같은 쿼리로 BM25 같은 키워드 검색과 임베딩 기반 벡터 검색을 함께 실행한 뒤, 두 순위 목록을 하나로 합치는 검색 방식입니다. 합칠 때는 주로 역순위 융합(RRF)이나 정규화한 점수의 가중합을 씁니다. - [멱등성 키(idempotency key)](https://broadrouter.com/ko/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): 멱등성 키(idempotency key)는 결제나 구매처럼 상태를 바꾸는 요청에 클라이언트가 붙이는 고유한 값으로, 보통 UUID 같은 무작위 ID입니다. 같은 요청이 같은 키로 다시 오면 서버는 작업을 두 번 수행하지 않고 처음 결과를 돌려줍니다. - [이미지 토큰(image tokens)](https://broadrouter.com/ko/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): 이미지 토큰(image tokens)은 비전 언어 모델이 프롬프트 속 이미지를 표현하는 데 쓰는 입력 토큰입니다. 제공업체는 보통 이미지를 리사이즈한 뒤 고정 크기 패치나 타일로 나눠 토큰 수를 계산하고, 텍스트 입력 토큰처럼 청구하므로 큰 이미지일수록 비용이 늘어납니다. - [추론 제공업체(Inference provider)](https://broadrouter.com/ko/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): 추론 제공업체(inference provider)는 머신러닝 모델, 주로 오픈 웨이트 LLM을 자체 소유, 장기 임차 또는 단기 임대한 GPU에서 실행하고, API로 이용 권한을 판매하는 회사입니다. 요금은 보통 토큰, 요청, 시간 단위로 매기며, 서빙과 확장, 하드웨어 관리를 맡습니다. - [입력 토큰과 출력 토큰(input and output tokens)](https://broadrouter.com/ko/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): 입력 토큰은 언어 모델이 요청에서 읽는 토큰으로, 시스템 프롬프트, 대화 기록, 도구 정의, 이미지가 포함됩니다. 출력 토큰은 모델이 응답으로 생성하는 토큰으로, 추론도 포함됩니다. LLM API는 보통 둘에 따로 가격을 매기며, 대개 출력이 더 비쌉니다. - [INT4(4비트 정수 양자화)](https://broadrouter.com/ko/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4는 16단계를 표현할 수 있는 4비트 정수 형식입니다. LLM 추론에서는 보통 4비트 가중치 전용 양자화(W4A16)를 뜻합니다. 가중치를 스케일 계수와 함께 INT4로 저장했다가 계산할 때 16비트로 풀어 쓰며, 16비트 가중치의 약 4분의 1 메모리만 씁니다. - [IP 허용 목록(IP allow list)](https://broadrouter.com/ko/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): IP 허용 목록(IP allow list)은 목록에 있는 IP 주소나 CIDR 범위에서 온 요청만 받고 나머지는 모두 거부하는 보안 통제입니다. API에서는 키가 유출되어도 목록에 없는 네트워크에서는 작동하지 않으므로 피해를 줄여 줍니다. - [LLM 탈옥(jailbreak)](https://broadrouter.com/ko/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): 탈옥(jailbreak)은 언어 모델이 안전 학습이나 모델을 배포한 애플리케이션의 규칙을 무시하게 만들어, 평소라면 거부할 콘텐츠를 생성하거나 행동을 하도록 꾸민 입력입니다. OWASP는 탈옥을 프롬프트 인젝션의 한 형태로 봅니다. - [JSON 모드(JSON mode)](https://broadrouter.com/ko/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): JSON 모드는 OpenAI 방식 API의 response_format {"type": "json_object"} 같은 LLM API 설정으로, 모델의 응답을 파싱 가능한 유효한 JSON으로 만듭니다. JSON Schema를 쓰는 구조화된 출력과 달리, 객체에 어떤 키와 타입, 값이 들어갈지는 보장하지 않습니다. - [JSON Schema(JSON 스키마)](https://broadrouter.com/ko/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema는 JSON으로 작성하는 어휘로, 타입, 필수 필드, 허용 값, 중첩 같은 JSON 데이터의 구조를 기술하고 검증합니다. 현재 버전은 2020-12입니다. LLM API는 이를 사용해 도구 파라미터와 구조화된 출력이 따라야 할 정확한 형태를 정의합니다. - [KV 캐시(KV cache)](https://broadrouter.com/ko/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): KV 캐시(KV cache)는 트랜스포머 언어 모델이 이미 처리한 모든 토큰에 대해 계산한 어텐션 키와 값을 저장하는 메모리입니다. 새 토큰은 전체 시퀀스를 다시 계산하지 않고 이를 재사용합니다. 컨텍스트 길이에 따라 커지며, GPU 하나가 처리할 수 있는 요청 수를 제한하는 경우가 많습니다. - [지연 시간(latency)](https://broadrouter.com/ko/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): 언어 모델 API에서 지연 시간(latency)은 요청을 보낸 시점부터 마지막 출력 토큰을 받을 때까지 걸리는 시간입니다. 첫 토큰까지의 시간(TTFT)과 생성 시간을 더한 값이며, 보통 평균이 아니라 p50, p95 같은 백분위수로 보고합니다. - [LLM-as-a-judge(LLM 평가자)](https://broadrouter.com/ko/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge는 언어 모델이 다른 모델의 출력을 채점하는 평가 방법으로, 답변 하나를 루브릭에 따라 점수 매기거나 두 답변 중 더 나은 쪽을 고릅니다. 훨씬 적은 비용으로 사람의 판단에 근접하지만, 첫 번째 답이나 더 긴 답을 선호하는 등 알려진 편향이 있습니다. - [LLM 벤치마크(LLM benchmark)](https://broadrouter.com/ko/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): LLM 벤치마크(LLM benchmark)는 채점 방법이 정해진, 대개 공개된 고정 작업 세트로, 지식 문제, 코딩, 음성 인식처럼 한 가지 능력에서 언어 모델을 측정하고 비교하는 데 씁니다. 벤치마크 점수는 그 테스트에서의 상대적 능력을 보여 줄 뿐, 내 워크로드에서의 성능을 보여 주지는 않습니다. - [LLM 라우터(LLM router)](https://broadrouter.com/ko/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): LLM 라우터(LLM router)는 애플리케이션과 하나 이상의 언어 모델 API 사이에 위치해, 요청마다 어떤 모델이나 제공업체가 처리할지 정하는 소프트웨어나 서비스입니다. 판단 기준으로는 비용, 지연 시간, 가용성, 작업 유형 같은 규칙을 씁니다. - [Logprobs(로그 확률)](https://broadrouter.com/ko/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Logprobs(로그 확률)는 언어 모델이 생성한 토큰에 부여한 로그 확률입니다. 이를 반환하는 API는 출력 토큰마다 모델이 그 토큰을 얼마나 가능성 있다고 판단했는지, 그리고 원하면 그 위치에서 가능성이 높았던 다른 후보 토큰도 보여 줍니다. - [LoRA(저랭크 적응, low-rank adaptation)](https://broadrouter.com/ko/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA(low-rank adaptation, 저랭크 적응)는 사전 학습된 모델의 가중치를 고정하고, 선택한 레이어에 작은 저랭크 행렬 한 쌍만 학습시키는 파라미터 효율적 파인튜닝 방법입니다. 학습된 변화량은 작은 어댑터로 저장하거나 가중치에 병합할 수 있어 추론 지연 시간이 늘지 않습니다. - [Lost in the middle(중간 정보 손실)](https://broadrouter.com/ko/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle은 언어 모델이 긴 입력에서 관련 정보가 처음이나 끝 근처에 있을 때 가장 잘 활용하고, 중간에 있을 때 가장 못 활용하는 경향을 말합니다. Liu 외의 2023년 연구에서 이름이 붙었습니다. - [max_tokens(최대 출력 토큰)](https://broadrouter.com/ko/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens는 모델이 응답 하나에서 생성할 수 있는 최대 토큰 수를 정하는 LLM API 파라미터입니다. 생성이 이 한도에 닿으면 멈추고, 응답에는 잘렸다는 표시가 붙습니다. 예를 들어 finish_reason이 "length"로 설정됩니다. - [MCP(Model Context Protocol)](https://broadrouter.com/ko/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP(Model Context Protocol)는 JSON-RPC 2.0 위에 만들어진 개방형 프로토콜로, LLM 애플리케이션이 외부 도구와 데이터에 연결하는 방식을 표준화합니다. MCP 서버는 도구, 리소스, 프롬프트를 제공하고, 호스트 애플리케이션 안의 MCP 클라이언트가 이를 찾아 모델의 도구 호출을 서버에 전달합니다. - [전문가 혼합(MoE, mixture of experts)](https://broadrouter.com/ko/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): 전문가 혼합(MoE, mixture of experts)은 한 레이어에 여러 전문가 서브네트워크를 두고, 학습된 라우터가 각 토큰을 그중 몇 개에만 보내는 신경망 설계입니다. MoE 언어 모델은 전체 파라미터 수는 크지만 토큰마다 작은 활성 부분집합만 사용합니다. - [모델 지원 중단(model deprecation)](https://broadrouter.com/ko/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): 모델 지원 중단(model deprecation)은 API 제공업체가 모델 제공을 멈추겠다고 발표하는 것으로, 보통 종료(retirement 또는 shutdown) 날짜와 권장 대체 모델을 함께 밝힙니다. 그 날짜 이후에는 해당 모델로 보낸 요청이 실패하므로, 애플리케이션은 그 전에 다른 모델 ID로 옮겨야 합니다. - [모델 폴백(model fallback)](https://broadrouter.com/ko/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): 모델 폴백(model fallback)은 실패한 LLM 요청을 순서가 정해진 목록의 다른 모델이나 제공업체로 다시 보내는 안정성 패턴입니다. 오류, 속도 제한, 컨텍스트 길이 오류, 콘텐츠 정책 거부가 계기가 될 수 있으며, 클라이언트나 API 앞단의 게이트웨이에서 실행됩니다. - [모델 라이선스(model license)](https://broadrouter.com/ko/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): 모델 라이선스(model license)는 모델의 가중치, 코드, 출력을 사용하고 수정하고 공유할 수 있는 법적 조건입니다. 오픈 웨이트 모델은 Apache 2.0이나 MIT 같은 표준 소프트웨어 라이선스, 또는 대규모 상업 서비스 제한 같은 조건을 추가할 수 있는 자체 모델 라이선스를 씁니다. - [현대 표준 아랍어(MSA, Modern Standard Arabic)](https://broadrouter.com/ko/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): 현대 표준 아랍어(MSA)는 아랍 세계 전역에서 글, 뉴스, 교육, 격식 있는 연설에 쓰는 표준화된 아랍어입니다. 누구도 일상 모국어로 쓰지 않습니다. 사람들은 이집트 아랍어나 걸프 아랍어 같은 지역 방언으로 말하고, 격식 있는 자리에서 MSA로 바꿉니다. - [멀티 에이전트 시스템(multi-agent system)](https://broadrouter.com/ko/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): 멀티 에이전트 시스템(multi-agent system)은 각자 프롬프트, 도구, 컨텍스트를 가진 여러 에이전트가 하나의 작업을 나눠 처리하는 LLM 애플리케이션입니다. 흔한 설계는 오케스트레이터-워커 방식으로, 리드 에이전트가 계획을 세우고 워커 에이전트를 병렬로 실행한 뒤 결과를 합칩니다. - [멀티모달 AI(multimodal AI)](https://broadrouter.com/ko/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): 멀티모달 AI(multimodal AI)는 텍스트, 이미지, 영상, 오디오처럼 두 가지 이상의 데이터 유형을 다루는 모델이나 시스템입니다. 사진과 질문을 함께 읽거나, 음성을 받아 텍스트로 답할 수 있습니다. 가장 흔한 형태는 비전-언어 모델입니다. - [개체명 인식(NER, named entity recognition)](https://broadrouter.com/ko/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): 개체명 인식(NER)은 텍스트에서 이름을 찾아 사람, 조직, 장소 같은 유형을 붙이는 작업입니다. 전통적인 시스템은 모든 토큰에 태그를 붙이고, LLM에는 보통 개체 목록을 JSON으로 반환하게 한 뒤 원문과 대조해 확인합니다. - [광학 문자 인식(OCR)](https://broadrouter.com/ko/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): 광학 문자 인식(OCR)은 이미지, 스캔본, PDF 속 인쇄되거나 손으로 쓴 글자를 찾아 기계가 읽을 수 있는 문자로 변환하는 기술입니다. 전통적인 OCR은 텍스트와 페이지 내 위치를 출력하고, 비전-언어 모델은 텍스트를 읽고 그에 관한 질문에 바로 답할 수 있습니다. - [Ogg Opus(오그 오퍼스)](https://broadrouter.com/ko/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus는 Opus 코덱(RFC 6716)으로 인코딩해 Ogg 컨테이너(RFC 7845)에 저장한 오디오입니다. 낮은 비트레이트의 음성과 실시간 대화형 오디오용으로 설계되었으며, WhatsApp이 음성 메시지에 쓰는 형식으로 보통 .ogg 또는 .opus 파일입니다. - [오픈 웨이트 모델(open-weight model)](https://broadrouter.com/ko/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): 오픈 웨이트 모델은 학습된 가중치를 공개해 누구나 내려받아 실행할 수 있게 한 머신러닝 모델로, 보통 조건이 붙은 라이선스로 배포됩니다. 학습 코드와 학습 데이터에 대한 상세한 정보까지 공개해 다른 사람이 모델을 연구하고 재현할 수 있게 하는 완전한 오픈 소스 AI와는 다릅니다. - [OpenAI 호환 API(OpenAI-compatible API)](https://broadrouter.com/ko/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): OpenAI 호환 API는 POST /v1/chat/completions처럼 OpenAI API와 같은 요청 및 응답 형식을 받는 HTTP API입니다. 그래서 OpenAI SDK와 도구는 base URL, API 키, 모델 이름만 바꾸면 이 API를 그대로 사용할 수 있습니다. - [병렬 도구 호출(parallel tool calls)](https://broadrouter.com/ko/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): 병렬 도구 호출(parallel tool calls)은 모델이 한 번의 응답에서 도구 호출을 둘 이상 반환하는 것입니다. 예를 들어 파일 세 개를 한꺼번에 읽습니다. OpenAI 형식에서는 tool_calls의 여러 항목으로 오며, parallel_tool_calls를 false로 설정하면 턴당 호출이 0개나 1개로 제한됩니다. - [종량제(Pay-as-you-go)](https://broadrouter.com/ko/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): 종량제(pay-as-you-go)는 구독이나 최소 약정 없이 토큰, 요청, 오디오 분 같은 단위로 사용량을 측정해 쓴 만큼만 내는 요금 방식입니다. 대부분의 LLM API가 선불 잔액에서 차감하거나 정기 청구서를 발행하는 식으로 이렇게 과금합니다. - [PDPL(사우디·UAE 개인정보 보호법)](https://broadrouter.com/ko/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL(Personal Data Protection Law)은 칙령 M/19로 공포되고 SDAIA가 집행하는 사우디아라비아 개인정보 보호법과, UAE의 2021년 연방 법령 제45호를 함께 부르는 통칭입니다. 두 법 모두 개인정보 처리를 규율하고, 해외 LLM API로 보내는 프롬프트를 포함한 국외 이전에 조건을 둡니다. - [PII(개인 식별 정보)](https://broadrouter.com/ko/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): PII(개인 식별 정보)는 이름, 신분 번호, 생체 기록처럼 개인의 신원을 구별하거나 추적하는 데 쓸 수 있는 정보와, 그 사람과 연결되거나 연결될 수 있는 기타 정보입니다. GDPR은 더 넓은 용어인 개인 데이터를 씁니다. LLM 작업에서 PII 마스킹은 프롬프트가 내 시스템을 떠나기 전에 이를 제거합니다. - [프리필과 디코드(prefill and decode)](https://broadrouter.com/ko/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): 프리필과 디코드(prefill and decode)는 LLM 추론의 두 단계입니다. 프리필에서 모델은 프롬프트 전체를 병렬로 처리해 KV 캐시를 채우고, 디코드에서는 그 캐시를 재사용하며 답변을 토큰 하나씩 생성합니다. 프리필은 주로 첫 토큰까지의 시간을, 디코드는 출력 속도를 결정합니다. - [선불 크레딧(prepaid credit)](https://broadrouter.com/ko/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): 선불 크레딧(prepaid credit)은 사용 전에 API 계정에 넣어 두는 금액입니다. 요청마다 비용이 잔액에서 차감되며, 잔액이 바닥나면 충전할 때까지 종량제 요청이 중단됩니다. 지출은 넣어 둔 금액을 넘지 않습니다. - [프롬프트 엔지니어링(prompt engineering)](https://broadrouter.com/ko/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): 프롬프트 엔지니어링은 언어 모델이 요구 사항에 맞는 출력을 일관되게 내도록, 모델에 주는 지시문을 작성하고 구조화하고 테스트하는 과정입니다. 역할 메시지, 명확한 지시, 예시, 형식 지정, 그리고 테스트 케이스로 결과를 측정하는 일까지 포함합니다. - [프롬프트 인젝션(prompt injection)](https://broadrouter.com/ko/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): 프롬프트 인젝션(prompt injection)은 사용자가 입력하거나 웹 페이지, 이메일, 파일, 도구 결과에 숨겨진 텍스트가 언어 모델의 동작을 개발자의 의도와 다르게 바꾸는 공격입니다. OWASP는 LLM 애플리케이션 Top 10에서 이를 1위(LLM01:2025)로 꼽습니다. - [양자화(quantization)](https://broadrouter.com/ko/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): 양자화(quantization)는 신경망의 숫자, 주로 가중치를 16비트 대신 8비트나 4비트 같은 낮은 정밀도 형식으로 저장해, 모델이 메모리를 덜 쓰고 대개 더 빠르게 실행되도록 하는 과정입니다. 좋은 방법은 정확도 손실을 작게 유지합니다. - [RAG(검색 증강 생성)](https://broadrouter.com/ko/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG(검색 증강 생성, retrieval-augmented generation)는 애플리케이션이 먼저 검색 인덱스나 벡터 인덱스 같은 외부 저장소에서 질문과 관련된 구절을 찾고, 이를 프롬프트에 추가해 언어 모델이 기억에만 의존하지 않고 그 텍스트를 바탕으로 답하게 하는 패턴입니다. - [속도 제한(rate limit)](https://broadrouter.com/ko/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): 속도 제한(rate limit)은 API가 일정 시간 동안 클라이언트가 사용할 수 있는 양에 두는 상한으로, 분당 요청 수나 분당 토큰 수 같은 기준이 있습니다. 한도를 넘은 요청은 사용량이 다시 한도 아래로 내려갈 때까지 거부되며, 보통 HTTP 429 Too Many Requests가 반환됩니다. - [ReAct(추론과 행동, reason and act)](https://broadrouter.com/ko/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct(추론과 행동)는 Yao 외가 2022년에 소개한 프롬프팅 패턴으로, 언어 모델이 추론 과정과 검색이나 도구 호출 같은 행동을 번갈아 수행합니다. 모델은 각 관찰 결과를 읽고 계획을 고친 뒤, 최종 답변을 낼 수 있을 때까지 이를 반복합니다. - [추론 토큰(reasoning tokens)](https://broadrouter.com/ko/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): 추론 토큰(reasoning tokens)은 추론 모델이 눈에 보이는 답변을 쓰기 전에 문제를 풀기 위해 생성하는 토큰입니다. 보통 출력 토큰으로 과금되고, 출력 한도와 컨텍스트 윈도우에 포함되며, API에 따라 숨겨지거나 요약되어 제공됩니다. - [요청 타임아웃(request timeout)](https://broadrouter.com/ko/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): 요청 타임아웃(request timeout)은 HTTP 클라이언트가 요청을 포기하기 전까지 서버를 기다리는 최대 시간입니다. 클라이언트는 보통 연결을 여는 연결 타임아웃과 데이터를 기다리는 읽기 타임아웃으로 나눕니다. 긴 답변은 시간이 걸리므로 LLM 호출에는 넉넉한 읽기 타임아웃이 필요합니다. - [리랭킹(reranking)](https://broadrouter.com/ko/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): 리랭킹(reranking)은 검색의 두 번째 단계로, 더 정확하지만 느린 모델(주로 크로스 인코더)이 1차 검색에서 나온 질의와 후보 문서 쌍마다 점수를 매겨 순서를 다시 정합니다. 그래서 가장 관련성 높은 몇 개만 언어 모델에 전달됩니다. - [RTL 텍스트(오른쪽에서 왼쪽 쓰기, right-to-left text)](https://broadrouter.com/ko/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): RTL(right-to-left) 텍스트는 아랍어나 히브리어처럼 오른쪽에서 왼쪽으로 표시되는 문자입니다. 컴퓨터는 이를 읽는 순서인 논리 순서로 저장하며, RTL 텍스트가 왼쪽에서 오른쪽으로 쓰는 텍스트, 숫자, 모델 이름, 코드와 섞이면 유니코드 양방향 알고리즘이 표시 순서를 정합니다. - [샘플링 레이트(sample rate)](https://broadrouter.com/ko/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): 샘플링 레이트(sample rate)는 1초 동안 기록하는 오디오 샘플 수로, 헤르츠(Hz) 단위로 나타냅니다. 음성 인식 모델은 보통 16,000Hz(16kHz) 모노 오디오를 기대합니다. 샘플링 레이트는 비트 깊이, 채널 수와 함께 비압축 오디오 파일의 분당 크기를 결정합니다. - [seed 파라미터(시드 값)](https://broadrouter.com/ko/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): seed 파라미터는 요청과 함께 보내는 정수로, 같은 요청을 반복할 때 API가 같은 토큰을 샘플링하도록 시도하게 합니다. OpenAI는 이를 최선 노력(best effort)으로 설명합니다. 결정성은 보장되지 않으며, system_fingerprint 필드로 백엔드 구성이 바뀐 시점을 알 수 있습니다. - [서버리스 추론(serverless inference)](https://broadrouter.com/ko/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): 서버리스 추론은 플랫폼이 엔드포인트 뒤의 컴퓨팅을 자동으로 시작하고 확장하고 중지하는 머신러닝 모델 실행 방식입니다. 유휴 시에는 0까지 줄이는 경우가 많고, 예약한 머신이 아니라 사용량으로 청구합니다. 그 대가로 콜드 스타트 지연이 생길 수 있습니다. - [SLA(서비스 수준 협약, service level agreement)](https://broadrouter.com/ko/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): SLA(서비스 수준 협약)는 제공업체가 측정 가능한 서비스 수준, 보통 월간 가동률을 계약으로 약속하고, 이를 지키지 못했을 때의 구제 수단, 보통 서비스 크레딧을 정해 둔 것입니다. SLA는 무엇을 다운타임으로 보는지, 무엇이 제외되는지, 어떻게 청구하는지를 정의합니다. - [추측 디코딩(speculative decoding)](https://broadrouter.com/ko/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): 추측 디코딩(speculative decoding)은 작고 빠른 초안(draft) 모델이 다음 토큰 여러 개를 제안하고, 큰 대상(target) 모델이 이를 순전파 한 번으로 검증해 받아들인 토큰을 유지하는 추론 기법입니다. 출력 분포를 바꾸지 않고 생성 속도를 높입니다. - [음성 인식(Speech-to-text, STT)](https://broadrouter.com/ko/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): 음성 인식(STT)은 자동 음성 인식(ASR)이라고도 하며, 말소리 오디오를 글로 변환하는 기술입니다. 음성 인식 API는 오디오 파일이나 스트림을 받아 전사 텍스트를 반환하며, 보통 오디오 분당 요금이 붙습니다. 정확도는 단어 오류율로 측정하며 억양, 방언, 녹음 품질에 따라 달라집니다. - [지출 한도(spend limit)](https://broadrouter.com/ko/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): 지출 한도(spend limit)는 API 계정, 프로젝트 또는 키가 일정 기간, 보통 한 달 동안 쓸 수 있는 금액의 상한입니다. 사용량이 한도에 도달하면 제공업체는 기간이 초기화되거나 한도가 올라갈 때까지 이후 요청을 거부하며, 흔히 HTTP 429 또는 400 오류를 반환합니다. - [SRT 자막(SRT subtitles)](https://broadrouter.com/ko/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT(SubRip)는 일반 텍스트 자막 형식입니다. 각 자막은 순번, hh:mm:ss,mmm --> hh:mm:ss,mmm 형태로 쓴 시작·종료 시간, 한 줄 이상의 텍스트, 빈 줄로 이루어집니다. WebVTT는 HTML 비디오에서 쓰는 비슷한 웹 형식입니다. - [서버 전송 이벤트(SSE, Server-Sent Events)](https://broadrouter.com/ko/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): 서버 전송 이벤트(SSE, Server-Sent Events)는 서버가 text/event-stream 콘텐츠 타입의 오래 유지되는 HTTP 응답 하나로 클라이언트에 텍스트 이벤트 스트림을 보내는 웹 표준입니다. 각 이벤트는 data: 같은 필드 줄 하나 이상으로 이루어지며 빈 줄로 끝납니다. - [중지 시퀀스(stop sequences)](https://broadrouter.com/ko/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): 중지 시퀀스(stop sequences)는 요청과 함께 보내는 문자열로, 모델이 그중 하나를 생성하는 즉시 생성을 멈추게 합니다. OpenAI는 이 필드를 stop이라 부르며 최대 4개까지 받고, Anthropic은 stop_sequences라고 부릅니다. 일치한 텍스트는 보통 반환되는 출력에서 빠집니다. - [스트리밍(streaming)](https://broadrouter.com/ko/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): 스트리밍은 언어 모델의 응답을 마지막에 한 번에 돌려주는 대신, 아직 생성 중일 때 작은 청크로 나눠 돌려주는 방식입니다. OpenAI 호환 API는 요청에서 stream을 true로 설정하면 서버 전송 이벤트(SSE)로 스트리밍합니다. - [구조화된 출력(Structured output)](https://broadrouter.com/ko/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): 구조화된 출력(structured output)은 모델의 응답을 개발자가 지정한 형식, 주로 JSON Schema에 맞추도록 제한해 코드가 안정적으로 파싱할 수 있게 하는 LLM API 기능입니다. 유효한 JSON만 보장하고 특정 구조는 보장하지 않는 JSON 모드보다 더 엄격합니다. - [서브에이전트(subagent)](https://broadrouter.com/ko/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): 서브에이전트(subagent)는 메인 에이전트가 집중된 작업 하나를 처리하려고 시작하는 별도의 에이전트입니다. 자체 컨텍스트 윈도우, 시스템 프롬프트, 도구 접근 권한으로 작업을 수행하고, 결과나 요약만 메인 에이전트에 돌려주므로 메인 대화가 작게 유지됩니다. - [하위 처리자(subprocessor)](https://broadrouter.com/ko/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): 하위 처리자(subprocessor)는 처리자가 컨트롤러를 대신해 개인정보 처리를 수행하도록 맡기는 제3자입니다. LLM API에서는 GPU 호스트, 파트너 추론 제공업체, 백업 용량이 하위 처리자가 될 수 있습니다. GDPR 제28조에 따라 컨트롤러의 승인이 필요하며 같은 데이터 보호 의무를 집니다. - [시스템 프롬프트(system prompt)](https://broadrouter.com/ko/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): 시스템 프롬프트는 애플리케이션이 대화보다 앞서 언어 모델에 보내는 지시문으로, 보통 system 또는 developer 역할의 메시지로 전달됩니다. 모델의 역할, 규칙, 어조, 출력 형식을 정하며, 모델은 이와 충돌하는 사용자 메시지보다 시스템 프롬프트를 따르도록 되어 있습니다. - [타 마르부타(taa marbuta, ة)](https://broadrouter.com/ko/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): 타 마르부타(ة, U+0629)는 단어 끝에만 나타나는 아랍어 글자로, 여성 명사와 형용사를 표시하는 경우가 많습니다. 점 두 개가 붙은 헤(ه, U+0647)처럼 생겨서 입력된 텍스트에서 둘이 섞이며, 많은 검색 정규화 도구가 ة를 ه로 바꿉니다. - [Temperature(온도)](https://broadrouter.com/ko/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): Temperature(온도)는 언어 모델 출력의 무작위성을 조절하는 샘플링 파라미터입니다. 모델은 토큰 점수를 확률로 바꾸기 전에 temperature로 나눕니다. 값이 낮으면 가장 가능성 높은 토큰이 압도적으로 선택되고, 값이 높으면 분포가 평평해져 가능성이 낮은 토큰도 더 자주 선택됩니다. - [음성 합성(TTS, text-to-speech)](https://broadrouter.com/ko/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): TTS(text-to-speech), 즉 음성 합성은 글로 쓴 텍스트를 음성 오디오로 변환합니다. TTS API는 텍스트, 목소리, MP3나 WAV 같은 출력 형식을 받아 생성된 음성을 반환합니다. 오디오를 텍스트로 바꾸는 음성 인식(STT)의 반대 방향입니다. - [처리량(throughput)](https://broadrouter.com/ko/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): LLM 서빙에서 처리량(throughput)은 시스템이 모든 동시 요청에 걸쳐 단위 시간당 끝내는 전체 작업량으로, 보통 초당 출력 토큰 수나 초당 요청 수로 측정합니다. 요청 하나가 토큰을 받는 속도인 스트림당 속도와는 다른 개념입니다. - [첫 토큰까지의 시간(TTFT)](https://broadrouter.com/ko/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): 첫 토큰까지의 시간(TTFT)은 언어 모델에 요청을 보낸 시점부터 첫 출력 토큰이 도착할 때까지의 시간입니다. 네트워크 시간, 서버의 대기열 시간, 그리고 모델이 프롬프트 전체를 읽는 단계인 프리필(prefill)이 포함되므로, 프롬프트가 길고 부하가 클수록 늘어납니다. - [타임스탬프(음성 인식, ASR timestamps)](https://broadrouter.com/ko/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): 음성 인식의 타임스탬프는 전사본의 각 부분에 붙는 초 단위 시작·종료 시간입니다. 세그먼트 타임스탬프는 구절이나 문장 단위이고, 단어 타임스탬프는 단어마다 붙습니다. 자막, 녹음 내 검색, 텍스트와 화자 정렬에 필요합니다. - [토큰(token)](https://broadrouter.com/ko/glossary/token/) (translation of https://broadrouter.com/glossary/token/): 토큰은 언어 모델이 읽고 쓰는 텍스트 단위로, 모델의 토크나이저가 만들어 내는 단어, 단어의 일부, 문장 부호 또는 바이트 시퀀스입니다. LLM API는 사용량을 토큰 단위로 집계하고 제한하며, 대개 토큰 단위로 과금합니다. - [토크나이저(tokenizer)](https://broadrouter.com/ko/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): 토크나이저(tokenizer)는 텍스트를 언어 모델이 읽는 정수 토큰 ID 시퀀스로 변환하고, 생성된 ID를 다시 텍스트로 바꾸는 구성 요소입니다. 모델마다 자체 토크나이저가 있으므로, 같은 텍스트도 모델에 따라 토큰 수가 달라질 수 있습니다. - [초당 토큰 수(TPS, tokens per second)](https://broadrouter.com/ko/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): 초당 토큰 수(TPS)는 언어 모델이 출력 토큰을 생성하는 속도로, 보통 요청 하나에 대해 첫 토큰이 도착한 뒤부터 잽니다. 응답이 시작된 뒤 얼마나 빨리 스트리밍되는지를 결정합니다. 모든 동시 요청을 합하면 시스템 처리량이 되며, 첫 토큰 전의 대기 시간은 따로 측정합니다. - [도구 호출(tool calling)](https://broadrouter.com/ko/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): 도구 호출(tool calling)은 함수 호출(function calling)이라고도 하며, 모델이 개발자가 정의한 함수를 실행해 달라는 구조화된 요청을 JSON 인자와 함께 반환하는 LLM API 기능입니다. 함수는 애플리케이션이 실행하고 그 결과를 다시 보냅니다. 모델이 직접 코드를 실행하지는 않습니다. - [tool_choice(도구 선택 파라미터)](https://broadrouter.com/ko/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice는 도구 사용을 제어하는 요청 파라미터입니다. OpenAI Chat Completions에서는 auto(모델이 결정), required(도구를 하나 이상 반드시 호출), none(도구 사용 안 함), 또는 강제할 함수 이름을 지정할 수 있습니다. Anthropic의 대응 옵션은 auto, any, tool, none입니다. - [Top-p(뉴클리어스 샘플링)](https://broadrouter.com/ko/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Top-p(뉴클리어스 샘플링, nucleus sampling)는 언어 모델이 가능성이 가장 높은 토큰부터 모아 확률 합이 0.9 같은 임곗값 p에 도달하는 가장 작은 집합에서만 다음 토큰을 샘플링하고, 가능성이 낮은 나머지 꼬리는 버리는 디코딩 방식입니다. - [유니코드 정규화(Unicode normalization)](https://broadrouter.com/ko/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): 유니코드 정규화는 텍스트를 표준 형식으로 바꿔, 똑같아 보이는 문자열이 같은 코드 포인트를 쓰게 하는 작업입니다. 형식은 NFD, NFC, NFKD, NFKC 네 가지이며, K 형식은 아랍어 표시형(presentation form)이나 합자 같은 호환 문자도 일반 글자로 바꿉니다. - [음성 활동 감지(VAD)](https://broadrouter.com/ko/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): 음성 활동 감지(VAD)는 오디오를 짧은 프레임으로 나눠 각 프레임이 음성인지 아닌지 분류하는 신호 처리 단계입니다. 음성 시스템은 이를 이용해 무음을 건너뛰고, 긴 녹음을 쉬는 지점에서 나누고, 화자가 말을 마쳤는지 감지하고, 배경 소음을 전사하지 않도록 합니다. - [벡터 데이터베이스(vector database)](https://broadrouter.com/ko/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): 벡터 데이터베이스는 임베딩 벡터를 원본 텍스트, 메타데이터와 함께 저장하고 최근접 이웃 쿼리에 답하는 데이터 저장소입니다. 쿼리 벡터가 주어지면 저장된 벡터 중 가장 비슷한 것을 돌려주며, 보통 코사인 거리, 유클리드 거리, 내적으로 유사도를 잽니다. - [비전-언어 모델(VLM)](https://broadrouter.com/ko/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): 비전-언어 모델(VLM)은 이미지와 텍스트를 함께 입력받아 텍스트를 출력하는 모델입니다. 그래서 그림을 설명하고, 그림에 관한 질문에 답하고, 문서, 차트, 스크린샷을 읽을 수 있습니다. VLM은 image-text-to-text 모델 또는 멀티모달 모델이라고도 합니다. - [vLLM](https://broadrouter.com/ko/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM은 UC 버클리 Sky Computing Lab에서 처음 개발된 LLM 추론·서빙용 오픈소스 라이브러리입니다. PagedAttention으로 KV 캐시 메모리를 관리하고, 들어오는 요청을 연속적으로 배칭하며, 다양한 양자화 형식을 지원하고, OpenAI 호환 HTTP 서버를 제공합니다. - [어휘 크기(vocabulary size)](https://broadrouter.com/ko/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): 어휘 크기(vocabulary size)는 언어 모델 토크나이저에 들어 있는 서로 다른 토큰의 수로, 토큰마다 고유한 ID와 임베딩이 있습니다. 어휘가 크면 흔한 텍스트를 더 적은 토큰으로 인코딩할 수 있지만, 임베딩 층과 출력 층이 커지고, 어휘 학습에 쓰인 종류의 텍스트에서만 효과가 있습니다. - [WAV](https://broadrouter.com/ko/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV(Waveform Audio File Format, WAVE)는 Microsoft와 IBM의 RIFF 형식을 바탕으로 한 오디오 컨테이너입니다. 보통 비압축 선형 PCM 샘플을 담기 때문에 크기를 예측하기 쉽습니다. 초당 크기는 샘플링 레이트 × 샘플당 바이트 수 × 채널 수입니다. - [단어 오류율(WER)](https://broadrouter.com/ko/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): 단어 오류율(WER)은 음성 인식의 표준 정확도 지표입니다. 전사 결과를 올바른 정답 텍스트로 바꾸는 데 필요한 단어 치환, 삭제, 삽입 횟수를 정답 텍스트의 단어 수로 나눈 값이며, 낮을수록 좋습니다. - [데이터 비보관(Zero data retention, ZDR)](https://broadrouter.com/ko/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): 데이터 비보관(ZDR)은 AI 제공업체가 프롬프트와 출력을 요청 처리에만 사용하고, 처리 후에는 그 내용을 로그, 캐시, 학습 데이터 어디에도 저장하지 않는 데이터 처리 정책입니다. 토큰 수나 처리 시간처럼 내용이 없는 기록은 보통 계속 보관됩니다. ### Russian (ru, Русский): 294 translated pages - [Articles in Russian](https://broadrouter.com/ru/blog/) - [Цены на LLM API в октябре 2026: 3 повышения и ни одного снижения](https://broadrouter.com/ru/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 изменения на страницах цен 5 из 7 провайдеров LLM API в октябре 2026: 3 повышения, ни одного снижения, 9 новых моделей с ценой и 11 удалённых. Ссылки на архив. - [Сколько стоит один запуск ИИ-агента для кода: шаги, контекст, кэш](https://broadrouter.com/ru/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): Почему каждый шаг агента для программирования дороже предыдущего, как кэширование меняет итог и когда сессия с фиксированной ценой выгоднее оплаты за токены. Расчёт. - [Огласовки в арабском тексте через LLM: промпт, проверка, DER](https://broadrouter.com/ru/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): Как расставить арабские огласовки (ташкиль) через LLM API: промпт без правки букв, проверка ответа, оценка по DER и WER и когда лучше взять CATT. - [Проверка арабской орфографии и грамматики через LLM: руководство](https://broadrouter.com/ru/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Как исправлять арабскую орфографию и грамматику через LLM API: промпт с минимальными правками, защита от переписывания, классы ошибок и оценка в стиле M2. - [Протокол арабской встречи по записи: конвейер на API](https://broadrouter.com/ru/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Как превратить запись встречи на арабском в протокол с решениями и задачами: распознать речь, очистить, свести через LLM в JSON и проверить имена и даты. - [Извлечение именованных сущностей из арабского с LLM: JSON и даты](https://broadrouter.com/ru/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Извлечение арабских имён, организаций, мест и дат по хиджре в JSON с помощью LLM: промпт, проверка спанов без учёта огласовок, перевод дат и скрипт оценки. - [Сравнение API для OCR арабского: Google, Azure, Textract, Mistral, LLM](https://broadrouter.com/ru/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): Какие OCR API официально поддерживают арабский, печатный и рукописный, что каждый возвращает и сколько стоит 1000 страниц, в сравнении с vision-LLM API. - [Арабские субтитры (SRT) из API распознавания речи: что нужно](https://broadrouter.com/ru/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): Что нужно файлу SRT, какие API распознавания речи возвращают для арабского временные метки или SRT, и проверенный конвейер субтитров из API, отдающего только текст. - [Классификация арабского текста с LLM: тональность, интенты, тикеты](https://broadrouter.com/ru/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): Классификация арабских отзывов, сообщений и обращений через LLM API: метки, промпт с учётом диалектов, вывод в JSON, проверка по macro F1 и цена на объёме. - [Нормализация арабского текста для поиска и RAG: что менять, а что нет](https://broadrouter.com/ru/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Ташкиль, формы алифа, та марбута, хамза, персидские буквы, цифры и Unicode: какие правила нормализации арабского помогают поиску, какие вредят, и код на Python. - [Суммаризация арабского текста с LLM: промпты, длина и оценка](https://broadrouter.com/ru/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): Как кратко пересказать арабские статьи, отчёты и диалектные сообщения через LLM API: промпты, контроль длины, проверка чисел и почему ROUGE даёт арабскому 0. - [Альтернативы Artificial Analysis: где сравнивать LLM в 2026 году](https://broadrouter.com/ru/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Сравнение альтернатив Artificial Analysis: Arena, Epoch AI, Vals AI, рейтинги OpenRouter, Hugging Face и данные о ценах — что измеряет каждый сайт и когда он нужен. - [GPU-парк и подход к инференсу BroadNet — по описанию самой BroadNet](https://broadrouter.com/ru/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): Что BroadNet рассказывает о железе за BroadNet Inference: три уровня GPU от B300 до устройств у клиента, пропускная способность на GPU и уроки её тестов. - [Кэширование ответов LLM API: кэш точных совпадений на Python](https://broadrouter.com/ru/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Проверенный кэш точных совпадений для ответов LLM API на Python: что входит в ключ, что нельзя кэшировать, TTL, приватность и как заранее измерить долю попаданий. - [Claude Code на открытых моделях: прокси, нативные эндпоинты, лимиты](https://broadrouter.com/ru/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code работает через Anthropic Messages API. Как запустить его на открытых моделях через прокси LiteLLM или нативный эндпоинт и что при этом ломается. - [Ручная проверка в распознавании документов с помощью ИИ](https://broadrouter.com/ru/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): Как устроить ручную проверку в конвейере распознавания документов: правила валидации, что отдавать человеку, случайный аудит, журнал аудита и доля проверки. - [Закон Египта о защите данных (№ 151 от 2020 года) и LLM API](https://broadrouter.com/ru/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Персональные данные египтян в зарубежном LLM API или API распознавания речи: что говорят Закон № 151 и Решение № 816 о передаче, лицензиях и утечках. - [Перед дообучением LLM: сформулируйте, какую проблему решаете](https://broadrouter.com/ru/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): Когда проблему решают промпт, RAG или оценка и дообучение не нужно, какие данные нужны для обучения и как проверить fine-tune, прежде чем его заказывать. - [Как исправить невалидный JSON от LLM: причины, починка, повтор](https://broadrouter.com/ru/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): Почему LLM возвращают сломанный JSON (блоки кода, лишний текст, обрезка по max_tokens, висячие запятые) и проверенная функция на Python, чинящая только безопасное. - [Как часто меняются цены на LLM API? 3 года архивных страниц цен](https://broadrouter.com/ru/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): Мы прочитали 308 архивных страниц цен 11 провайдеров LLM API за 2023–2026 годы. 80% моделей ни разу не меняли цену; 75% всех изменений сделала DeepInfra. - [Как снизить расходы на LLM API: чек-лист из 10 шагов по порядку](https://broadrouter.com/ru/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): Сокращайте счёт за LLM API в правильном порядке: замер, кэш префикса, сокращение промптов, лимит выхода и рассуждений, выбор модели, batch и фиксированная цена. - [Отказоустойчивость LLM API: резервный провайдер в вашем коде](https://broadrouter.com/ru/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): Как переключаться между OpenAI-совместимыми LLM-провайдерами: на каких ошибках, circuit breaker, ID моделей, стриминг, цена и равные условия приватности. - [Снижения цен на LLM API в 2023–2026 годах: крупнейшие, с источниками](https://broadrouter.com/ru/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): Крупнейшие снижения цен на LLM API с 2023 по 2026 год со ссылками на архив до и после: o3, Codestral, Gemini 1.5, а также повышения, которые мало кто заметил. - [Цены на LLM API только падают? 52 повышения и как учесть их в бюджете](https://broadrouter.com/ru/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): Цены на LLM API растут: в архивах прайс-листов за 2024–2026 годы мы нашли 52 повышения цены на ту же модель, медиана 40%. Все источники и правило для бюджета. - [Лимиты запросов LLM API у провайдеров: RPM, TPM и параллельность](https://broadrouter.com/ru/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): Лимиты запросов LLM API на начальных уровнях 13 провайдеров по их документации: OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks и другие. Что такое RPM и TPM. - [SLA LLM API: что обещают Azure, Bedrock, Google и OpenAI](https://broadrouter.com/ru/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): Что каждый провайдер LLM API обещает письменно: процент аптайма, что считается простоем, сервисные кредиты и цели по задержке — по страницам SLA. - [Пакетная обработка LLM без Batch API: возобновляемый раннер](https://broadrouter.com/ru/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Как прогнать от 10 000 до 1 000 000 промптов через chat completions без Batch API: лимит параллельности, темп запросов, повторы, чекпойнты и расчёт времени. - [Safety Map: как 39 LLM обращаются с конфиденциальными данными](https://broadrouter.com/ru/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet проверила 39 LLM от 14 лабораторий на учётных данных, ПДн, медданных и финансах: что переносится, что нет, утечки через инструменты и рассуждения. - [Как проверить LLM на утечку конфиденциальных данных: метод BroadNet](https://broadrouter.com/ru/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): Метод BroadNet: как до развёртывания проверить, не сливает ли LLM учётные данные, ПДн, медицинские или финансовые данные. Категории, утечки, 3+ прогона. - [Переход с OpenAI API: разбор кода на Python, Node и cURL](https://broadrouter.com/ru/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): Какие изменения в коде нужны для перехода с OpenAI на OpenAI-совместимый API: клиент, Responses → Chat Completions, инструменты, стриминг, эмбеддинги и smoke-тест. - [Какие модели BroadNet протестировала и отбраковала и почему](https://broadrouter.com/ru/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet публикует отвергнутые модели с причинами: чат-модели, не прошедшие проверку на утечку учётных данных, и проигравшие модели речи, OCR и эмбеддингов. - [Как спроектировать многоязычный ИИ-продукт](https://broadrouter.com/ru/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): Проектирование ИИ-продукта для арабского, английского и испанского: локаль интерфейса и язык модели, RTL и смешанный текст, числа, даты, определение языка, цена. - [Открытые или закрытые модели: стоимость API по данным о ценах](https://broadrouter.com/ru/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): Дешевле ли модели с открытыми весами через API, чем GPT, Claude или Gemini? Цены за токен со страниц провайдеров, разброс между хостами и расчёт на пользователя. - [История цен API OpenAI, Anthropic и Google, 2023–2026](https://broadrouter.com/ru/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): Цены API GPT, Claude и Gemini рядом по архивным страницам цен за 2023–2026 годы: какие линейки подешевели, какие подорожали, у каждой цены — ссылка на архив. - [Базовые URL OpenAI-совместимых API: 14 провайдеров с проверкой](https://broadrouter.com/ru/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): base_url для 14 OpenAI-совместимых LLM API по документации самих провайдеров: какие эндпоинты они перечисляют и какие особенности мешают простой замене. - [OpenClaw и Hermes Agent на дешёвой модели: настройка и расчёт токенов](https://broadrouter.com/ru/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): Подключите OpenClaw или Hermes Agent к OpenAI-совместимому эндпоинту, поймите, почему проверки жгут токены, и снизьте счёт лёгким контекстом и дешёвой моделью. - [Приватный эндпоинт начинается с чётко описанной нагрузки](https://broadrouter.com/ru/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): Что определить, прежде чем запрашивать приватный эндпоинт LLM: нагрузка, токены в секунду, параллельность, контекст, регион, лицензия и бюджет — и как его проверить. - [Промпт-инъекции в ИИ-агентах для программирования: рабочая защита](https://broadrouter.com/ru/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): Где промпт-инъекция добирается до агента для программирования и какую защиту описывают вендоры и исследователи: песочница, сеть, белые списки, одобрение. - [Размер чанка и перекрытие в RAG: как они влияют на поиск и цену](https://broadrouter.com/ru/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): Как размер чанка и перекрытие меняют полноту RAG, размер индекса, стоимость эмбеддингов и число токенов в каждом запросе к LLM: опубликованные тесты и расчёты. - [max_tokens для моделей с рассуждениями: как рассчитать бюджет](https://broadrouter.com/ru/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Почему модели с рассуждениями возвращают пустой ответ при низком max_tokens и как подобрать бюджет по перцентилям, контекстному окну и повтору с ростом лимита. - [Запись аудио в браузере и отправка в API распознавания речи](https://broadrouter.com/ru/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): Запишите микрофон через MediaRecorder, переведите WebM или MP4 из Safari в WAV 16 кГц и отправьте в API транскрибации через небольшой сервер, где хранится ключ. - [Сокращаем токены промпта: замеры для инструкций, инструментов, данных](https://broadrouter.com/ru/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): Измеренная экономия токенов: системный промпт, few-shot, схемы инструментов, форматы данных, история чата и арабская огласовка на реальных токенизаторах. - [Одна модель — разные провайдеры: насколько отличаются цены LLM API](https://broadrouter.com/ru/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): У 119 моделей, которые продают несколько хостов, самая высокая прайсовая цена в медиане в 1,76 раза выше самой низкой. Данные 41 провайдера, ноябрь 2026 г. - [PDPL Саудовской Аравии и ОАЭ и LLM API: чек-лист передачи данных](https://broadrouter.com/ru/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Персональные данные в LLM API из Саудовской Аравии или ОАЭ: что говорят правила PDPL о передаче, что спросить у провайдера и когда важен инференс внутри страны. - [Текст, речь или изображения? Начните с задачи](https://broadrouter.com/ru/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): Как выбрать текстовый чат, ввод изображений или транскрибацию аудио по задаче — с моделями и лимитами BroadNet и тем, чего у неё нет, например TTS. - [Транскрибация аудио больше 25 МБ: разрезать, отправить, склеить](https://broadrouter.com/ru/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): Как расшифровать запись больше лимита API распознавания речи: уменьшить файл, разрезать по паузам через ffmpeg, отправить части параллельно и склеить текст. - [Перевод строк приложения на арабский через LLM: плейсхолдеры и плюралы](https://broadrouter.com/ru/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): Как перевести i18n JSON на арабский через LLM API: заморозить плейсхолдеры, получить все шесть форм числа, проверять синтаксис ICU в CI и изолировать имена. - [Перевод арабского аудио в английский текст по API: рабочий пайплайн](https://broadrouter.com/ru/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): Арабские записи в английский текст за два вызова API: расшифровка с language=ar, затем перевод через LLM, с проверкой чисел, нарезкой, ценой и оценкой качества. - [Live, по запросу, исследования: как читать каталог моделей](https://broadrouter.com/ru/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): Как понять, какие модели из каталога можно вызвать сегодня, какие требуют запроса, а какие только тестировались — на примере каталога BroadNet и /v1/models. - [Uno в vLLM: что измерило исследование инференса BroadNet](https://broadrouter.com/ru/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet встроила самодрафтовый адаптер Uno в vLLM и измерила его на H100: 2,58x пропускной способности на одном потоке, 1,74x на 32. Метод, DOI и ограничения. - [Uno на Gemma 4 26B: исследование BroadNet по mixture-of-experts](https://broadrouter.com/ru/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet обучила первый самодрафтовый адаптер Uno для MoE-модели Gemma 4 26B A4B. Что измерено, когда самодрафтинг окупается, и DOI препринта. - [Бюджет токенов агента: куда уходят токены в цикле агента](https://broadrouter.com/ru/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): Системный промпт, схемы инструментов, чтение файлов, история и рассуждения: как цикл агента тратит токены, как измерить каждую часть по usage и что сокращать первым. - [WER для арабского: что такое доля ошибок в словах в ASR](https://broadrouter.com/ru/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): Что такое WER, как его считать и почему для арабского он так чувствителен к огласовкам, нормализации написания и диалектам. С кодом на Python и реальными числами. - [Распознавание арабских диалектов: Залив, Египет, Левант, Магриб](https://broadrouter.com/ru/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Почему распознавание речи хуже справляется с арабским языком Залива, Египта, Леванта и Магриба и как собрать тестовый набор по диалектам на своём аудио. - [Арабско-английский перевод через LLM: промпты, глоссарии, оценка](https://broadrouter.com/ru/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): Перевод с арабского на английский и обратно через LLM API: промпты, сохраняющие тон, глоссарии терминов, числа и имена, и как оценивать качество перевода. - [Арабский Залива, египетский и левантийский: как LLM понимают диалекты](https://broadrouter.com/ru/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): Как LLM справляются с арабским Залива, египетским и левантийским: где ошибаются, как писать промпт под диалект, как тестировать и что BroadNet пишет о Qwen3.8 27B. - [OCR арабского или vision-LLM: как читать арабские документы и счета](https://broadrouter.com/ru/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): Когда для арабских счетов, чеков и форм брать классический OCR, а когда vision-LLM: ловушки точности, вывод в JSON, цена за страницу и пайплайн проверки. - [Промпт-инжиниринг для арабского: регистр, RTL, цифры и смешанный текст](https://broadrouter.com/ru/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): Как писать промпты, чтобы LLM выдавала хороший арабский: регистр и диалект, цифры и даты, чистое RTL-форматирование и смешение арабского с английским. - [RAG на арабском: нормализация, чанкинг и поиск, который работает](https://broadrouter.com/ru/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): RAG по арабским документам: нормализуйте алиф, йа и та марбута, разумно делите текст на чанки, берите мультиязычные эмбеддинги и добавьте поиск по словам. - [API распознавания арабской речи: как выбрать в 2026 году](https://broadrouter.com/ru/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): Как выбрать API для транскрибации арабского: WER, диалекты, лимиты на файлы, цена за час аудио и конфиденциальность, плюс план теста на ваших записях. - [Архитектура арабского чат-бота WhatsApp: от голосовых к ответам](https://broadrouter.com/ru/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): Рабочая архитектура арабского ассистента в WhatsApp: вебхук, очередь, расшифровка голосовых, ответ LLM, правила политики и затраты, с кодом на Python. - [Арабизи и LLM: как работать с франко-арабским и цифрами 3 и 7](https://broadrouter.com/ru/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): Арабизи — арабский латиницей с цифрами вроде 3 для ع и 7 для ح. Как его распознать, перевести в арабскую графику с помощью LLM и правильно ответить. - [Лучшая LLM для Aider: как выбрать модель и подключить её](https://broadrouter.com/ru/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Как выбрать лучшую LLM для Aider: формат правок, размер контекста и цена часа работы, часовой тест на своём репозитории и точная настройка OPENAI_API_BASE. - [Лучшая открытая модель для агентов программирования: как выбрать](https://broadrouter.com/ru/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B или Gemma 4 для агента программирования? Выбор по вызову инструментов, контексту, рассуждениям и цене задачи, с расчётом. - [Кэшированные входные токены: как снизить счёт за LLM API](https://broadrouter.com/ru/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): Вход из кэша стоит лишь часть цены обычного входа. Как тарифицируется кэширование промптов, где смотреть cached_tokens и как строить промпт под кэш. - [Самый дешёвый vision LLM API: цена за изображение и за страницу](https://broadrouter.com/ru/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): Сколько стоит вызов vision LLM за изображение и за страницу документа: расчёт токенов для Gemma 4, Qwen3.8 27B и DeepSeek V4.1 Flash и где DeepInfra дешевле. - [Как выбрать провайдера LLM API: чек-лист из 20 пунктов](https://broadrouter.com/ru/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): Двадцать вопросов перед выбором LLM API: цены, лимиты, совместимость с OpenAI, приватность, модели, поддержка и аптайм, плюс таблица оценки и план теста. - [Как выбрать LLM API для арабского: что проверить](https://broadrouter.com/ru/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): Как выбрать LLM для арабских приложений: понимание диалектов, качество текста, перевод, арабские документы и изображения, цена и тестовый набор за один день. - [Cline, Roo Code или Aider для команды на открытых моделях (2026)](https://broadrouter.com/ru/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): Cline, Roo Code (в архиве, продолжение — Zoo Code) и Aider: что важно с открытыми моделями — вызов инструментов, контекст, MCP, git и подтверждения. - [Как посчитать токены до отправки запроса к LLM](https://broadrouter.com/ru/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): Как оценить число токенов в промпте до вызова LLM API: почему токенизаторы считают по-разному, замеры на пяти языках и безопасный запас для max_tokens. - [Как снизить расходы агента для программирования на LLM: 10 способов](https://broadrouter.com/ru/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Десять конкретных способов сократить расходы Cline, Aider, opencode или собственного агента для программирования на токены: расчёт экономии и как её измерить. - [Альтернативы DeepSeek API: 5 способов работать с моделями DeepSeek](https://broadrouter.com/ru/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): Модели DeepSeek через официальный API, агрегатор, каталог облака, свои GPU или BroadNet Inference: плюсы и минусы по цене, конфиденциальности и лимитам. - [Цены DeepSeek V4.1 Flash API: за токены, вход из кэша и тарифы](https://broadrouter.com/ru/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): Цена DeepSeek V4.1 Flash API за 1 млн входных, кэшированных и выходных токенов, тарифы на месяц, неделю и час и сравнение с DeepSeek, DeepInfra и Fireworks. - [Как рассчитать стоимость LLM API до запуска функции](https://broadrouter.com/ru/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): Пошаговая модель затрат на новую ИИ-функцию: пользователи × действия × вызовы × токены, с кэшем, рассуждениями, повторами, лимитами и расчётом на месяц. - [Как оценить LLM перед сменой провайдера](https://broadrouter.com/ru/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): Соберите небольшой eval-набор из реального трафика, оцените его проверками в коде и LLM-судьёй без перекосов и сравните качество, цену и задержку до миграции. - [Распознавание счетов, чеков и форм в JSON с помощью vision LLM](https://broadrouter.com/ru/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): Рабочий конвейер для счетов, чеков и форм: подготовка изображений, промпт со схемой JSON, проверка Pydantic, арифметические проверки, повторы и ручная проверка. - [Фиксированный тариф или оплата за токены в LLM API: когда выгоднее](https://broadrouter.com/ru/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): Когда тариф LLM с фиксированной ценой выгоднее оплаты за токены? Четыре расчёта месячного счёта для DeepSeek V4.1 Flash, точка окупаемости и как измерить свою. - [Бесплатный LLM API: на что на самом деле хватит стартового гранта](https://broadrouter.com/ru/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): Что реально даёт бесплатный стартовый грант в LLM API: токены, минуты аудио и почасовые ключи в BroadNet в сравнении с бесплатными тарифами Gemini и OpenRouter. - [GDPR и LLM API: 15 вопросов, которые стоит задать провайдеру](https://broadrouter.com/ru/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): Чек-лист GDPR для LLM API: DPA, субпроцессоры, хранение, обучение, международная передача и уведомление об утечках, со статьями GDPR к каждому вопросу. - [GLM 5.3 или DeepSeek V4.1 Flash: что выбрать и когда](https://broadrouter.com/ru/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): GLM 5.3 и DeepSeek V4.1 Flash: цена, контекст, инструменты, структурированный вывод, изображения и рассуждения, расчёты нагрузок и простое правило выбора. - [Ошибка 429 в LLM API: экспоненциальная задержка, jitter и лимитеры](https://broadrouter.com/ru/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): Чем 429 отличается от 5xx, экспоненциальная задержка с full jitter, ограничение параллельности и очереди для LLM API, с кодом на Python и Node для продакшена. - [Почасовой ключ LLM API: фиксированная цена за час](https://broadrouter.com/ru/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): Почасовой ключ BroadNet даёт час Gemma 4 VL 26B в одном потоке по фиксированной цене. Что в него входит, его лимиты, кому он подходит и когда окупается. - [Хранение данных в LLM API: OpenAI, Anthropic, Gemini и Mistral](https://broadrouter.com/ru/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): Что OpenAI, Anthropic, Gemini API, Mistral и BroadNet официально говорят о хранении промптов из API и обучении на них. Цитаты с их страниц, ноябрь 2026 г. - [Безопасность API-ключей LLM: хранение, ротация и белые списки IP](https://broadrouter.com/ru/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): Как не допустить API-ключ LLM в код и логи, выдать каждой нагрузке свой ключ, ротировать по графику, привязать к IP серверов и отозвать утёкший ключ за секунды. - [Цены на LLM API: вход, кэш, выход и токены рассуждений](https://broadrouter.com/ru/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): Как устроена оплата LLM API за токены: вход, вход из кэша, выход, рассуждения, изображения и поминутное аудио — с примерами расчёта по реальному прайс-листу. - [Конфиденциальные данные в LLM API: маскирование, минимизация, логи](https://broadrouter.com/ru/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): Как безопасно отправлять в LLM API данные клиентов, медицинские и финансовые данные: классифицировать поля, маскировать обратимыми метками, ZDR и чистые логи. - [Контроль расходов на LLM API: бюджеты, лимиты ключей, защита от циклов](https://broadrouter.com/ru/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): Как не дать счёту за LLM API выйти из-под контроля: предоплаченный баланс, ключ на приложение, белые списки IP, логи, бюджеты в коде, лимиты циклов и тарифы. - [Логирование LLM API: стоимость по пользователям и функциям без ПДн](https://broadrouter.com/ru/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): Логируйте каждый вызов LLM API с токенами, стоимостью, задержкой и ID запроса, считайте расходы по пользователям и функциям, маскируйте ПДн до записи. Код на Python. - [Стоимость LLM на пользователя: юнит-экономика ИИ-функции в SaaS](https://broadrouter.com/ru/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): Как посчитать стоимость ИИ-функции на активного пользователя в месяц, почему тяжёлые пользователи важнее среднего и как построить тарифы SaaS, чтобы счёт окупался. - [Цены провайдеров инференса на открытые модели: сравнение 6 сервисов](https://broadrouter.com/ru/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): Как OpenRouter, Together AI, Fireworks, Groq, DeepInfra и BroadNet берут плату за открытые модели: токены, кэш, batch, аренда GPU и тарифы. Цены со ссылками. - [Задержка LLM: TTFT, токены в секунду и полное время ответа](https://broadrouter.com/ru/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): Что измеряют время до первого токена, токены в секунду и полная задержка, как замерить каждую метрику на клиенте с кодом и какие меры какую цифру улучшают. - [Стриминг LLM: чанки SSE, дельты, usage и отмена](https://broadrouter.com/ru/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Как устроен стриминг chat completions на уровне протокола, как собрать ответ из дельт, взять usage из последнего чанка и отменить поток в Python, Node и curl. - [Длинный контекст или RAG: когда 1 млн токенов заменяет векторную БД](https://broadrouter.com/ru/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): Когда контекстное окно на 1 млн токенов заменяет векторную базу данных, а когда RAG всё ещё выигрывает: сравнение по цене запроса, задержке, полноте и поддержке. - [Управление контекстом долгоживущих агентов с окном 1 млн токенов](https://broadrouter.com/ru/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): Как сохранить точность и низкую цену долгих сессий агента с окном 1 млн токенов: история только с дописыванием, промпты под кэш, когда сжимать и что убирать первым. - [Долгоживущие ИИ-агенты: таймауты, повторы и чекпойнты](https://broadrouter.com/ru/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): Как держать LLM-агента в работе часами: таймаут на вызов, стриминг для поиска зависаний, безопасные повторы, идемпотентные инструменты и чекпойнт после каждого шага. - [MCP-серверы с OpenAI-совместимыми моделями (у любого провайдера)](https://broadrouter.com/ru/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP работает с любой LLM: клиент превращает инструменты MCP в вызовы функций. Как подключить MCP к открытым моделям в Cline, Continue, opencode, n8n и своём коде. - [Когда контекстное окно на 1 млн токенов оправдано (а когда нет)](https://broadrouter.com/ru/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): Что помещается в 1 048 576 токенов, сколько это стоит на DeepSeek V4.1 Flash, как рассуждения и кэширование меняют расчёт и когда RAG всё ещё лучше. - [Сколько на самом деле стоят мультиагентные системы: токены или сессии](https://broadrouter.com/ru/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): Мультиагентные системы сжигают в разы больше токенов, чем один агент. Сессии тарифа против оплаты за токены, скрытые множители и что когда выгоднее. - [Многоязычный API распознавания речи: 14 языков, один эндпоинт](https://broadrouter.com/ru/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): Транскрибация английского, испанского, французского, китайского, арабского и ещё 9 языков через один OpenAI-совместимый эндпоинт. Когда брать арабскую модель. - [ИИ-агент n8n на дешёвой LLM: настройка, выбор модели и цена запуска](https://broadrouter.com/ru/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): AI Agent в n8n на недорогой OpenAI-совместимой модели: какую модель взять для инструментов, какие настройки экономят токены и расчёт на 10 000 запусков. - [Открытые или закрытые модели ИИ для бизнеса: цена, контроль, риски](https://broadrouter.com/ru/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): Открытые или закрытые LLM для бизнеса? Сравниваем цену, контроль, приватность, разрыв в качестве и привязку к поставщику, три способа запуска открытых моделей. - [Переход на OpenAI-совместимый API: чек-лист миграции](https://broadrouter.com/ru/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): Сменить base_url — самое простое. Ломаются ID моделей, токены рассуждений, usage в стриминге, ошибки, лимиты запросов и ключи. Чек-лист с примерами кода. - [OpenRouter или фиксированный тариф для ИИ-агентов: что дешевле?](https://broadrouter.com/ru/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter берёт за токены плюс комиссию за кредиты, фиксированный тариф — за сессию. Расчёт для агентов, которые работают весь день, комиссии и лимиты OpenRouter. - [Как безопасно запускать агентов для программирования на ночь](https://broadrouter.com/ru/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): Практичная схема автономных запусков агента для программирования: песочница, контрольные точки в git, файл задачи, лимит шагов, ревью утром и сессии тарифа. - [Как структурировать промпты, чтобы префикс кэшировался](https://broadrouter.com/ru/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): Как выстроить промпт LLM, чтобы кэшируемый префикс не менялся: системный промпт, схемы инструментов, документы, дописываемая история и тест-доказательство. - [Токены рассуждений: почему вы платите за выход, которого не видите](https://broadrouter.com/ru/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): Скрытые размышления моделей оплачиваются как выход. Как их измерить, сколько они стоят на каждой модели и как сократить: reasoning_effort, max_tokens, маршрутизация. - [Параллельные агенты для программирования на безлимитном LLM API](https://broadrouter.com/ru/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Субагенты, оркестраторы и пакетные воркеры на безлимитном DeepSeek V4.1 Flash: сколько сессий тарифа нужно, как распределить их по ключам и ограничить пересечения. - [Своя LLM на GPU или API: реальная стоимость и точка окупаемости](https://broadrouter.com/ru/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): Цены аренды GPU с официальных страниц, формула стоимости 1 млн токенов на своём GPU, загрузка, нужная, чтобы обогнать API, и забытые расходы на эксплуатацию. - [Цены API распознавания речи: сравнение стоимости минуты в 2026 году](https://broadrouter.com/ru/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): Цены за минуту и за 1000 часов у OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq и BroadNet Transcribe по официальным страницам, плюс доп. опции. - [JSON mode или json_schema: надёжный структурированный вывод LLM](https://broadrouter.com/ru/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON mode даёт валидный JSON, json_schema — нужную вам структуру. Как запросить оба режима у OpenAI-совместимых моделей, валидировать через Pydantic и чинить ответ. - [Токены на слово: арабский и английский — почему арабский дороже](https://broadrouter.com/ru/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): Почему для того же смысла арабский часто требует больше токенов LLM, чем английский, почему «токены на слово» вводят в заблуждение и скрипт для замера по usage. - [Вызов инструментов в открытых моделях: схемы, параллельность, повторы](https://broadrouter.com/ru/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): Function calling на OpenAI-совместимых открытых моделях: советы по схемам, параллельные вызовы, пять типичных сбоев и цикл на Python с валидацией и повторами. - [Транскрибация записей колл-центра: конвейер с поддержкой арабского](https://broadrouter.com/ru/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): Разделите стерео по каналам, нарежьте реплики, распознайте арабскую и английскую речь, скройте ПДн и получите сводку звонка от LLM. Код на Python и цены. - [Открытые модели в агентах для программирования: Aider, Cline, Codex](https://broadrouter.com/ru/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Как подключить Aider, Cline, Roo Code, opencode, Continue и Codex CLI к любому OpenAI-совместимому эндпоинту: точные названия полей и настройки, которые всё ломают. - [Голосовые в текст для бизнеса: из WhatsApp и Telegram в тикеты](https://broadrouter.com/ru/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): Как превращать голосовые клиентов из WhatsApp Business и Telegram-ботов в расшифровки и тикеты поддержки: скачать, транскрибировать, извлечь поля через LLM. - [Недельный тариф LLM для коротких проектов: Orbit, час или месяц](https://broadrouter.com/ru/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): Когда безлимитный DeepSeek V4.1 Flash на неделю выгоднее часовой сессии, месячного тарифа или оплаты за токены: расчёт для хакатонов, спринтов и миграций. - [Что на самом деле значит «безлимит» в тарифе LLM API](https://broadrouter.com/ru/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Счётчик токенов, скользящие 5-часовые окна, недельные лимиты, лимиты запросов и параллельность. Семь вопросов: безлимитен ли «безлимит» для вашего агента. - [Альтернатива Whisper API: когда переходить и как её протестировать](https://broadrouter.com/ru/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): Когда стоит искать замену OpenAI Whisper API: WER на вашем аудио, языки, цена за час, лимиты и приватность. С кодом для тестирования на своих записях. - [Нулевое хранение данных (ZDR) в LLM API: чек-лист из 10 вопросов](https://broadrouter.com/ru/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): Что должно означать «нулевое хранение данных» для LLM API: промпты, ответы, логи, обучение, субподрядчики и аудит. Десять вопросов до того, как отправлять данные. - [Guides in Russian](https://broadrouter.com/ru/guides/) - [OpenAI Python SDK и BroadNet Inference](https://broadrouter.com/ru/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): Как подключить официальный OpenAI Python SDK к BroadNet Inference, поменяв только base_url: установка, ключ, первый запрос и стриминг ответа. - [OpenAI Node.js SDK и BroadNet Inference](https://broadrouter.com/ru/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): Как подключить официальный OpenAI SDK для Node.js и TypeScript к BroadNet Inference через baseURL: установка, первый запрос, стриминг и ID запросов. - [BroadNet Inference в браузере: без регистрации](https://broadrouter.com/ru/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): Проверьте распознавание арабской речи, чтение документов и чат с Gemma 4 в браузерной песочнице BroadNet без регистрации и карты, а затем повторите запрос через API. - [Первый запрос к BroadNet Inference через curl](https://broadrouter.com/ru/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): Первые запросы к BroadNet Inference через curl: chat completion, стриминг ответа и транскрибация аудио. Как читать ответ, usage и ошибки. - [Стриминг chat completions и обрыв соединения](https://broadrouter.com/ru/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): Как работает стриминг в BroadNet Inference: Server-Sent Events, маркер [DONE], расход токенов в последнем чанке и что делать, если поток оборвался раньше. - [Почему рассуждающая модель возвращает пустой ответ](https://broadrouter.com/ru/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3 и Qwen3.8 27B рассуждают перед ответом, а рассуждения учитываются в max_tokens. Как перестать получать пустые ответы. - [Коды ошибок BroadNet Inference и когда повторять запрос](https://broadrouter.com/ru/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): Что означает каждый HTTP-статус API BroadNet Inference, какие ошибки повторять с экспоненциальной задержкой, какие сначала исправлять и что отправить в поддержку. - [Привязка API-ключей к серверам через белый список IP](https://broadrouter.com/ru/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): Защитите ключи BroadNet Inference белым списком IP, отделите ключи инференса от учётных данных портала и отзовите утёкший ключ за секунды. С примерами curl. - [Как устроены безлимитные тарифы LLM API для агентов](https://broadrouter.com/ru/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): У безлимитных тарифов DeepSeek V4.1 Flash нет счётчика токенов и недельного лимита. Разбираем, как на деле работают сессии, параллельные агенты, ключи и оплата. - [Aider на безлимитной DeepSeek V4.1 Flash](https://broadrouter.com/ru/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Как подключить агента Aider к BroadNet Inference через OpenAI-совместимую настройку и работать на безлимитном тарифе DeepSeek V4.1 Flash с фиксированной ценой. - [LangChain и BroadNet Inference](https://broadrouter.com/ru/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): Как подключить ChatOpenAI из LangChain к BroadNet Inference через base_url и api_key и что LangChain сохраняет, а что отбрасывает из ответов моделей с рассуждениями. - [Cline и BroadNet Inference: провайдер OpenAI Compatible](https://broadrouter.com/ru/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Как подключить расширение Cline для VS Code к BroadNet Inference через провайдер OpenAI Compatible: поля, ID моделей, размеры контекста и частые ошибки. - [Roo Code и BroadNet Inference: провайдер OpenAI Compatible](https://broadrouter.com/ru/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code закрыт в мае 2026 года. Настройки для существующих установок и форков и куда переходить дальше (Cline или форк Zoo Code) с BroadNet Inference. - [Continue и BroadNet Inference: config.yaml и apiBase](https://broadrouter.com/ru/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): Как добавить модели BroadNet Inference в расширение Continue для IDE через provider openai и apiBase в config.yaml, включая capabilities tool_use и длину контекста. - [BroadNet Inference в opencode: свой провайдер](https://broadrouter.com/ru/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): Как добавить BroadNet Inference в агента для программирования opencode как свой OpenAI-совместимый провайдер в opencode.jsonc: алиасы моделей, возможности и лимиты. - [Как подключить Open WebUI к BroadNet Inference](https://broadrouter.com/ru/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Как добавить BroadNet Inference в Open WebUI как OpenAI-совместимое подключение, внести ID моделей в список и исправить ошибки проверки, пустые ответы и 429. - [BroadNet Inference в LibreChat как custom endpoint](https://broadrouter.com/ru/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): Как добавить BroadNet Inference в LibreChat как OpenAI-совместимый custom endpoint в librechat.yaml: ключ в .env, список моделей, titleModel и типичные ошибки. - [BroadNet Inference в n8n: OpenAI credential и Base URL](https://broadrouter.com/ru/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): Как направить OpenAI credential в n8n на BroadNet Inference через поле Base URL и использовать узел OpenAI Chat Model по ID с выключенным Responses API. - [Vercel AI SDK и BroadNet Inference: OpenAI-совместимый провайдер](https://broadrouter.com/ru/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): Как вызывать BroadNet Inference из Vercel AI SDK через createOpenAICompatible из @ai-sdk/openai-compatible: generateText, streamText, инструменты и reasoning effort. - [API распознавания арабской речи: пошаговое руководство](https://broadrouter.com/ru/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): Транскрибация арабского аудио через BroadNet Transcribe Arabic и OpenAI-совместимый эндпоинт: форматы, поле language, примеры на Python и curl, цена. - [LiteLLM и BroadNet Inference: SDK и прокси](https://broadrouter.com/ru/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): Как вызывать BroadNet Inference через LiteLLM с префиксом модели openai/ и api_base — в Python SDK или в config.yaml прокси, с подходящими лимитами запросов. - [OpenAI Agents SDK на BroadNet Inference](https://broadrouter.com/ru/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Как запустить OpenAI Agents SDK для Python на BroadNet Inference через AsyncOpenAI и OpenAIChatCompletionsModel и отключить трейсинг, которому нужен ключ OpenAI. - [Голосовые сообщения WhatsApp в текст через API](https://broadrouter.com/ru/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): Конвертируйте голосовые WhatsApp в .ogg или записи телефона в .m4a в WAV через ffmpeg и расшифруйте их на арабском или ещё 13 языках одним вызовом API. - [Расшифровка голосовых сообщений WhatsApp в n8n](https://broadrouter.com/ru/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): Workflow n8n, который получает голосовое сообщение WhatsApp, скачивает его у Meta, конвертирует через ffmpeg и расшифровывает на арабском или ещё 13 языках. - [Как читать арабские счета и чеки vision-моделью](https://broadrouter.com/ru/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Отправьте фото счёта или чека vision-модели в BroadNet Inference и получите итог, дату и поставщика в JSON; вопросы можно задавать на арабском или английском. - [Zed и BroadNet Inference: OpenAI-совместимый провайдер](https://broadrouter.com/ru/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Как добавить BroadNet Inference в редактор Zed как OpenAI-совместимого провайдера через settings.json: ID моделей, лимиты токенов, capabilities и переменная ключа. - [Kilo Code и BroadNet Inference: провайдер OpenAI Compatible](https://broadrouter.com/ru/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Как подключить Kilo Code к BroadNet Inference как свой провайдер OpenAI Compatible: в диалоге VS Code или в kilo.json, с ID моделей, лимитами и флагами инструментов. - [BroadNet Inference в goose: пользовательский провайдер](https://broadrouter.com/ru/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): Как подключить ИИ-агента goose к BroadNet Inference как пользовательский провайдер OpenAI Compatible: через goose Desktop, goose configure или JSON-файл. - [Pydantic AI и BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/ru/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Запуск агентов Pydantic AI на BroadNet Inference через OpenAIChatModel и OpenAICompatibleProvider: вызов инструментов, структурированный вывод и профиль модели. - [CrewAI и BroadNet Inference как свой OpenAI-эндпоинт](https://broadrouter.com/ru/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): Как направить агентов CrewAI на BroadNet Inference через LLM(custom_openai=True, base_url=...) и выбрать модели для инструментов, длинного контекста и тарифа. - [LlamaIndex и BroadNet Inference через OpenAILike](https://broadrouter.com/ru/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): Как подключить LlamaIndex к BroadNet Inference через класс OpenAILike и правильно задать api_base, context_window, is_chat_model и is_function_calling_model. - [BroadNet Inference в Dify через плагин OpenAI-API-compatible](https://broadrouter.com/ru/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): Как подключить Dify к BroadNet Inference через плагин провайдера моделей OpenAI-API-compatible: точные значения полей, размеры контекста и настройки вызова функций. - [Jan и BroadNet Inference: свой OpenAI-совместимый провайдер](https://broadrouter.com/ru/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Как добавить BroadNet Inference в десктопное приложение Jan как свой OpenAI-совместимый провайдер, вписать ID моделей вручную и включить инструменты и vision. - [SillyTavern и BroadNet Inference: Custom (OpenAI-compatible)](https://broadrouter.com/ru/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): Как подключить SillyTavern к BroadNet Inference через источник Custom (OpenAI-compatible): нужные ID моделей, постобработка промпта и лимиты токенов. - [Spring AI и BroadNet Inference: OpenAI base-url](https://broadrouter.com/ru/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Как направить чат-стартер OpenAI из Spring AI на BroadNet Inference через spring.ai.openai.base-url, ID модели BroadNet и max-tokens и вызвать его из ChatClient. - [BroadNet Inference из Go: SDK openai-go](https://broadrouter.com/ru/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): Как подключить официальный SDK openai-go v3 к BroadNet Inference через option.WithBaseURL: chat completions, стриминг с usage и обработка ошибок на Go. - [Агенты LangGraph на BroadNet Inference](https://broadrouter.com/ru/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): Как запустить StateGraph или create_agent из LangGraph на BroadNet Inference через ChatOpenAI, base_url и ID модели BroadNet: инструменты, стриминг и память. - [Пайплайны и агенты Haystack на BroadNet Inference](https://broadrouter.com/ru/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Как направить OpenAIChatGenerator из Haystack на BroadNet Inference: api_base_url, ID модели и ключ через Secret — для пайплайнов, инструментов и агентов. - [Semantic Kernel и BroadNet Inference](https://broadrouter.com/ru/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Как подключить Microsoft Semantic Kernel к BroadNet Inference: в Python через клиент AsyncOpenAI, в C# через свой эндпоинт — для чата, плагинов и агентов. - [Агенты Mastra на BroadNet Inference](https://broadrouter.com/ru/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): Как запустить агентов Mastra на BroadNet Inference через объект model с id, url и apiKey и почему ID модели нужен двойной префикс broadnet/. - [Структурированный вывод через Instructor на BroadNet Inference](https://broadrouter.com/ru/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Как извлекать объекты Pydantic из BroadNet Inference с Instructor: from_openai или from_provider со своим base_url и правильный выбор Mode. - [DSPy и BroadNet Inference: укажите model_type="chat"](https://broadrouter.com/ru/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): Как настроить dspy.LM для BroadNet Inference с префиксом openai/, api_base и model_type="chat": DSPy 3.5 отправляет модели OpenAI в Responses API. - [Оценка моделей BroadNet в promptfoo](https://broadrouter.com/ru/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): Как запускать evals в promptfoo и сравнивать модели BroadNet Inference бок о бок через провайдер openai:chat, apiBaseUrl и apiKeyEnvar, а также с грейдерами. - [smolagents на BroadNet Inference через OpenAIModel](https://broadrouter.com/ru/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Как подключить smolagents от Hugging Face к BroadNet Inference через OpenAIModel и api_base: CodeAgent на любой модели, ToolCallingAgent на GLM 5.3 и Qwen3.8 27B. - [Flowise и BroadNet Inference: узел OpenAI Custom Model](https://broadrouter.com/ru/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Как подключить BroadNet Inference к чатфлоу Flowise через узел OpenAI Custom Model: Base Path, учётные данные OpenAI API и ID модели BroadNet вручную. - [Langflow и BroadNet Inference: компонент OpenAI](https://broadrouter.com/ru/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Как подключить BroadNet Inference в Langflow через поле OpenAI API Base компонента OpenAI и ID модели вручную и почему провайдер OpenAI Compatible его не покажет. - [Glossary entries in Russian](https://broadrouter.com/ru/glossary/) - [ИИ-агент (AI agent)](https://broadrouter.com/ru/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): ИИ-агент — это система, в которой языковая модель сама направляет свои шаги к цели: решает, какие инструменты вызвать, читает результаты и повторяет, пока не выдаст итоговый ответ или не упрётся в лимит. Фиксированный сценарий, который вызывает модель в заранее заданных местах, обычно называют не агентом, а workflow. - [API-ключ (API key)](https://broadrouter.com/ru/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): API-ключ — это секретная строка, которая идентифицирует и авторизует клиента, вызывающего API. LLM API обычно ожидают его в заголовке Authorization: Bearer, и любой, у кого есть ключ, может отправлять запросы, оплачиваемые аккаунтом владельца ключа. - [Арабские огласовки (ташкиль, tashkeel)](https://broadrouter.com/ru/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): Арабские огласовки, или ташкиль, — это небольшие знаки над буквами или под ними, которые обозначают краткие гласные, удвоенные согласные и отсутствие гласного. Они необязательны: в большинстве арабских текстов, включая новости, их нет, и читатель или программа восстанавливает гласные по контексту. - [Определение арабского диалекта (dialect identification)](https://broadrouter.com/ru/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): Определение арабского диалекта — это задача установить, на каком варианте арабского написан текст или сделана запись: египетском, диалектах Персидского залива, левантийском, магрибском или современном литературном арабском, на уровне региона, страны или города. Его используют для маршрутизации сообщений, выбора промптов и разбиения наборов для оценки. - [Нормализация арабского текста (Arabic text normalization)](https://broadrouter.com/ru/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): Нормализация арабского текста — это шаг, который перед сопоставлением приводит варианты написания одного слова к единой форме: удаляет огласовки и татвиль, унифицирует формы алифа и часто заменяет ى на ي и ة на ه. Её применяют к поисковым ключам и при оценке, а не к тексту, который показывают. - [Арабизи (Arabizi)](https://broadrouter.com/ru/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): Арабизи — это неформальная запись арабского латинскими буквами, где арабские звуки без латинского аналога передаются цифрами, например 3 вместо ع. Он распространён в чатах и соцсетях, в основном диалектный, не имеет стандартной орфографии и часто перемешан с английским или французским. - [AWQ (activation-aware weight quantization)](https://broadrouter.com/ru/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (activation-aware weight quantization) — это метод посттренировочной квантизации, который сжимает веса LLM до низкой точности, обычно до 4-битных целых. По статистике активаций на небольшом калибровочном наборе он находит самые важные каналы весов и масштабирует их, чтобы они меньше теряли в точности. Обратное распространение ему не нужно. - [Базовый URL (base URL)](https://broadrouter.com/ru/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): Базовый URL — это фиксированный корневой адрес API, например https://api.example.com/v1, который клиент подставляет перед каждым путём эндпоинта, например /chat/completions. В OpenAI-совместимых SDK смена базового URL вместе с API-ключом и ID модели направляет тот же код к другому провайдеру. - [Batch API (пакетная обработка запросов)](https://broadrouter.com/ru/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Batch API — это асинхронный эндпоинт LLM, который принимает сразу много запросов, обычно в виде файла, обрабатывает их не немедленно, а в пределах заданного окна, и возвращает результаты вместе. Такие провайдеры, как OpenAI и Anthropic, берут за пакетные запросы меньше, чем в своих стандартных API. - [BM25](https://broadrouter.com/ru/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) — это функция ранжирования для поиска по ключевым словам, которая оценивает документ по тому, как часто в нём встречаются слова запроса, насколько эти слова редки во всей коллекции и какова длина документа. Это сходство по умолчанию в Lucene и Elasticsearch. - [Кодирование пар байтов (byte pair encoding, BPE)](https://broadrouter.com/ru/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Кодирование пар байтов (BPE) — это алгоритм подсловной токенизации, который начинает с отдельных символов или байтов и раз за разом объединяет самую частую соседнюю пару в новую запись словаря, пока не достигнет заданного размера словаря. Большинство современных токенизаторов LLM используют BPE или его байтовый вариант. - [Вход из кэша (cached input tokens)](https://broadrouter.com/ru/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): Вход из кэша (cached input tokens) — это токены промпта, которые провайдер LLM недавно обработал и может взять из кэша, а не вычислять заново; обычно это повторяющийся префикс, например системный промпт или документ. Провайдеры тарифицируют их дешевле, чем новые входные токены. - [Цепочка рассуждений (chain-of-thought, CoT)](https://broadrouter.com/ru/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Цепочка рассуждений (chain-of-thought, CoT) — это техника промптинга, при которой языковой модели показывают или поручают записывать промежуточные шаги рассуждения перед окончательным ответом. Она повышает точность в многошаговых задачах, например в арифметике и логике, а модели рассуждений теперь генерируют её автоматически. - [Доля ошибок в символах (character error rate, CER)](https://broadrouter.com/ru/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Доля ошибок в символах (CER) — это мера точности транскрибации или OCR на уровне символов: число замен, удалений и вставок, нужных, чтобы превратить результат в эталонный текст, делённое на число символов в эталоне. Чем меньше, тем лучше; 0 означает полное совпадение. - [Chat Completions API](https://broadrouter.com/ru/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Chat Completions API — это HTTP-эндпоинт OpenAI, POST /v1/chat/completions, который принимает имя модели и список сообщений с ролями и возвращает следующее сообщение модели. Формат его запроса и ответа стал общим стандартом, который повторяют OpenAI-совместимые провайдеры и SDK. - [Шаблон чата (chat template)](https://broadrouter.com/ru/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): Шаблон чата — это правило форматирования, обычно написанное на Jinja и поставляемое вместе с токенизатором модели, которое превращает список сообщений чата с ролями в единую последовательность токенов, ожидаемую моделью, включая её специальные управляющие токены. Неверный шаблон может резко ухудшить ответы модели. - [Чанкинг (chunking) в RAG](https://broadrouter.com/ru/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Чанкинг — это шаг конвейера поиска, на котором документы разбиваются на небольшие фрагменты, чанки, перед векторизацией и индексацией. Размер чанка задаёт, сколько текста в каждом фрагменте, а перекрытие повторяет часть текста между соседними чанками, чтобы мысль, разрезанная границей, не потерялась. - [Переключение кодов (code-switching)](https://broadrouter.com/ru/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Переключение кодов — это использование двух или более языков либо вариантов одного языка в пределах одного разговора или предложения, устно или письменно. Например, арабоязычные люди вставляют английские слова в диалектную речь или переходят с современного литературного арабского на свой диалект и обратно. - [ИИ-агент для программирования (AI coding agent)](https://broadrouter.com/ru/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): ИИ-агент для программирования — это инструмент, в котором языковая модель работает с кодовой базой в цикле: читает файлы, правит их, запускает команды, например тесты, читает вывод и решает, каким будет следующий шаг, пока задача не выполнена или не достигнут лимит. - [Параллельность (concurrency)](https://broadrouter.com/ru/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): Параллельность (concurrency) — это число запросов, которые клиент одновременно держит в работе. LLM API часто ограничивают её на ключ или тариф отдельно от лимита запросов в минуту, потому что каждый открытый запрос занимает мощности GPU, пока ответ не будет готов. - [Сжатие контекста (context compaction)](https://broadrouter.com/ru/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): Сжатие контекста — это приём, при котором диалог, подходящий к пределу контекстного окна, суммаризируется, и сессия продолжается с этого резюме: агент сохраняет ключевые решения и состояние, а сырую историю, например старый вывод инструментов, отбрасывает. - [Контекстная инженерия (context engineering)](https://broadrouter.com/ru/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): Контекстная инженерия — это работа по отбору и поддержанию набора токенов, который языковая модель видит в каждом запросе: системного промпта, описаний инструментов, истории, найденных документов и заметок. Цель — дать модели минимальный набор самой ценной информации, нужной, чтобы хорошо выполнить задачу. - [Деградация контекста (context rot)](https://broadrouter.com/ru/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): Деградация контекста (context rot) — это постепенная потеря точности и надёжности языковых моделей по мере удлинения входа, даже если задача не меняется, а вход далеко не заполняет контекстное окно. Термин закрепился после технического отчёта Chroma 2025 года, в котором проверили 18 моделей. - [Контекстное окно (context window)](https://broadrouter.com/ru/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): Контекстное окно — это максимальное число токенов, с которым языковая модель может работать в одном запросе. В него входят промпт, история диалога, описания инструментов, изображения и сгенерированный ответ, включая токены рассуждений. Всё, что не помещается, приходится обрезать, сжимать в краткое изложение или делить на несколько запросов. - [Локализация данных (data residency)](https://broadrouter.com/ru/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): Локализация данных (data residency) — это требование или обязательство хранить и обрабатывать данные в определённой стране или регионе. Для LLM API она касается того, где обрабатываются промпты во время инференса и где лежит сохранённый контент; данные аккаунта и биллинга в неё часто не входят. - [Арабские диалекты (dialectal Arabic)](https://broadrouter.com/ru/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): Арабские диалекты — это повседневный разговорный арабский язык региона или города, например египетский, диалекты Персидского залива, левантийский или магрибский. Они отличаются от современного стандартного арабского звуками, словами и грамматикой, не имеют стандартной орфографии и всё чаще пишутся в интернете, поэтому языковым моделям с ними сложнее. - [Диаризация дикторов (speaker diarization)](https://broadrouter.com/ru/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): Диаризация дикторов — это задача разбиения аудиозаписи на сегменты по говорящим, то есть ответ на вопрос «кто и когда говорил». Каждый отрезок речи получает анонимную метку диктора, например SPEAKER_00, и обычно её совмещают с расшифровкой, чтобы показать, кто произнёс каждую реплику. - [Дистилляция знаний (knowledge distillation)](https://broadrouter.com/ru/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): Дистилляция, или дистилляция знаний, — это метод обучения, при котором меньшая модель-ученик учится воспроизводить выходы большей модели-учителя, например её распределения вероятностей или сгенерированные ответы. В итоге ученик сохраняет значительную часть качества учителя, а запуск обходится дешевле. - [Соглашение об обработке данных (DPA)](https://broadrouter.com/ru/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): Соглашение об обработке данных (DPA) — это договор между контролёром и обработчиком, который определяет, как обработчик обращается с персональными данными. Его требует статья 28(3) GDPR и перечисляет содержание: документированные инструкции, конфиденциальность, безопасность, правила для субобработчиков, удаление или возврат данных и аудиты. - [Эмбеддинги (embeddings)](https://broadrouter.com/ru/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): Эмбеддинги — это списки чисел с плавающей точкой (векторы), которые модель выдаёт для фрагмента текста, изображения или другого входа. Близкие по смыслу входы получают близкие векторы, поэтому расстояние между векторами позволяет ранжировать результаты поиска, группировать документы или находить похожие объекты. - [Eval (оценка LLM)](https://broadrouter.com/ru/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Eval (сокращение от evaluation, «оценка») — это воспроизводимый тест языковой модели или промпта: фиксированный набор входов, описание того, как выглядит хороший ответ, и оценщик (код, человек или другая модель), который выставляет оценку каждому ответу, чтобы версии можно было сравнивать. - [Экспоненциальная задержка (exponential backoff)](https://broadrouter.com/ru/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): Экспоненциальная задержка (exponential backoff) — это стратегия повторов, при которой клиент после каждой неудачной попытки ждёт дольше, обычно удваивая паузу до потолка, прежде чем попробовать снова. Случайный джиттер в каждой паузе разносит повторы множества клиентов, чтобы они не били в сервер одновременно. - [Few-shot промптинг (few-shot prompting)](https://broadrouter.com/ru/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): Few-shot промптинг — это приём, при котором в промпт перед реальным входом добавляют несколько готовых примеров «вход — выход», чтобы языковая модель по ним поняла задачу и формат. Веса модели не меняются; в zero-shot примеров нет, в one-shot — один. - [Дообучение (fine-tuning)](https://broadrouter.com/ru/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): Дообучение (fine-tuning) — это продолжение обучения предобученной языковой модели на меньшем наборе данных под конкретную задачу, например на примерах диалогов. Веса модели меняются, и она надёжнее выполняет эту задачу или держит стиль без длинных инструкций и примеров в каждом промпте. - [finish_reason (причина остановки)](https://broadrouter.com/ru/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason — это поле каждого варианта (choice) в ответе chat completion, которое сообщает, почему модель прекратила генерацию. В формате OpenAI значения такие: stop, length, tool_calls, content_filter и устаревшее function_call. В Messages API от Anthropic то же самое передаёт поле stop_reason. - [Безлимитный тариф LLM (flat-rate LLM plan)](https://broadrouter.com/ru/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): Безлимитный тариф LLM, или тариф с фиксированной ценой, — это доступ к API языковой модели за фиксированную плату за период, например месяц, вместо оплаты за токены. Обычно такие тарифы ограничивают что-то другое — параллельность, скорость, набор моделей или окна использования, поэтому ограничения важны не меньше цены. - [FP8](https://broadrouter.com/ru/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 — это 8-битный формат чисел с плавающей точкой для глубокого обучения, определённый в двух кодировках: E4M3 с 4 битами экспоненты и 3 битами мантиссы и E5M2 с 5 битами экспоненты и 2 битами мантиссы. По сравнению с 16-битными форматами он вдвое сокращает память и нативно поддерживается современными GPU. - [Штрафы за частоту и присутствие (frequency/presence penalty)](https://broadrouter.com/ru/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): Штрафы за частоту и присутствие — это параметры сэмплирования, которые снижают оценку токенов, уже встречавшихся в выводе. Штраф за присутствие применяется, как только токен появился хотя бы раз; штраф за частоту растёт с каждым повтором. В API OpenAI оба принимают значения от -2.0 до 2.0, по умолчанию 0. - [GDPR (Общий регламент по защите данных)](https://broadrouter.com/ru/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): GDPR (General Data Protection Regulation, Регламент (ЕС) 2016/679) — это закон ЕС об обработке персональных данных. Он применяется с 25 мая 2018 года и распространяется на организации, учреждённые в ЕС, и на некоторые за его пределами. Для LLM API он регулирует договоры с обработчиками, международную передачу данных и их минимизацию. - [GGUF](https://broadrouter.com/ru/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF — это бинарный формат файлов для хранения языковых моделей и инференса в средах на базе GGML, таких как llama.cpp. Один файл GGUF содержит веса, часто квантизированные, и метаданные для загрузки модели, включая токенизатор и параметры архитектуры; его можно отобразить в память (mmap) для быстрой загрузки. - [GPU-час (GPU-hour)](https://broadrouter.com/ru/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): GPU-час — это использование одного графического процессора в течение одного часа. Облачные провайдеры тарифицируют арендованные GPU-инстансы почасово или посекундно, а исследователи в GPU-часах указывают, сколько вычислений потребовала задача. Вы платите за время машины независимо от того, загружена она или нет. - [Заземление (grounding)](https://broadrouter.com/ru/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): Заземление (grounding) — это практика привязки ответа языковой модели к конкретным проверяемым источникам информации, например к переданным документам, результатам поиска или записям базы данных, с инструкцией отвечать на их основе. Заземлённые ответы можно проследить до источника, и в них реже встречаются выдуманные факты. - [Ограничители (guardrails)](https://broadrouter.com/ru/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): Ограничители (guardrails) — это проверки, которые приложение выполняет вокруг вызова языковой модели, чтобы удерживать её в заданных рамках. Входные проверяют то, что попадает в модель, выходные — то, что из неё выходит, а ограничители инструментов проверяют действия до их выполнения. Проваленная проверка блокирует, переписывает или помечает запрос. - [Галлюцинация (hallucination)](https://broadrouter.com/ru/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): Галлюцинация — это ответ языковой модели, который звучит правдоподобно, но ложен, выдуман или не подкреплён переданными ей данными: например, несуществующая ссылка на источник, неверное число или выдуманный метод API. Она возникает потому, что модель предсказывает вероятный текст, а не проверяет факты. - [Хамза (hamza)](https://broadrouter.com/ru/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): Хамза (ء) — арабская буква, обозначающая гортанную смычку. Она пишется отдельно или на подставке: алиф (أ إ), вав (ؤ) или йа (ئ), а آ сочетает алиф с маддой. Поскольку в набранном тексте её пишут непоследовательно, поисковые системы и оценка распознавания речи часто сводят формы алифа к ا. - [Гибридный поиск (hybrid search)](https://broadrouter.com/ru/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): Гибридный поиск — это метод извлечения, который по одному запросу выполняет поиск по ключевым словам, например BM25, и векторный поиск по эмбеддингам, а затем сливает два ранжированных списка в один, часто с помощью reciprocal rank fusion (RRF) или взвешенной суммы нормализованных оценок. - [Ключ идемпотентности (idempotency key)](https://broadrouter.com/ru/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Ключ идемпотентности — это уникальное значение, обычно случайный ID вроде UUID, которое клиент прикладывает к запросу, меняющему состояние, например к платежу или покупке. Если тот же запрос приходит снова с тем же ключом, сервер возвращает исходный результат, а не выполняет действие второй раз. - [Токены изображений (image tokens)](https://broadrouter.com/ru/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): Токены изображений — это входные токены, которыми визуально-языковая модель представляет картинку в промпте. Провайдеры вычисляют их по размеру изображения, обычно после масштабирования разбивая его на патчи или плитки фиксированного размера, и тарифицируют как текстовые входные токены, поэтому большие изображения стоят дороже. - [Провайдер инференса (inference provider)](https://broadrouter.com/ru/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): Провайдер инференса — это компания, которая запускает модели машинного обучения, часто LLM с открытыми весами, на собственных, взятых в лизинг или арендованных GPU и продаёт к ним доступ через API, обычно с оплатой за токен, запрос или единицу времени. Он отвечает за обслуживание запросов, масштабирование и оборудование. - [Входные и выходные токены (input and output tokens)](https://broadrouter.com/ru/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): Входные токены — это токены, которые языковая модель читает в запросе: системный промпт, история диалога, описания инструментов и изображения. Выходные токены — это токены, которые она генерирует в ответ, включая рассуждения. LLM API обычно тарифицируют их по отдельности, и выход, как правило, дороже. - [INT4](https://broadrouter.com/ru/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 — это 4-битный целочисленный формат, способный представить 16 уровней. В инференсе LLM под ним обычно понимают 4-битную квантизацию только весов (W4A16): веса хранятся в INT4 с коэффициентами масштабирования и разворачиваются в 16 бит во время вычислений, занимая примерно четверть памяти 16-битных весов. - [Белый список IP (IP allow list)](https://broadrouter.com/ru/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): Белый список IP — это механизм безопасности, который принимает запросы только с перечисленных IP-адресов или диапазонов CIDR и отклоняет все остальные. В API он ограничивает ущерб от утёкшего ключа, потому что ключ не работает из сетей, которых нет в списке. - [Джейлбрейк LLM (jailbreak)](https://broadrouter.com/ru/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): Джейлбрейк — это ввод, составленный так, чтобы языковая модель проигнорировала своё обучение безопасности или правила развернувшего её приложения и выдала контент или выполнила действия, от которых обычно отказывается. OWASP рассматривает джейлбрейк как разновидность промпт-инъекции. - [JSON mode (режим JSON)](https://broadrouter.com/ru/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): JSON mode — это настройка API языковой модели (в OpenAI-совместимых API — response_format {"type": "json_object"}), которая делает ответ модели валидным, разбираемым JSON. В отличие от structured outputs с JSON Schema, она не гарантирует, какие ключи, типы и значения будут в объекте. - [JSON Schema (JSON-схема)](https://broadrouter.com/ru/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema — это словарь, записанный на самом JSON, для описания и валидации структуры JSON-данных: типов, обязательных полей, допустимых значений и вложенности. Текущая версия — 2020-12. API языковых моделей используют его, чтобы задать параметры инструментов и точную форму структурированного вывода. - [KV-кэш (KV cache)](https://broadrouter.com/ru/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): KV-кэш — это память, в которой трансформерная языковая модель хранит ключи и значения внимания, вычисленные для каждого уже обработанного токена, чтобы каждый новый токен использовал их повторно, а не пересчитывал всю последовательность. Он растёт с длиной контекста и часто ограничивает число запросов, которые может обслужить GPU. - [Задержка (latency)](https://broadrouter.com/ru/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): Задержка (latency) в API языковой модели — это время от отправки запроса до получения последнего выходного токена. Она складывается из времени до первого токена и времени генерации и обычно указывается в перцентилях, например p50 и p95, а не как среднее. - [LLM-as-a-judge (LLM в роли судьи)](https://broadrouter.com/ru/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge — это метод оценки, при котором языковая модель оценивает ответы других моделей: либо выставляет балл одному ответу по рубрике, либо выбирает лучший из двух. Он приближается к человеческому суждению при гораздо меньших затратах, но имеет известные перекосы, например предпочитает первый или более длинный ответ. - [Бенчмарк LLM (LLM benchmark)](https://broadrouter.com/ru/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): Бенчмарк LLM — это фиксированный, обычно публичный набор задач со способом подсчёта баллов, по которому измеряют и сравнивают языковые модели в одном навыке: ответах на вопросы на знания, программировании или распознавании речи. Балл на бенчмарке показывает относительные способности на этом тесте, а не то, как модель справится с вашей нагрузкой. - [LLM-роутер (LLM router)](https://broadrouter.com/ru/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): LLM-роутер — это программа или сервис между приложением и одним или несколькими API языковых моделей, который для каждого запроса решает, какая модель или какой провайдер его обработает, по правилам вроде стоимости, задержки, доступности или типа задачи. - [Logprobs (логарифмы вероятностей токенов)](https://broadrouter.com/ru/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Logprobs — это логарифмы вероятностей, которые языковая модель присвоила сгенерированным токенам. API, который их возвращает, показывает для каждого выходного токена, насколько вероятным его сочла модель, и при желании — самые вероятные альтернативные токены на этой позиции. - [LoRA (низкоранговая адаптация)](https://broadrouter.com/ru/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation, низкоранговая адаптация) — это метод параметро-эффективного дообучения, который замораживает веса предобученной модели и обучает пару небольших низкоранговых матриц для выбранных слоёв. Выученное изменение можно хранить как небольшой адаптер или влить в веса, не добавляя задержки при инференсе. - [Потеря в середине (lost in the middle)](https://broadrouter.com/ru/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle («потеря в середине») — это наблюдаемая склонность языковых моделей лучше всего использовать нужную информацию, когда она стоит ближе к началу или концу длинного ввода, и хуже всего — когда она находится в середине. Название эффекту дало исследование Liu и соавторов 2023 года. - [max_tokens](https://broadrouter.com/ru/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens — это параметр LLM API, который задаёт максимальное число токенов, которое модель может сгенерировать в одном ответе. Когда генерация доходит до лимита, она останавливается, а ответ помечается как обрезанный, например значением finish_reason "length". - [MCP (Model Context Protocol)](https://broadrouter.com/ru/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol) — это открытый протокол на основе JSON-RPC 2.0, который стандартизирует подключение LLM-приложений к внешним инструментам и данным. MCP-сервер предоставляет инструменты, ресурсы и промпты, а MCP-клиент внутри приложения-хоста находит их и передаёт серверу вызовы инструментов от модели. - [Смесь экспертов (mixture of experts, MoE)](https://broadrouter.com/ru/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Смесь экспертов (MoE) — это архитектура нейросети, в которой слой содержит много подсетей-экспертов, а обучаемый роутер отправляет каждый токен лишь в несколько из них. У MoE-модели большое общее число параметров, но на каждый токен она задействует небольшое активное подмножество. - [Вывод модели из эксплуатации (model deprecation)](https://broadrouter.com/ru/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): Вывод модели из эксплуатации (deprecation) — это объявление провайдера API о том, что модель перестанет предоставляться; обычно в нём указана дата отключения и рекомендуемая замена. После этой даты запросы к модели завершаются ошибкой, поэтому приложения должны заранее перейти на другой ID модели. - [Резервная модель (model fallback)](https://broadrouter.com/ru/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): Model fallback (переключение на резервную модель) — это паттерн надёжности, при котором неудавшийся запрос к LLM повторяется на другой модели или у другого провайдера из упорядоченного списка. Поводом может быть ошибка, лимит запросов, превышение длины контекста или отказ по контент-политике. Он работает в вашем клиенте или в шлюзе перед API. - [Лицензия модели (model license)](https://broadrouter.com/ru/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): Лицензия модели — это юридические условия, на которых веса, код и выходные данные модели можно использовать, изменять и распространять. Модели с открытыми весами выходят под стандартными лицензиями на ПО, например Apache 2.0 или MIT, или под собственными лицензиями, которые могут добавлять условия, например ограничения для крупных коммерческих сервисов. - [Современный стандартный арабский (MSA, фусха)](https://broadrouter.com/ru/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): Современный стандартный арабский (MSA) — это стандартизированная форма арабского языка для письма, новостей, образования и официальной речи во всём арабском мире. Ни для кого он не родной повседневный язык: люди говорят на региональных диалектах, например египетском или заливском, и переходят на MSA в официальной обстановке. - [Мультиагентная система (multi-agent system)](https://broadrouter.com/ru/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): Мультиагентная система — это LLM-приложение, в котором несколько агентов, каждый со своим промптом, инструментами и контекстом, решают части одной задачи. Распространённая схема — оркестратор и исполнители: ведущий агент планирует, параллельно запускает агентов-исполнителей и объединяет их результаты. - [Мультимодальный ИИ (multimodal AI)](https://broadrouter.com/ru/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): Мультимодальный ИИ — это модель или система, которая работает больше чем с одним типом данных, например с текстом, изображениями, видео и аудио. Мультимодальная модель может прочитать фото вместе с вопросом или принять речь и ответить текстом. Самая распространённая разновидность — визуально-языковая модель. - [Распознавание именованных сущностей (NER)](https://broadrouter.com/ru/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): Распознавание именованных сущностей (NER) — это задача найти в тексте имена и присвоить каждому тип, например «персона», «организация» или «место». Классические системы размечают каждый токен; у LLM вместо этого обычно просят вернуть список сущностей в JSON, который затем сверяют с исходным текстом. - [Оптическое распознавание символов (OCR)](https://broadrouter.com/ru/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): Оптическое распознавание символов (OCR) — это технология, которая находит печатный или рукописный текст на изображении, скане или в PDF и превращает его в машиночитаемые символы. Классический OCR выдаёт текст и его положение на странице; визуально-языковые модели могут прочитать текст и сразу ответить на вопросы о нём. - [Ogg Opus](https://broadrouter.com/ru/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus — это аудио, закодированное кодеком Opus (RFC 6716) и упакованное в контейнер Ogg (RFC 7845). Формат создан для речи и интерактивного аудио на низких битрейтах, и именно в нём WhatsApp хранит голосовые сообщения, обычно в файлах .ogg или .opus. - [Модель с открытыми весами (open-weight model)](https://broadrouter.com/ru/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): Модель с открытыми весами — это модель машинного обучения, чьи обученные веса опубликованы, чтобы любой мог их скачать и запустить, обычно по лицензии с условиями. Она отличается от полностью открытого (open source) ИИ, где публикуются ещё код обучения и подробные сведения об обучающих данных, чтобы другие могли изучить и воспроизвести модель. - [OpenAI-совместимый API](https://broadrouter.com/ru/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): OpenAI-совместимый API — это HTTP API, который принимает те же форматы запросов и ответов, что и OpenAI API, например POST /v1/chat/completions. Поэтому SDK и инструменты OpenAI работают с ним после замены всего трёх значений: базового URL, API-ключа и названия модели. - [Параллельные вызовы инструментов (parallel tool calls)](https://broadrouter.com/ru/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Параллельные вызовы инструментов — это когда модель возвращает больше одного вызова инструмента в одном ответе, например три чтения файлов сразу. В формате OpenAI они приходят несколькими элементами tool_calls, а parallel_tool_calls со значением false ограничивает модель нулём или одним вызовом за ход. - [Оплата по факту (pay-as-you-go)](https://broadrouter.com/ru/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): Оплата по факту (pay-as-you-go) — это модель ценообразования, при которой вы платите только за то, что использовали, по счётчику в единицах вроде токенов, запросов или минут аудио, без подписки и минимальных обязательств. Так тарифицирует большинство LLM API — с предоплаченного баланса или по периодическому счёту. - [PDPL (законы о персональных данных Саудовской Аравии и ОАЭ)](https://broadrouter.com/ru/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) — это общее название закона Саудовской Аравии о персональных данных, принятого королевским указом M/19, за соблюдением которого следит SDAIA, и федерального декрета-закона ОАЭ № 45 от 2021 года. Оба регулируют обработку персональных данных и задают условия передачи за рубеж, включая промпты, отправленные в иностранные LLM API. - [Персональные данные (PII)](https://broadrouter.com/ru/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): PII (personally identifiable information) — это информация, с помощью которой можно отличить или отследить личность человека, например имя, номер удостоверения или биометрические данные, а также другая информация, связанная или связываемая с ним. В GDPR используется более широкий термин «персональные данные». В работе с LLM маскирование PII удаляет их из промптов до того, как те покинут ваши системы. - [Prefill и decode (префилл и декодирование)](https://broadrouter.com/ru/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill и decode — это две фазы инференса LLM. На этапе prefill модель параллельно обрабатывает весь промпт и заполняет KV-кэш; на этапе decode она генерирует ответ по одному токену, повторно используя этот кэш. Prefill в основном определяет время до первого токена, decode — скорость вывода. - [Предоплаченный баланс (prepaid credit)](https://broadrouter.com/ru/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): Предоплаченный баланс — это деньги, внесённые на аккаунт API до начала использования. Каждый запрос списывает с баланса свою стоимость, а когда баланс заканчивается, запросы с оплатой за использование прекращаются, пока вы его не пополните. Расходы не могут превысить внесённую сумму. - [Промпт-инжиниринг (prompt engineering)](https://broadrouter.com/ru/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): Промпт-инжиниринг — это процесс написания, структурирования и проверки инструкций для языковой модели, чтобы она стабильно выдавала результат, отвечающий вашим требованиям. Он включает роли сообщений, ясные инструкции, примеры, форматирование и оценку результатов на тестовых примерах. - [Промпт-инъекция (prompt injection)](https://broadrouter.com/ru/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): Промпт-инъекция — это атака, при которой текст, переданный языковой модели пользователем или спрятанный на веб-странице, в письме, файле или результате инструмента, меняет поведение модели вопреки замыслу разработчика. OWASP ставит её на первое место (LLM01:2025) в своём Top 10 для LLM-приложений. - [Квантизация (quantization)](https://broadrouter.com/ru/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): Квантизация — это хранение чисел нейросети, в первую очередь её весов, в формате пониженной точности, например 8- или 4-битном вместо 16-битного, чтобы модель занимала меньше памяти и часто работала быстрее. Хорошие методы сохраняют потерю точности небольшой. - [RAG (генерация, дополненная поиском)](https://broadrouter.com/ru/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG (retrieval-augmented generation, генерация, дополненная поиском) — это подход, при котором приложение сначала находит во внешнем хранилище, например поисковом или векторном индексе, фрагменты, относящиеся к вопросу, а затем добавляет их в промпт, чтобы языковая модель отвечала по этому тексту, а не только по памяти. - [Лимит запросов (rate limit)](https://broadrouter.com/ru/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): Лимит запросов (rate limit) — это ограничение, которое API накладывает на объём использования клиентом за отрезок времени, например число запросов или токенов в минуту. Запросы сверх лимита отклоняются, обычно с HTTP 429 Too Many Requests, пока использование снова не опустится ниже порога. - [ReAct (reason and act, «рассуждай и действуй»)](https://broadrouter.com/ru/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (reason and act) — это паттерн промптинга, предложенный Yao et al. в 2022 году, в котором языковая модель чередует цепочки рассуждений с действиями, например поиском или вызовами инструментов. Она читает каждое наблюдение, обновляет план и повторяет цикл, пока не сможет дать итоговый ответ. - [Токены рассуждений (reasoning tokens)](https://broadrouter.com/ru/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): Токены рассуждений — это токены, которые модель с рассуждениями генерирует, чтобы проработать задачу, прежде чем написать видимый ответ. Обычно они оплачиваются как выходные токены, учитываются в лимите на ответ и в контекстном окне, а API может их скрывать или выдавать в сжатом виде. - [Тайм-аут запроса (request timeout)](https://broadrouter.com/ru/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): Тайм-аут запроса — это наибольшее время, которое HTTP-клиент ждёт сервер, прежде чем отказаться от запроса. Обычно клиенты делят его на тайм-аут подключения, для установки соединения, и тайм-аут чтения, для ожидания данных. Вызовам LLM нужны щедрые тайм-ауты чтения: длинные ответы требуют времени. - [Реранкинг (reranking)](https://broadrouter.com/ru/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Реранкинг — это второй этап поиска, на котором более точная, но медленная модель, обычно кросс-энкодер, оценивает каждую пару из запроса и документа-кандидата, найденного первичным поиском, и переупорядочивает их, чтобы в языковую модель попали только несколько самых релевантных. - [Текст справа налево (RTL, right-to-left)](https://broadrouter.com/ru/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): Текст справа налево (RTL) — это письмо, например арабское или иврит, которое отображается справа налево. Компьютер хранит его в логическом порядке чтения, а когда RTL-текст смешивается с текстом слева направо, цифрами, названиями моделей или кодом, порядок отображения определяет двунаправленный алгоритм Unicode. - [Частота дискретизации (sample rate)](https://broadrouter.com/ru/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): Частота дискретизации — это число отсчётов звука, записываемых за секунду, в герцах (Гц). Модели распознавания речи обычно ожидают моно-аудио с частотой 16 000 Гц (16 кГц). Вместе с разрядностью и числом каналов частота дискретизации определяет, сколько весит минута несжатого аудио. - [Параметр seed (сид)](https://broadrouter.com/ru/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): Параметр seed — это целое число, которое вы передаёте в запросе, чтобы API постарался выбрать те же токены при повторе того же запроса. OpenAI описывает это как best effort: детерминизм не гарантирован, а поле system_fingerprint показывает, когда изменилась конфигурация бэкенда. - [Бессерверный инференс (serverless inference)](https://broadrouter.com/ru/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): Бессерверный инференс — это способ запуска моделей машинного обучения, при котором платформа сама запускает, масштабирует и останавливает вычислительные ресурсы за эндпоинтом, часто до нуля в простое, и берёт плату за использование, а не за зарезервированные машины. Обратная сторона — возможная задержка холодного старта. - [SLA (соглашение об уровне обслуживания)](https://broadrouter.com/ru/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): SLA (соглашение об уровне обслуживания, service level agreement) — это договорное обязательство провайдера обеспечить измеримый уровень сервиса, обычно процент доступности за месяц, с заранее определённой компенсацией, как правило сервисными кредитами, если он не выдержан. SLA определяет, что считается простоем, что исключено и как подать заявку. - [Спекулятивное декодирование (speculative decoding)](https://broadrouter.com/ru/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): Спекулятивное декодирование — это техника инференса, при которой небольшая быстрая черновая модель предлагает несколько следующих токенов, а большая целевая модель проверяет их за один прямой проход и оставляет принятые. Генерация ускоряется, а распределение выходных токенов не меняется. - [Распознавание речи (speech-to-text)](https://broadrouter.com/ru/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): Распознавание речи (speech-to-text, автоматическое распознавание речи, ASR) — это технология, которая превращает устную речь в письменный текст. API распознавания речи принимает аудиофайл или поток и возвращает расшифровку, обычно с оплатой за минуту аудио. Точность измеряют долей ошибок в словах (WER), и она зависит от акцента, диалекта и качества записи. - [Лимит расходов (spend limit)](https://broadrouter.com/ru/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): Лимит расходов — это потолок суммы, которую аккаунт, проект или ключ API может потратить за период, обычно за месяц. Когда расход его достигает, провайдер отклоняет новые запросы, часто с ошибкой HTTP 429 или 400, пока не начнётся новый период или лимит не поднимут. - [Субтитры SRT (SRT subtitles)](https://broadrouter.com/ru/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) — это текстовый формат субтитров. Каждый титр состоит из порядкового номера, времени начала и конца в виде hh:mm:ss,mmm --> hh:mm:ss,mmm, одной или нескольких строк текста и пустой строки. WebVTT — похожий веб-формат, который использует видео в HTML. - [Server-sent events (SSE)](https://broadrouter.com/ru/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Server-sent events (SSE) — это веб-стандарт, по которому сервер передаёт клиенту поток текстовых событий в одном долгоживущем HTTP-ответе с типом содержимого text/event-stream. Каждое событие — одна или несколько строк-полей, например data:, и завершается пустой строкой. - [Стоп-последовательности (stop sequences)](https://broadrouter.com/ru/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): Стоп-последовательности — это строки, которые вы передаёте в запросе, чтобы модель прекратила генерацию, как только выведет одну из них. В OpenAI поле называется stop и принимает до 4 последовательностей, в Anthropic — stop_sequences. Совпавший текст обычно не попадает в возвращаемый ответ. - [Стриминг (streaming)](https://broadrouter.com/ru/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): Стриминг — это способ возвращать ответ языковой модели небольшими частями (чанками) прямо во время генерации, а не одним ответом в конце. OpenAI-совместимые API передают поток через server-sent events, если в запросе stream равен true. - [Структурированный вывод (structured output)](https://broadrouter.com/ru/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): Структурированный вывод — это функция LLM API, которая ограничивает ответ модели форматом, заданным разработчиком, обычно JSON Schema, чтобы код мог надёжно его разобрать. Он строже, чем JSON mode, который гарантирует корректный JSON, но не конкретную структуру. - [Субагент (subagent)](https://broadrouter.com/ru/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): Субагент — это отдельный агент, которого основной агент запускает для одной узкой задачи. Он работает со своим контекстным окном, системным промптом и доступом к инструментам, выполняет работу и возвращает основному агенту только результат или краткую сводку, благодаря чему основной диалог остаётся компактным. - [Субобработчик (subprocessor)](https://broadrouter.com/ru/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): Субобработчик — это третья сторона, которую обработчик привлекает к обработке персональных данных от имени контролёра. У LLM API субобработчиками могут быть GPU-хосты, партнёрские провайдеры инференса и резервные мощности. По статье 28 GDPR им нужно разрешение контролёра и те же обязательства по защите данных. - [Системный промпт (system prompt)](https://broadrouter.com/ru/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): Системный промпт — это текст инструкций, который приложение отправляет языковой модели перед диалогом, обычно как сообщение с ролью system или developer. Он задаёт роль модели, правила, тон и формат ответа, и при конфликте с сообщениями пользователя модель должна следовать именно ему. - [Та марбута (ة, taa marbuta)](https://broadrouter.com/ru/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): Та марбута (ة, U+0629) — это арабская буква, которая встречается только в конце слова и часто обозначает женский род существительных и прилагательных. Она выглядит как ха (ه, U+0647) с двумя точками, поэтому в набранном тексте их путают, и многие поисковые нормализаторы заменяют ة на ه. - [Температура (temperature)](https://broadrouter.com/ru/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): Температура — это параметр сэмплирования, который определяет, насколько случайным будет ответ языковой модели. Модель делит оценки токенов на температуру, прежде чем превратить их в вероятности: при низких значениях доминирует самый вероятный токен, при высоких распределение выравнивается и менее вероятные токены выбираются чаще. - [Синтез речи (text-to-speech, TTS)](https://broadrouter.com/ru/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): Синтез речи (text-to-speech, TTS) — это преобразование письменного текста в звучащую речь. API TTS принимает текст, голос и формат вывода, например MP3 или WAV, и возвращает сгенерированную речь. Это обратная задача по отношению к распознаванию речи, которое превращает аудио в текст. - [Пропускная способность (throughput)](https://broadrouter.com/ru/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): Пропускная способность (throughput) в обслуживании LLM — это общий объём работы, который система выполняет за единицу времени по всем параллельным запросам; обычно её измеряют в выходных токенах или запросах в секунду. Она отличается от скорости потока — того, как быстро токены получает отдельный запрос. - [Время до первого токена (TTFT)](https://broadrouter.com/ru/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): Время до первого токена (TTFT) — это время от отправки запроса языковой модели до прихода первого выходного токена. Оно включает сетевое время, очередь на сервере и prefill — этап, на котором модель читает весь промпт, поэтому растёт с длиной промпта и нагрузкой. - [Временные метки в распознавании речи (timestamps)](https://broadrouter.com/ru/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): Временные метки в распознавании речи — это время начала и конца в секундах, привязанное к частям расшифровки. Метки сегментов охватывают фразу или предложение, метки слов — каждое слово. Они нужны для субтитров, поиска внутри записей и сопоставления текста с говорящими. - [Токен (token)](https://broadrouter.com/ru/glossary/token/) (translation of https://broadrouter.com/glossary/token/): Токен — это единица текста, которую языковая модель читает и пишет: слово, часть слова, знак препинания или последовательность байтов, которую выдаёт токенизатор модели. LLM API считают, ограничивают и обычно тарифицируют использование в токенах. - [Токенизатор (tokenizer)](https://broadrouter.com/ru/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): Токенизатор — это компонент, который преобразует текст в последовательность целочисленных ID токенов, которую читает языковая модель, и превращает сгенерированные ID обратно в текст. У каждой модели свой токенизатор, поэтому один и тот же текст может дать разное число токенов на разных моделях. - [Токены в секунду (tokens per second, TPS)](https://broadrouter.com/ru/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): Токены в секунду (TPS) — это скорость, с которой языковая модель генерирует выходные токены; обычно её измеряют для одного запроса после прихода первого токена. Она определяет, как быстро стримится начавшийся ответ. В сумме по всем параллельным запросам это пропускная способность системы; ожидание первого токена измеряют отдельно. - [Вызов инструментов (tool calling)](https://broadrouter.com/ru/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): Вызов инструментов (function calling) — это функция LLM API, при которой модель возвращает структурированный запрос на выполнение функции, заданной разработчиком, с аргументами в JSON. Приложение выполняет функцию и отправляет результат обратно; сама модель код никогда не выполняет. - [tool_choice](https://broadrouter.com/ru/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice — это параметр запроса, который управляет использованием инструментов. В Chat Completions от OpenAI он может быть auto (модель решает сама), required (модель обязана вызвать один или несколько инструментов), none (без инструментов) или именем функции, вызов которой принудителен. Аналогичные варианты у Anthropic — auto, any, tool и none. - [Top-p (ядерное сэмплирование, nucleus sampling)](https://broadrouter.com/ru/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Top-p, или ядерное сэмплирование (nucleus sampling), — это метод декодирования, при котором языковая модель выбирает следующий токен только из наименьшего набора самых вероятных токенов, чья суммарная вероятность достигает порога p, например 0,9, а маловероятный «хвост» отбрасывает. - [Нормализация Unicode (Unicode normalization)](https://broadrouter.com/ru/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Нормализация Unicode — это приведение текста к стандартной форме, чтобы одинаково выглядящие строки состояли из одних и тех же кодовых точек. Форм четыре: NFD, NFC, NFKD и NFKC; формы K вдобавок сводят символы совместимости, например арабские формы представления и лигатуры, к обычным буквам. - [Детекция речевой активности (VAD)](https://broadrouter.com/ru/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): Детекция речевой активности (voice activity detection, VAD) — это этап обработки сигнала, который помечает короткие кадры аудио как речь или не речь. Речевые системы используют её, чтобы пропускать тишину, делить длинные записи по паузам, понимать, что собеседник договорил, и не транскрибировать фоновый шум. - [Векторная база данных (vector database)](https://broadrouter.com/ru/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): Векторная база данных — это хранилище, которое держит векторы эмбеддингов, часто вместе с исходным текстом и метаданными, и отвечает на запросы ближайших соседей: по вектору запроса возвращает самые похожие сохранённые векторы, обычно по косинусному расстоянию, евклидову расстоянию или скалярному произведению. - [Визуально-языковая модель (VLM)](https://broadrouter.com/ru/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): Визуально-языковая модель (vision-language model, VLM) — это модель, которая принимает на вход изображения вместе с текстом и выдаёт текст, поэтому может описывать картинки, отвечать на вопросы о них и читать документы, графики и скриншоты. VLM также называют моделями image-text-to-text или мультимодальными моделями. - [vLLM](https://broadrouter.com/ru/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM — это библиотека с открытым кодом для инференса и сервинга LLM, изначально созданная в Sky Computing Lab Калифорнийского университета в Беркли. Она управляет памятью KV-кэша с помощью PagedAttention, непрерывно батчит входящие запросы, поддерживает многие форматы квантизации и предоставляет OpenAI-совместимый HTTP-сервер. - [Размер словаря (vocabulary size)](https://broadrouter.com/ru/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): Размер словаря — это число различных токенов в токенизаторе языковой модели, у каждого из которых свой ID и свой эмбеддинг. Больший словарь может кодировать распространённый текст меньшим числом токенов, но увеличивает слои эмбеддингов и выхода и помогает только для текста, на котором словарь обучали. - [WAV](https://broadrouter.com/ru/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format, или WAVE) — это аудиоконтейнер на основе формата RIFF от Microsoft и IBM. Обычно он содержит несжатые отсчёты линейного PCM, поэтому его размер предсказуем: частота дискретизации × байт на отсчёт × число каналов за каждую секунду. - [Доля ошибок в словах (word error rate, WER)](https://broadrouter.com/ru/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): Доля ошибок в словах (word error rate, WER) — это стандартная метрика точности распознавания речи: число замен, пропусков и вставок слов, нужных, чтобы превратить расшифровку в правильный эталон, делённое на число слов в эталоне. Чем меньше, тем лучше. - [Нулевое хранение данных (ZDR)](https://broadrouter.com/ru/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): Нулевое хранение данных (ZDR) — это политика обработки данных, при которой AI-провайдер обрабатывает промпты и ответы только для выполнения запроса и после этого не сохраняет их содержимое ни в логах, ни в кэшах, ни в обучающих данных. Записи без содержимого, например число токенов и время выполнения, обычно всё же хранятся. ### Vietnamese (vi, Tiếng Việt): 294 translated pages - [Articles in Vietnamese](https://broadrouter.com/vi/blog/) - [Thay đổi giá API LLM tháng 10/2026: 3 lần tăng, không lần giảm nào](https://broadrouter.com/vi/blog/llm-api-price-changes-october-2026/) (translation of https://broadrouter.com/blog/llm-api-price-changes-october-2026/): 23 thay đổi trên trang giá của 5 trong 7 nhà cung cấp API LLM tháng 10/2026: 3 lần tăng, không giảm, 9 model có giá mới và 11 bị gỡ. Kèm link bản lưu cho từng mục. - [Chi phí một lần chạy AI coding agent: số bước, context và cache](https://broadrouter.com/vi/blog/ai-coding-agent-run-cost/) (translation of https://broadrouter.com/blog/ai-coding-agent-run-cost/): Vì sao lần chạy coding agent càng dài thì mỗi bước càng đắt, cache thay đổi tổng chi phí ra sao, và khi nào phiên giá cố định thắng trả theo token. Kèm phép tính. - [Thêm dấu cho tiếng Ả Rập bằng LLM: prompt, kiểm tra, DER](https://broadrouter.com/vi/blog/arabic-diacritization-llm/) (translation of https://broadrouter.com/blog/arabic-diacritization-llm/): Khôi phục dấu tiếng Ả Rập (tashkeel) bằng API LLM: prompt giữ nguyên từng chữ cái, bước kiểm tra loại bỏ chỉnh sửa, chấm DER và WER, và khi nào dùng CATT. - [Sửa ngữ pháp và chính tả tiếng Ả Rập bằng LLM: hướng dẫn](https://broadrouter.com/vi/blog/arabic-grammar-spelling-correction-llm/) (translation of https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/): Sửa chính tả, ngữ pháp tiếng Ả Rập bằng API LLM: prompt sửa tối thiểu, chặn viết lại, chỉnh sửa theo từ phân loại theo lỗi và chấm điểm kiểu M2. - [Biên bản họp tiếng Ả Rập từ bản ghi âm: một pipeline API](https://broadrouter.com/vi/blog/arabic-meeting-minutes-from-recordings/) (translation of https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/): Biến ghi âm cuộc họp tiếng Ả Rập thành biên bản có quyết định và đầu việc: chép lời, làm sạch, tóm tắt bằng LLM ra JSON, rồi kiểm tra tên và ngày tháng. - [Trích xuất thực thể tiếng Ả Rập bằng LLM: JSON, ngày tháng, kiểm tra](https://broadrouter.com/vi/blog/arabic-named-entity-extraction-llm/) (translation of https://broadrouter.com/blog/arabic-named-entity-extraction-llm/): Trích xuất tên người, tổ chức, địa danh và ngày Hijri tiếng Ả Rập ra JSON bằng LLM: prompt, kiểm tra span bỏ qua dấu nguyên âm, quy đổi ngày và script chấm điểm. - [So sánh API OCR tiếng Ả Rập: Google, Azure, Textract, Mistral, LLM](https://broadrouter.com/vi/blog/arabic-ocr-api-comparison/) (translation of https://broadrouter.com/blog/arabic-ocr-api-comparison/): API OCR nào chính thức hỗ trợ tiếng Ả Rập in và viết tay, mỗi API trả về gì và tốn bao nhiêu mỗi 1.000 trang, so với các API LLM vision. - [Phụ đề tiếng Ả Rập (SRT) từ API chuyển giọng nói: cần những gì](https://broadrouter.com/vi/blog/arabic-subtitles-srt-api/) (translation of https://broadrouter.com/blog/arabic-subtitles-srt-api/): File SRT cần gì, API chuyển giọng nói thành văn bản nào trả timestamp hay SRT cho tiếng Ả Rập, và một pipeline đã test để làm phụ đề từ API chỉ trả văn bản. - [Phân loại văn bản tiếng Ả Rập bằng LLM: cảm xúc, ý định, ticket](https://broadrouter.com/vi/blog/arabic-text-classification-llm/) (translation of https://broadrouter.com/blog/arabic-text-classification-llm/): Phân loại đánh giá, tin nhắn và ticket tiếng Ả Rập bằng API LLM: thiết kế nhãn, prompt theo phương ngữ, output JSON, test macro F1 và chi phí khi chạy số lượng lớn. - [Chuẩn hóa văn bản tiếng Ả Rập cho tìm kiếm và RAG: đổi gì, giữ gì](https://broadrouter.com/vi/blog/arabic-text-normalization-search/) (translation of https://broadrouter.com/blog/arabic-text-normalization-search/): Tashkeel, dạng alef, ta marbuta, vị trí hamza, chữ Ba Tư, chữ số và Unicode: quy tắc chuẩn hóa tiếng Ả Rập nào giúp tìm kiếm, quy tắc nào gây hại, kèm code Python. - [Tóm tắt văn bản tiếng Ả Rập bằng LLM: prompt, độ dài và chấm điểm](https://broadrouter.com/vi/blog/arabic-text-summarization-llm/) (translation of https://broadrouter.com/blog/arabic-text-summarization-llm/): Tóm tắt bài báo, báo cáo, tin nhắn phương ngữ tiếng Ả Rập qua LLM API: mẫu prompt, kiểm soát số từ, kiểm tra con số và vì sao ROUGE mặc định cho tiếng Ả Rập điểm 0. - [Lựa chọn thay thế Artificial Analysis: nơi so sánh LLM năm 2026](https://broadrouter.com/vi/blog/artificial-analysis-alternatives/) (translation of https://broadrouter.com/blog/artificial-analysis-alternatives/): Các lựa chọn thay thế Artificial Analysis: Arena, Epoch AI, Vals AI, bảng xếp hạng OpenRouter, Hugging Face và dữ liệu giá, theo thứ mỗi bên đo và khi nào nên dùng. - [Phần cứng GPU và cách phục vụ model của BroadNet, theo lời BroadNet](https://broadrouter.com/vi/blog/broadnet-gpu-fleet-and-serving/) (translation of https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/): Những gì BroadNet nói về phần cứng đằng sau BroadNet Inference: ba tầng GPU từ B300 đến thiết bị tại chỗ khách hàng, thông lượng mỗi GPU và bài học từ thử nghiệm. - [Cache phản hồi API LLM để giảm chi phí: cache khớp chính xác (Python)](https://broadrouter.com/vi/blog/cache-llm-responses-cut-cost/) (translation of https://broadrouter.com/blog/cache-llm-responses-cut-cost/): Cache khớp chính xác đã kiểm thử cho phản hồi API LLM bằng Python: key gồm gì, thứ không bao giờ cache, TTL, quyền riêng tư và cách đo tỷ lệ trúng trước khi xây. - [Claude Code với model mở: proxy, endpoint gốc và giới hạn](https://broadrouter.com/vi/blog/claude-code-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/claude-code-with-openai-compatible-models/): Claude Code dùng Anthropic Messages API. Cách chạy nó với model mở qua proxy LiteLLM hoặc endpoint gốc, và những gì có thể hỏng trên đường đi. - [Đưa bước người duyệt vào quy trình AI xử lý tài liệu](https://broadrouter.com/vi/blog/document-review-queues/) (translation of https://broadrouter.com/blog/document-review-queues/): Thiết kế bước người duyệt trong pipeline AI tài liệu: quy tắc kiểm tra, tài liệu nào cần gửi cho người, mẫu kiểm toán ngẫu nhiên, nhật ký kiểm toán và tỷ lệ duyệt. - [Luật bảo vệ dữ liệu Ai Cập (Luật 151 năm 2020) và API LLM](https://broadrouter.com/vi/blog/egypt-data-protection-law-llm-api/) (translation of https://broadrouter.com/blog/egypt-data-protection-law-llm-api/): Gửi dữ liệu cá nhân của người Ai Cập tới API LLM hay chuyển giọng nói nước ngoài: Luật 151/2020 và Quyết định 816/2025 nói gì về chuyển dữ liệu, giấy phép và rò rỉ. - [Trước khi fine-tune: xác định vấn đề cần giải quyết](https://broadrouter.com/vi/blog/fine-tuning-readiness/) (translation of https://broadrouter.com/blog/fine-tuning-readiness/): Khi nào prompt, RAG hay đánh giá đã giải quyết được vấn đề và không cần fine-tune, cần dữ liệu huấn luyện gì, và cách thử một bản fine-tune trước khi yêu cầu. - [Sửa JSON không hợp lệ từ LLM: nguyên nhân, sửa lỗi, thử lại](https://broadrouter.com/vi/blog/fix-invalid-json-from-llm/) (translation of https://broadrouter.com/blog/fix-invalid-json-from-llm/): Vì sao LLM trả về JSON hỏng (code fence, câu dẫn, bị cắt ở max_tokens, dấu phẩy thừa) và một hàm Python đã test, chỉ sửa phần an toàn và từ chối phần còn lại. - [Giá API LLM thay đổi thường xuyên đến mức nào? 3 năm trang giá lưu trữ](https://broadrouter.com/vi/blog/how-often-llm-api-prices-change/) (translation of https://broadrouter.com/blog/how-often-llm-api-prices-change/): Chúng tôi đọc 308 trang giá lưu trữ của 11 nhà cung cấp API LLM, 2023–2026. 80% model chưa từng đổi giá; DeepInfra chiếm 75% tổng số lần thay đổi giá niêm yết. - [Cách giảm chi phí API LLM: checklist 10 bước theo đúng thứ tự](https://broadrouter.com/vi/blog/how-to-reduce-llm-api-costs/) (translation of https://broadrouter.com/blog/how-to-reduce-llm-api-costs/): Giảm hóa đơn API LLM đúng thứ tự: đo lường, cache phần đầu prompt, rút gọn prompt, giới hạn đầu ra và suy luận, chọn model, batch, và khi nào nên dùng gói cố định. - [Failover API LLM: chuyển sang nhà cung cấp thứ hai bằng code của bạn](https://broadrouter.com/vi/blog/llm-api-failover-fallback/) (translation of https://broadrouter.com/blog/llm-api-failover-fallback/): Cách failover giữa các nhà cung cấp LLM tương thích OpenAI: lỗi nào cần chuyển, circuit breaker, ánh xạ model ID, streaming, chi phí và điều khoản dữ liệu. - [Giảm giá API LLM 2023–2026: những lần giảm lớn nhất, kèm nguồn](https://broadrouter.com/vi/blog/llm-api-price-cuts-2023-2026/) (translation of https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/): Những lần giảm giá API LLM lớn nhất 2023–2026, mỗi lần kèm link Internet Archive trước và sau: o3, Codestral, Gemini 1.5, và cả những lần tăng giá ít ai để ý. - [Giá API LLM chỉ giảm? 52 lần tăng giá và cách lập ngân sách](https://broadrouter.com/vi/blog/llm-api-price-increases/) (translation of https://broadrouter.com/blog/llm-api-price-increases/): Giá API LLM vẫn có thể tăng: 52 lần tăng giá niêm yết trên cùng model trong trang giá lưu trữ 2024–2026, trung vị 40%. Kèm nguồn và quy tắc ngân sách. - [Giới hạn tốc độ API LLM theo nhà cung cấp: so sánh RPM, TPM, đồng thời](https://broadrouter.com/vi/blog/llm-api-rate-limits-by-provider/) (translation of https://broadrouter.com/blog/llm-api-rate-limits-by-provider/): Giới hạn tốc độ API LLM ở hạng đầu của 13 nhà cung cấp, đọc từ tài liệu chính thức: OpenAI, Anthropic, Gemini, Groq, DeepSeek, Fireworks…, kèm giải thích RPM, TPM. - [So sánh SLA API LLM: Azure, Bedrock, Google và OpenAI cam kết gì](https://broadrouter.com/vi/blog/llm-api-sla-uptime-compared/) (translation of https://broadrouter.com/blog/llm-api-sla-uptime-compared/): Mỗi nhà cung cấp API LLM cam kết gì bằng văn bản: tỷ lệ uptime, thế nào là downtime, tín dụng dịch vụ và mục tiêu độ trễ, đọc trực tiếp từ các trang SLA. - [Xử lý LLM hàng loạt không cần Batch API: runner chạy tiếp được](https://broadrouter.com/vi/blog/llm-batch-processing-without-batch-api/) (translation of https://broadrouter.com/blog/llm-batch-processing-without-batch-api/): Chạy 10.000 đến 1.000.000 prompt qua API chat completions khi không có Batch API: giới hạn đồng thời, điều tốc, retry, checkpoint chạy tiếp và tính thời gian. - [Safety Map: 39 LLM xử lý dữ liệu nhạy cảm như thế nào](https://broadrouter.com/vi/blog/llm-safety-map-39-models/) (translation of https://broadrouter.com/blog/llm-safety-map-39-models/): BroadNet khảo sát 39 LLM từ 14 phòng lab về credential, PII, dữ liệu y tế và tài chính: điều gì chuyển giao được, điều gì không, và rò rỉ qua tool call, suy luận. - [Kiểm tra độ an toàn dữ liệu nhạy cảm của LLM: phương pháp BroadNet](https://broadrouter.com/vi/blog/llm-sensitive-data-safety-evaluation-method/) (translation of https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/): Phương pháp BroadNet công bố để kiểm tra LLM có làm lộ credential, PII, dữ liệu y tế hay tài chính trước khi triển khai: nhóm dữ liệu, rò rỉ và bỏ sót, chạy 3+ lần. - [Chuyển khỏi OpenAI API: hướng dẫn code bằng Python, Node, cURL](https://broadrouter.com/vi/blog/migrate-from-openai-api-code-walkthrough/) (translation of https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/): Các thay đổi code để chuyển từ OpenAI sang API tương thích OpenAI: cài đặt client, từ Responses sang Chat Completions, công cụ, streaming, embeddings và smoke test. - [Những model BroadNet đã đánh giá và loại bỏ, kèm lý do](https://broadrouter.com/vi/blog/models-broadnet-ruled-out/) (translation of https://broadrouter.com/blog/models-broadnet-ruled-out/): BroadNet công bố các model đã thử và loại bỏ kèm lý do: model chat trượt bài kiểm tra an toàn thông tin xác thực, model giọng nói, OCR và embedding bị vượt mặt. - [Thiết kế sản phẩm AI đa ngôn ngữ: tiếng Ả Rập, Anh, Tây Ban Nha](https://broadrouter.com/vi/blog/multilingual-product-design/) (translation of https://broadrouter.com/blog/multilingual-product-design/): Thiết kế sản phẩm AI cho người dùng Ả Rập, Anh và Tây Ban Nha: locale giao diện và ngôn ngữ model, RTL và văn bản hỗn hợp, số, ngày, nhận diện ngôn ngữ, chi phí. - [Chi phí API model trọng số mở so với model đóng: dữ liệu giá nói gì](https://broadrouter.com/vi/blog/open-weight-vs-closed-model-api-cost/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/): Model trọng số mở qua API có rẻ hơn GPT, Claude hay Gemini? Giá theo token từ trang của chính nhà cung cấp, chênh lệch theo nhà host và chi phí mỗi người dùng. - [Lịch sử giá API OpenAI, Anthropic và Google, 2023–2026](https://broadrouter.com/vi/blog/openai-anthropic-google-api-price-history/) (translation of https://broadrouter.com/blog/openai-anthropic-google-api-price-history/): Giá API GPT, Claude và Gemini đặt cạnh nhau từ các trang giá lưu trữ, 2023–2026: dòng nào rẻ đi, dòng nào đắt lên, mỗi mức giá đều kèm link lưu trữ. - [Base URL của API tương thích OpenAI: 14 nhà cung cấp, đã kiểm tra](https://broadrouter.com/vi/blog/openai-compatible-api-base-urls/) (translation of https://broadrouter.com/blog/openai-compatible-api-base-urls/): base_url của 14 API LLM tương thích OpenAI, lấy từ tài liệu chính thức của từng bên, kèm các endpoint được liệt kê và những khác biệt làm hỏng việc đổi nhanh. - [OpenClaw và Hermes Agent với model giá rẻ: cài đặt và tính token](https://broadrouter.com/vi/blog/openclaw-hermes-agent-cheap-model/) (translation of https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/): Kết nối OpenClaw hoặc Hermes Agent với endpoint tương thích OpenAI, hiểu vì sao kiểm tra định kỳ đốt token, và giảm hóa đơn bằng context nhẹ, model rẻ hơn. - [Endpoint riêng bắt đầu từ một khối lượng công việc rõ ràng](https://broadrouter.com/vi/blog/private-endpoint-planning/) (translation of https://broadrouter.com/blog/private-endpoint-planning/): Cần xác định gì trước khi yêu cầu endpoint LLM riêng: khối lượng công việc, token mỗi giây, đồng thời, context, khu vực, giấy phép, ngân sách, rồi cách kiểm thử. - [Prompt injection trong coding agent: cách phòng thủ thực sự hiệu quả](https://broadrouter.com/vi/blog/prompt-injection-coding-agents/) (translation of https://broadrouter.com/blog/prompt-injection-coding-agents/): Prompt injection lọt vào coding agent từ đâu, và các lớp phòng thủ được nhà cung cấp, nhà nghiên cứu ghi nhận: sandbox, chặn mạng ra ngoài, allow-list và phê duyệt. - [Chunk size và overlap trong RAG: ảnh hưởng tới truy xuất và chi phí](https://broadrouter.com/vi/blog/rag-chunk-size-overlap/) (translation of https://broadrouter.com/blog/rag-chunk-size-overlap/): Chunk size và overlap thay đổi recall của RAG, kích thước index, chi phí embedding và số token mỗi truy vấn gửi tới LLM ra sao, kèm số liệu đã công bố. - [Đặt max_tokens cho model suy luận: phương pháp lập ngân sách token](https://broadrouter.com/vi/blog/reasoning-model-max-tokens-budget/) (translation of https://broadrouter.com/blog/reasoning-model-max-tokens-budget/): Vì sao model suy luận trả về câu trả lời trống khi max_tokens thấp, và cách định ngân sách từ phân vị đo được, context window và cơ chế retry tăng dần. - [Ghi âm trên trình duyệt và gửi tới API chuyển giọng nói thành văn bản](https://broadrouter.com/vi/blog/record-browser-audio-transcription-api/) (translation of https://broadrouter.com/blog/record-browser-audio-transcription-api/): Ghi âm micro bằng MediaRecorder, chuyển WebM hoặc MP4 của Safari sang WAV 16 kHz, rồi gửi tới API chép lời qua một server nhỏ giữ API key. - [Giảm token prompt: số đo thực cho system prompt, tool và dữ liệu](https://broadrouter.com/vi/blog/reduce-prompt-tokens/) (translation of https://broadrouter.com/blog/reduce-prompt-tokens/): Số token tiết kiệm được khi rút gọn system prompt, ví dụ few-shot, tool schema, định dạng dữ liệu, lịch sử chat và dấu tiếng Ả Rập, đếm bằng tokenizer thật. - [Cùng model, khác nhà host: giá API LLM chênh nhau bao nhiêu](https://broadrouter.com/vi/blog/same-llm-different-host-prices/) (translation of https://broadrouter.com/blog/same-llm-different-host-prices/): Với 119 model được nhiều nhà host bán, giá niêm yết đắt nhất có trung vị bằng 1,76 lần giá rẻ nhất. Dữ liệu từ trang giá của 41 nhà cung cấp, tháng 11/2026. - [PDPL Ả Rập Xê Út và UAE với API LLM: checklist chuyển dữ liệu](https://broadrouter.com/vi/blog/saudi-uae-pdpl-llm-api-checklist/) (translation of https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/): Gửi dữ liệu cá nhân tới API LLM từ Ả Rập Xê Út hoặc UAE: quy định chuyển dữ liệu của PDPL nói gì, nên hỏi nhà cung cấp gì, và khi nào cần inference trong nước. - [Văn bản, giọng nói hay vision? Hãy bắt đầu từ tác vụ](https://broadrouter.com/vi/blog/text-speech-vision/) (translation of https://broadrouter.com/blog/text-speech-vision/): Chọn chat văn bản, đầu vào ảnh hay chép lời âm thanh theo tác vụ, với các model và giới hạn BroadNet công bố, cùng những gì họ không cung cấp, như TTS. - [Chép lời âm thanh dài trên 25 MB: tách, gửi và ghép lại](https://broadrouter.com/vi/blog/transcribe-long-audio-files-api/) (translation of https://broadrouter.com/blog/transcribe-long-audio-files-api/): Cách chép lời bản ghi vượt giới hạn tệp của API chuyển giọng nói thành văn bản: thu nhỏ tệp, tách tại khoảng lặng bằng ffmpeg, gửi song song và ghép văn bản. - [Dịch chuỗi ứng dụng sang tiếng Ả Rập bằng LLM: placeholder và số nhiều](https://broadrouter.com/vi/blog/translate-app-strings-to-arabic-llm/) (translation of https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/): Dịch file i18n JSON sang tiếng Ả Rập bằng API LLM: cố định placeholder, sinh đủ sáu dạng số nhiều, kiểm tra cú pháp ICU trong CI và cô lập tên được chèn. - [Dịch âm thanh tiếng Ả Rập sang văn bản tiếng Anh bằng API](https://broadrouter.com/vi/blog/translate-arabic-audio-to-english/) (translation of https://broadrouter.com/blog/translate-arabic-audio-to-english/): Biến bản ghi tiếng Ả Rập thành văn bản tiếng Anh bằng hai lần gọi API: chép lời với language=ar, rồi dịch bằng LLM, kèm kiểm tra số, chia đoạn, chi phí và chấm điểm. - [Live, theo yêu cầu, nghiên cứu: đọc danh mục model cho đúng](https://broadrouter.com/vi/blog/understand-model-status/) (translation of https://broadrouter.com/blog/understand-model-status/): Cách phân biệt model trong danh mục nào gọi được ngay, model nào cần liên hệ và model nào chỉ mới được test, lấy danh mục của BroadNet và /v1/models làm ví dụ. - [Uno trong vLLM: nghiên cứu phục vụ của BroadNet đã đo những gì](https://broadrouter.com/vi/blog/uno-in-vllm-serving-study/) (translation of https://broadrouter.com/blog/uno-in-vllm-serving-study/): BroadNet tích hợp adapter tự soạn nháp Uno vào vLLM và đo trên H100: throughput gấp 2,58 lần bản thường ở 1 luồng, 1,74 lần ở 32 luồng. Phương pháp, DOI và giới hạn. - [Uno trên Gemma 4 26B: nghiên cứu mixture-of-experts của BroadNet](https://broadrouter.com/vi/blog/uno-mixture-of-experts-gemma-4-study/) (translation of https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/): BroadNet đã huấn luyện adapter tự soạn nháp Uno đầu tiên cho một model mixture-of-experts, Gemma 4 26B A4B. Họ đo gì, khi nào tự soạn nháp có lợi, và DOI. - [Ngân sách token của agent: token trong vòng lặp agent đi đâu](https://broadrouter.com/vi/blog/agent-token-budget/) (translation of https://broadrouter.com/blog/agent-token-budget/): System prompt, schema tool, đọc file, lịch sử và suy luận: vòng lặp agent tiêu token thế nào, cách đo từng phần từ dữ liệu usage và nên cắt gì trước. - [Tỷ lệ lỗi từ (WER) trong nhận dạng giọng nói tiếng Ả Rập](https://broadrouter.com/vi/blog/arabic-asr-wer-explained/) (translation of https://broadrouter.com/blog/arabic-asr-wer-explained/): WER là gì, tính thế nào, và vì sao WER tiếng Ả Rập rất nhạy với dấu nguyên âm, chuẩn hóa chính tả và phương ngữ. Kèm code Python và số liệu thật. - [Nhận dạng giọng nói phương ngữ Ả Rập: Vịnh, Ai Cập, Levant, Maghreb](https://broadrouter.com/vi/blog/arabic-dialect-speech-recognition/) (translation of https://broadrouter.com/blog/arabic-dialect-speech-recognition/): Vì sao chuyển giọng nói thành văn bản gặp khó với phương ngữ Ả Rập vùng Vịnh, Ai Cập, Levant, Maghreb, và cách xây bộ test theo phương ngữ cho âm thanh của bạn. - [Dịch Ả Rập–Anh bằng LLM: prompt, bảng thuật ngữ, đánh giá](https://broadrouter.com/vi/blog/arabic-english-translation-llm/) (translation of https://broadrouter.com/blog/arabic-english-translation-llm/): Dịch tiếng Ả Rập sang tiếng Anh và ngược lại bằng API LLM: prompt giữ giọng văn, bảng thuật ngữ cố định, số và tên riêng, và cách đánh giá chất lượng. - [Tiếng Ả Rập vùng Vịnh, Ai Cập và Levant: LLM xử lý phương ngữ ra sao](https://broadrouter.com/vi/blog/arabic-llm-gulf-egyptian-dialects/) (translation of https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/): LLM xử lý tiếng Ả Rập vùng Vịnh, Ai Cập và Levant thế nào: chỗ chúng sai, cách viết prompt cho phương ngữ, cách test, và BroadNet báo cáo gì về Qwen3.8 27B. - [OCR tiếng Ả Rập hay vision LLM: đọc tài liệu và hóa đơn tiếng Ả Rập](https://broadrouter.com/vi/blog/arabic-ocr-vision-llm/) (translation of https://broadrouter.com/blog/arabic-ocr-vision-llm/): Khi nào dùng OCR, khi nào dùng vision LLM cho hóa đơn, biên lai, biểu mẫu tiếng Ả Rập: bẫy độ chính xác, đầu ra JSON, chi phí mỗi trang và pipeline kiểm tra. - [Prompt engineering tiếng Ả Rập: văn phong, RTL, chữ số và văn bản trộn](https://broadrouter.com/vi/blog/arabic-prompt-engineering/) (translation of https://broadrouter.com/blog/arabic-prompt-engineering/): Viết prompt để LLM cho ra tiếng Ả Rập tốt: chọn văn phong và phương ngữ, kiểm soát chữ số và ngày tháng, giữ định dạng RTL sạch và xử lý văn bản trộn Ả Rập–Anh. - [RAG tiếng Ả Rập: chuẩn hóa, chia chunk và truy xuất hiệu quả](https://broadrouter.com/vi/blog/arabic-rag-search/) (translation of https://broadrouter.com/blog/arabic-rag-search/): RAG trên tài liệu tiếng Ả Rập: chuẩn hóa alef, ya và ta marbuta, chia chunk hợp lý, chọn embedding đa ngôn ngữ và kết hợp tìm kiếm từ khóa. - [API chuyển giọng nói tiếng Ả Rập thành văn bản: cách chọn năm 2026](https://broadrouter.com/vi/blog/arabic-speech-to-text-api-buyers-guide/) (translation of https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/): Cách chọn API chép lời tiếng Ả Rập: tỷ lệ lỗi từ, phương ngữ, giới hạn file, giá mỗi giờ âm thanh và quyền riêng tư, kèm kế hoạch test trên bản ghi của bạn. - [Kiến trúc chatbot WhatsApp tiếng Ả Rập: từ tin nhắn thoại đến trả lời](https://broadrouter.com/vi/blog/arabic-whatsapp-chatbot-architecture/) (translation of https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/): Kiến trúc chạy được cho trợ lý WhatsApp tiếng Ả Rập: webhook, hàng đợi, chép lời tin nhắn thoại, LLM trả lời, quy tắc chính sách và chi phí, kèm code Python. - [Arabizi và LLM: xử lý tiếng Ả Rập Franco với các số như 3 và 7](https://broadrouter.com/vi/blog/arabizi-llm-handling/) (translation of https://broadrouter.com/blog/arabizi-llm-handling/): Arabizi viết tiếng Ả Rập bằng chữ Latin, dùng số như 3 cho ع và 7 cho ح. Cách nhận diện, chuyển sang chữ Ả Rập bằng LLM và trả lời đúng cách. - [LLM tốt nhất cho Aider: cách chọn model và kết nối](https://broadrouter.com/vi/blog/best-llm-for-aider/) (translation of https://broadrouter.com/blog/best-llm-for-aider/): Cách chọn LLM tốt nhất cho Aider: tuân thủ edit format, kích thước context, chi phí mỗi giờ, bài thử một giờ trên repo của bạn và cấu hình OPENAI_API_BASE đúng. - [Model mở tốt nhất cho coding agent: chọn theo tool, context và chi phí](https://broadrouter.com/vi/blog/best-open-model-for-coding-agents/) (translation of https://broadrouter.com/blog/best-open-model-for-coding-agents/): GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B hay Gemma 4 cho coding agent của bạn? Chọn theo tool calling, context, suy luận và chi phí mỗi tác vụ, kèm phép tính. - [Token đầu vào từ cache: cách giảm hóa đơn API LLM](https://broadrouter.com/vi/blog/cached-input-tokens-explained/) (translation of https://broadrouter.com/blog/cached-input-tokens-explained/): Đầu vào từ cache chỉ tốn một phần nhỏ giá đầu vào thường. Cách tính phí prompt caching, cách đọc cached_tokens trong usage và cách cấu trúc prompt để cache trúng. - [API LLM vision rẻ nhất: chi phí mỗi ảnh và mỗi trang, tính cụ thể](https://broadrouter.com/vi/blog/cheapest-vision-llm-api/) (translation of https://broadrouter.com/blog/cheapest-vision-llm-api/): Một lần gọi LLM vision tốn bao nhiêu cho mỗi ảnh và mỗi trang tài liệu: phép tính token cho Gemma 4, Qwen3.8 27B và DeepSeek V4.1 Flash, và khi nào DeepInfra rẻ hơn. - [Cách chọn nhà cung cấp API LLM: checklist 20 điểm](https://broadrouter.com/vi/blog/choose-llm-api-provider-checklist/) (translation of https://broadrouter.com/blog/choose-llm-api-provider-checklist/): 20 câu hỏi cần đặt ra trước khi chọn API LLM: giá, giới hạn, tương thích OpenAI, quyền riêng tư, model, hỗ trợ và uptime, kèm bảng chấm điểm và kế hoạch thử nghiệm. - [Chọn API LLM cho tiếng Ả Rập: cần test những gì](https://broadrouter.com/vi/blog/choosing-an-llm-for-arabic/) (translation of https://broadrouter.com/blog/choosing-an-llm-for-arabic/): Cách chọn LLM cho ứng dụng tiếng Ả Rập: hiểu phương ngữ, chất lượng viết, dịch thuật, tài liệu và hình ảnh tiếng Ả Rập, chi phí, kèm bộ test làm xong trong một ngày. - [Cline, Roo Code hay Aider cho team dùng model mở (2026)](https://broadrouter.com/vi/blog/cline-vs-roo-code-vs-aider/) (translation of https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/): So sánh Cline, Roo Code (đã lưu trữ, tiếp nối bởi Zoo Code) và Aider theo những gì quan trọng với model mở: tool call, mức dùng context, MCP, git và phê duyệt. - [Cách đếm token trước khi gửi request tới LLM](https://broadrouter.com/vi/blog/count-tokens-before-sending/) (translation of https://broadrouter.com/blog/count-tokens-before-sending/): Ước tính số token của prompt theo ngôn ngữ và model trước khi gọi API LLM: vì sao tokenizer khác nhau, số đo thực tế với năm ngôn ngữ và biên max_tokens an toàn. - [10 cách giảm chi phí LLM cho coding agent, kèm mức tiết kiệm ước tính](https://broadrouter.com/vi/blog/cut-coding-agent-llm-costs/) (translation of https://broadrouter.com/blog/cut-coding-agent-llm-costs/): Mười cách cụ thể để giảm chi phí token của Cline, Aider, opencode hay coding agent tự viết, mỗi cách kèm phép tính mức tiết kiệm và cách đo kết quả. - [Thay thế API DeepSeek: 5 cách gọi model DeepSeek](https://broadrouter.com/vi/blog/deepseek-api-alternatives/) (translation of https://broadrouter.com/blog/deepseek-api-alternatives/): Gọi model DeepSeek qua API chính thức, aggregator, danh mục cloud, GPU của riêng bạn hoặc BroadNet Inference: đánh đổi về giá, quyền riêng tư và giới hạn. - [Giá API DeepSeek V4.1 Flash: theo token, đầu vào từ cache và gói](https://broadrouter.com/vi/blog/deepseek-v4-1-flash-api-pricing/) (translation of https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/): Giá API DeepSeek V4.1 Flash mỗi 1M token đầu vào, từ cache và đầu ra; gói cố định theo tháng, tuần, giờ; và so sánh với DeepSeek, DeepInfra, Fireworks. - [Cách ước tính chi phí API LLM trước khi ra mắt một tính năng](https://broadrouter.com/vi/blog/estimate-llm-api-cost-before-launch/) (translation of https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/): Mô hình chi phí từng bước cho một tính năng AI mới: người dùng × thao tác × lệnh gọi × token, kèm cache, suy luận, retry, giới hạn tốc độ và bảng ví dụ theo tháng. - [Cách đánh giá LLM trước khi đổi nhà cung cấp](https://broadrouter.com/vi/blog/evaluate-llm-before-switching/) (translation of https://broadrouter.com/blog/evaluate-llm-before-switching/): Dựng bộ eval nhỏ từ traffic thật, chấm bằng kiểm tra code và LLM chấm điểm đã giảm thiên lệch, rồi so sánh chất lượng, chi phí và độ trễ trước khi chuyển. - [Trích xuất hóa đơn, biên lai, biểu mẫu sang JSON bằng LLM vision](https://broadrouter.com/vi/blog/extract-data-from-documents-with-vision-llm/) (translation of https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/): Pipeline production cho hóa đơn, biên lai, biểu mẫu: chuẩn bị ảnh, prompt theo JSON schema, kiểm tra Pydantic, kiểm tra cộng tổng, retry và hàng đợi duyệt. - [API LLM giá cố định hay trả theo token: phép tính hòa vốn](https://broadrouter.com/vi/blog/flat-rate-vs-pay-per-token-llm-api/) (translation of https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/): Khi nào gói LLM giá cố định rẻ hơn trả theo token? Bốn ví dụ hóa đơn hằng tháng cho DeepSeek V4.1 Flash, khối lượng hòa vốn và cách tự đo mức dùng của bạn. - [Tín dụng API LLM miễn phí: khoản tín dụng ban đầu dùng được bao lâu](https://broadrouter.com/vi/blog/free-llm-api-credit/) (translation of https://broadrouter.com/blog/free-llm-api-credit/): Tín dụng API LLM miễn phí thực sự mua được gì: token, phút âm thanh và key theo giờ với tín dụng ban đầu của BroadNet, so với gói miễn phí của Gemini và OpenRouter. - [GDPR và API LLM: 15 câu hỏi cần đặt cho nhà cung cấp](https://broadrouter.com/vi/blog/gdpr-llm-api-checklist/) (translation of https://broadrouter.com/blog/gdpr-llm-api-checklist/): Checklist GDPR cho API LLM: DPA, bên xử lý phụ, lưu giữ, huấn luyện, chuyển dữ liệu quốc tế và thông báo vi phạm, kèm điều khoản GDPR sau mỗi câu hỏi. - [GLM 5.3 và DeepSeek V4.1 Flash: nên dùng model nào, khi nào](https://broadrouter.com/vi/blog/glm-5-3-vs-deepseek-v4-1-flash/) (translation of https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/): So sánh GLM 5.3 và DeepSeek V4.1 Flash về giá, context, tool calling, đầu ra có cấu trúc, hình ảnh và suy luận, kèm chi phí mẫu và quy tắc chọn đơn giản. - [Cách xử lý lỗi 429 của API LLM: backoff, jitter và giới hạn đồng thời](https://broadrouter.com/vi/blog/handle-llm-api-429-rate-limits/) (translation of https://broadrouter.com/blog/handle-llm-api-429-rate-limits/): 429 khác 5xx thế nào, exponential backoff với full jitter, giới hạn đồng thời và hàng đợi cho API LLM, kèm code Python và Node có thể đưa thẳng vào production. - [API key LLM theo giờ: giá cố định mỗi giờ, giải thích rõ](https://broadrouter.com/vi/blog/hourly-llm-api-key-explained/) (translation of https://broadrouter.com/blog/hourly-llm-api-key-explained/): Key theo giờ của BroadNet cho bạn một giờ dùng Gemma 4 VL 26B trên một luồng với giá cố định. Gồm những gì, giới hạn ra sao, hợp với ai và khi nào thì có lợi. - [So sánh lưu giữ dữ liệu API LLM: OpenAI, Anthropic, Gemini, Mistral](https://broadrouter.com/vi/blog/llm-api-data-retention-compared/) (translation of https://broadrouter.com/blog/llm-api-data-retention-compared/): OpenAI, Anthropic, Gemini API, Mistral và BroadNet chính thức nói gì về việc lưu prompt API và dùng chúng để huấn luyện, trích nguyên văn vào tháng 11/2026. - [Bảo mật API key LLM: lưu trữ, xoay vòng và danh sách IP cho phép](https://broadrouter.com/vi/blog/llm-api-key-security/) (translation of https://broadrouter.com/blog/llm-api-key-security/): Giữ API key LLM ngoài code và log, cấp key riêng cho từng workload, xoay vòng định kỳ, khóa key vào IP server và thu hồi key bị lộ chỉ trong vài giây. - [Giá API LLM: cách tính token đầu vào, cache, đầu ra và suy luận](https://broadrouter.com/vi/blog/llm-api-pricing-explained/) (translation of https://broadrouter.com/blog/llm-api-pricing-explained/): Giá API LLM theo token hoạt động thế nào: đầu vào, đầu vào từ cache, đầu ra, token suy luận, hình ảnh và âm thanh tính theo phút, kèm ví dụ từ bảng giá công khai. - [Gửi dữ liệu nhạy cảm tới API LLM: che, tối thiểu hóa, ghi log ít](https://broadrouter.com/vi/blog/llm-api-sensitive-data/) (translation of https://broadrouter.com/blog/llm-api-sensitive-data/): Cách gửi dữ liệu khách hàng, sức khỏe hay tài chính tới API LLM an toàn: phân loại trường, che bằng placeholder đảo ngược được, không lưu giữ dữ liệu, log sạch. - [Kiểm soát chi phí API LLM: ngân sách, giới hạn key và chặn vòng lặp](https://broadrouter.com/vi/blog/llm-api-spend-control/) (translation of https://broadrouter.com/blog/llm-api-spend-control/): Chặn hóa đơn API LLM tăng vọt: số dư trả trước, mỗi ứng dụng một key, danh sách IP cho phép, log mức dùng, ngân sách trong code, trần vòng lặp, key giá cố định. - [Ghi log sử dụng API LLM: chi phí theo người dùng, tính năng, ẩn PII](https://broadrouter.com/vi/blog/llm-api-usage-logging/) (translation of https://broadrouter.com/blog/llm-api-usage-logging/): Ghi log mọi lời gọi API LLM với token, chi phí, độ trễ và request ID, phân bổ chi tiêu theo người dùng và tính năng, và ẩn PII trước khi ghi. Có code Python. - [Chi phí LLM mỗi người dùng: unit economics cho tính năng AI SaaS](https://broadrouter.com/vi/blog/llm-cost-per-user-saas/) (translation of https://broadrouter.com/blog/llm-cost-per-user-saas/): Tính chi phí một tính năng AI trên mỗi người dùng đang hoạt động mỗi tháng, vì sao người dùng nặng quan trọng hơn mức trung bình, và cách định giá các hạng gói SaaS. - [So sánh giá API model mở: 6 nhà cung cấp inference tính phí ra sao](https://broadrouter.com/vi/blog/llm-inference-provider-pricing-compared/) (translation of https://broadrouter.com/blog/llm-inference-provider-pricing-compared/): OpenRouter, Together AI, Fireworks, Groq, DeepInfra và BroadNet tính giá model mở thế nào: theo token, cache, batch, giờ GPU và gói cố định, kèm nguồn giá. - [Độ trễ LLM: TTFT, token mỗi giây và thời gian end-to-end](https://broadrouter.com/vi/blog/llm-latency-ttft-throughput/) (translation of https://broadrouter.com/blog/llm-latency-ttft-throughput/): Thời gian tới token đầu tiên, token mỗi giây và độ trễ end-to-end đo gì, cách đo từng chỉ số từ phía client bằng code, và cách tối ưu nào tác động tới chỉ số nào. - [Giải thích streaming LLM: chunk SSE, delta, usage và hủy](https://broadrouter.com/vi/blog/llm-streaming-sse-explained/) (translation of https://broadrouter.com/blog/llm-streaming-sse-explained/): Streaming chat completion thực sự truyền đi thế nào, cách ghép câu trả lời từ delta, đọc usage ở chunk cuối và hủy gọn gàng bằng Python, Node và curl. - [Long context và RAG: khi 1M token thay được vector DB](https://broadrouter.com/vi/blog/long-context-vs-rag/) (translation of https://broadrouter.com/blog/long-context-vs-rag/): Khi nào context window 1M token có thể thay thế vector database và khi nào RAG vẫn thắng, so sánh theo chi phí mỗi truy vấn, độ trễ, recall và công bảo trì. - [Quản lý context cho agent chạy dài với cửa sổ 1M token](https://broadrouter.com/vi/blog/long-running-agent-context-management/) (translation of https://broadrouter.com/blog/long-running-agent-context-management/): Giữ phiên agent dài chính xác và rẻ với cửa sổ 1M token: lịch sử chỉ nối thêm, prompt thân thiện với cache, khi nào nén và nên cắt gì trước. - [Agent chạy dài: timeout, retry và checkpoint đứng vững](https://broadrouter.com/vi/blog/long-running-agents-timeouts-retries/) (translation of https://broadrouter.com/blog/long-running-agents-timeouts-retries/): Giữ một agent LLM chạy hàng giờ: timeout cho từng lệnh gọi, streaming để phát hiện treo, retry an toàn, tác dụng phụ idempotent của tool và checkpoint sau mỗi bước. - [Dùng MCP server với model tương thích OpenAI (mọi nhà cung cấp)](https://broadrouter.com/vi/blog/mcp-with-openai-compatible-models/) (translation of https://broadrouter.com/blog/mcp-with-openai-compatible-models/): MCP chạy được với mọi LLM: client biến công cụ MCP thành function call. Cách dùng MCP với model mở trong Cline, Continue, opencode, n8n hoặc code Python của bạn. - [Khi nào context window 1M token đáng dùng (và khi nào không)](https://broadrouter.com/vi/blog/million-token-context-window-guide/) (translation of https://broadrouter.com/blog/million-token-context-window-guide/): Những gì vừa trong 1.048.576 token, chi phí trên DeepSeek V4.1 Flash, suy luận và cache thay đổi phép tính ra sao, và khi nào RAG vẫn là lựa chọn tốt hơn. - [Chi phí thật của multi-agent: trả theo token hay phiên giá cố định](https://broadrouter.com/vi/blog/multi-agent-cost/) (translation of https://broadrouter.com/blog/multi-agent-cost/): Hệ thống multi-agent đốt token gấp nhiều lần một agent. Phép tính phiên cho gói giá cố định và trả theo token, các hệ số nhân ẩn, và khi nào mỗi cách thắng. - [API chuyển giọng nói thành văn bản: 14 ngôn ngữ, một endpoint](https://broadrouter.com/vi/blog/multilingual-speech-to-text-api/) (translation of https://broadrouter.com/blog/multilingual-speech-to-text-api/): Chép lời tiếng Anh, Tây Ban Nha, Pháp, Trung, Việt, Ả Rập và 8 ngôn ngữ khác qua một endpoint tương thích OpenAI giá theo phút, và khi nào dùng model tiếng Ả Rập. - [n8n AI Agent với LLM giá rẻ: cài đặt, model, chi phí mỗi lần chạy](https://broadrouter.com/vi/blog/n8n-ai-agent-cheap-llm/) (translation of https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/): Xây workflow n8n AI Agent trên model tương thích OpenAI giá thấp: chọn model nào cho tool, thiết lập nào giảm token, và phép tính chi phí cho 10.000 lần chạy. - [LLM open-weight hay đóng cho doanh nghiệp: chi phí, kiểm soát, rủi ro](https://broadrouter.com/vi/blog/open-weight-vs-closed-models/) (translation of https://broadrouter.com/blog/open-weight-vs-closed-models/): LLM open-weight hay model đóng cho doanh nghiệp? So sánh chi phí, kiểm soát, quyền riêng tư, khoảng cách chất lượng, sự phụ thuộc và ba cách chạy model mở. - [Chuyển sang API tương thích OpenAI: checklist khi chuyển đổi](https://broadrouter.com/vi/blog/openai-compatible-api-migration-checklist/) (translation of https://broadrouter.com/blog/openai-compatible-api-migration-checklist/): Đổi base_url là phần dễ. Thứ hay hỏng là model ID, token suy luận, usage khi streaming, lỗi, giới hạn tốc độ và key. Checklist kèm code cho từng bước. - [OpenRouter hay gói giá cố định cho AI agent: bên nào rẻ hơn?](https://broadrouter.com/vi/blog/openrouter-alternative-flat-rate/) (translation of https://broadrouter.com/blog/openrouter-alternative-flat-rate/): OpenRouter tính theo token cộng phí mua credit; gói cố định tính theo phiên. Chi phí cho agent chạy liên tục, phí và giới hạn của OpenRouter, và bên nào phù hợp. - [Cách chạy coding agent qua đêm không cần giám sát mà vẫn an toàn](https://broadrouter.com/vi/blog/overnight-coding-agent-runs/) (translation of https://broadrouter.com/blog/overnight-coding-agent-runs/): Thiết lập thực tế để chạy coding agent không giám sát: sandbox, checkpoint git, tệp nhiệm vụ, giới hạn số bước, review buổi sáng và phiên gói giá cố định qua đêm. - [Cách cấu trúc prompt để phần đầu được cache](https://broadrouter.com/vi/blog/prompt-structure-for-caching/) (translation of https://broadrouter.com/blog/prompt-structure-for-caching/): Cách sắp xếp prompt LLM để phần đầu được cache luôn giống hệt: system prompt, tool schema, tài liệu, lịch sử chỉ nối thêm, cùng một phép thử để chứng minh điều đó. - [Token suy luận: vì sao bạn trả cho nhiều đầu ra hơn những gì thấy](https://broadrouter.com/vi/blog/reasoning-tokens-cost/) (translation of https://broadrouter.com/blog/reasoning-tokens-cost/): Model suy luận tính phần suy nghĩ ẩn như token đầu ra. Cách đo, chi phí theo từng model, và cách cắt giảm bằng reasoning_effort, max_tokens và định tuyến. - [Cách chạy coding agent song song trên API LLM giá cố định](https://broadrouter.com/vi/blog/run-parallel-coding-agents-flat-rate/) (translation of https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/): Sub-agent, orchestrator và batch worker trên DeepSeek V4.1 Flash không giới hạn: cần bao nhiêu phiên, chia phiên cho các key và giới hạn request chồng lấn. - [Tự host LLM hay dùng API: chi phí thực và điểm hòa vốn](https://broadrouter.com/vi/blog/self-hosting-llm-vs-api-cost/) (translation of https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/): Giá thuê GPU từ trang chính thức, công thức tính chi phí mỗi 1M token trên GPU của bạn, mức sử dụng cần đạt để rẻ hơn API, và chi phí vận hành hay bị quên. - [So sánh giá API chuyển giọng nói thành văn bản: giá mỗi phút năm 2026](https://broadrouter.com/vi/blog/speech-to-text-api-pricing-compared/) (translation of https://broadrouter.com/blog/speech-to-text-api-pricing-compared/): Giá mỗi phút và mỗi 1.000 giờ của OpenAI, Google, AWS, Deepgram, AssemblyAI, Speechmatics, Groq và BroadNet Transcribe từ trang chính thức, kèm các khoản phụ phí. - [JSON mode và json_schema: đầu ra có cấu trúc đáng tin cậy](https://broadrouter.com/vi/blog/structured-output-json-schema/) (translation of https://broadrouter.com/blog/structured-output-json-schema/): JSON mode cho JSON hợp lệ; json_schema cho đúng cấu trúc. Cách yêu cầu từng loại trên model tương thích OpenAI, validate bằng Pydantic và sửa đầu ra lỗi. - [Số token mỗi từ: tiếng Ả Rập và tiếng Anh, vì sao tiếng Ả Rập đắt hơn](https://broadrouter.com/vi/blog/tokens-per-word-arabic-english/) (translation of https://broadrouter.com/blog/tokens-per-word-arabic-english/): Vì sao cùng một nội dung, tiếng Ả Rập thường tốn nhiều token LLM hơn tiếng Anh, vì sao số token mỗi từ gây hiểu lầm, và script đo tỷ lệ của bạn từ trường usage. - [Tool calling với model mở: schema, gọi song song, retry](https://broadrouter.com/vi/blog/tool-calling-open-models/) (translation of https://broadrouter.com/blog/tool-calling-open-models/): Function calling trên model mở tương thích OpenAI: mẹo viết schema, gọi công cụ song song, năm kiểu lỗi của tool call và một vòng lặp Python có validate và retry. - [Chép lời ghi âm tổng đài: pipeline sẵn sàng cho tiếng Ả Rập](https://broadrouter.com/vi/blog/transcribe-call-center-recordings/) (translation of https://broadrouter.com/blog/transcribe-call-center-recordings/): Tách cuộc gọi stereo theo kênh, cắt lượt nói, chép lời tiếng Ả Rập và tiếng Anh, che PII và tóm tắt từng cuộc gọi bằng LLM. Pipeline Python chạy được, kèm chi phí. - [Dùng model mở trong coding agent: Aider, Cline, opencode, Codex](https://broadrouter.com/vi/blog/use-open-models-in-coding-agents/) (translation of https://broadrouter.com/blog/use-open-models-in-coding-agents/): Trỏ Aider, Cline, Roo Code, opencode, Continue và Codex CLI tới bất kỳ endpoint tương thích OpenAI nào. Tên field chính xác cho từng công cụ và thiết lập hay hỏng. - [Tin nhắn thoại thành văn bản cho doanh nghiệp: WhatsApp, Telegram](https://broadrouter.com/vi/blog/voice-notes-to-text-for-business/) (translation of https://broadrouter.com/blog/voice-notes-to-text-for-business/): Biến tin nhắn thoại của khách hàng từ WhatsApp Business và bot Telegram thành bản chép lời và ticket hỗ trợ: tải về, chép lời, trích xuất bằng LLM, định tuyến. - [Gói LLM theo tuần cho dự án ngắn: Orbit, theo giờ hay theo tháng](https://broadrouter.com/vi/blog/weekly-llm-plan-short-projects/) (translation of https://broadrouter.com/blog/weekly-llm-plan-short-projects/): Khi nào gói DeepSeek V4.1 Flash không giới hạn một tuần có lợi hơn phiên theo giờ, gói tháng hay trả theo token, kèm phép tính cho hackathon, sprint và migration. - [“Không giới hạn” trong gói API LLM thực sự có nghĩa là gì](https://broadrouter.com/vi/blog/what-unlimited-llm-api-really-means/) (translation of https://broadrouter.com/blog/what-unlimited-llm-api-really-means/): Đồng hồ đo token, cửa sổ cuộn 5 giờ, trần tuần, giới hạn tốc độ và đồng thời. Bảy câu hỏi để biết gói AI “không giới hạn” có thật sự vô hạn với agent của bạn. - [Thay thế Whisper API: khi nào nên chuyển và cách test](https://broadrouter.com/vi/blog/whisper-api-alternative/) (translation of https://broadrouter.com/blog/whisper-api-alternative/): Khi nào nên dùng một API được host thay cho OpenAI Whisper API: tỷ lệ lỗi từ trên âm thanh của bạn, ngôn ngữ, giá mỗi giờ, giới hạn và quyền riêng tư, kèm code test. - [API LLM không lưu giữ dữ liệu: checklist 10 câu hỏi](https://broadrouter.com/vi/blog/zero-data-retention-llm-api-checklist/) (translation of https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/): “Không lưu giữ dữ liệu” nên có nghĩa gì với API LLM: prompt, đầu ra, log, huấn luyện, bên xử lý phụ và kiểm tra. Mười câu cần hỏi trước khi gửi dữ liệu. - [Guides in Vietnamese](https://broadrouter.com/vi/guides/) - [Dùng OpenAI Python SDK với BroadNet Inference](https://broadrouter.com/vi/guides/openai-python-sdk/) (translation of https://broadrouter.com/guides/openai-python-sdk/): Trỏ OpenAI Python SDK chính thức tới BroadNet Inference chỉ bằng cách đổi base_url. Cài đặt, đặt key, gửi request đầu tiên và stream câu trả lời. - [Dùng OpenAI Node.js SDK với BroadNet Inference](https://broadrouter.com/vi/guides/openai-node-sdk/) (translation of https://broadrouter.com/guides/openai-node-sdk/): Trỏ OpenAI SDK chính thức cho Node.js và TypeScript tới BroadNet Inference bằng baseURL. Cài đặt, gửi request đầu tiên, stream token và ghi log request ID. - [Dùng thử BroadNet Inference ngay trên trình duyệt, không cần tài khoản](https://broadrouter.com/vi/guides/try-broadnet-inference-playground/) (translation of https://broadrouter.com/guides/try-broadnet-inference-playground/): Thử chuyển giọng nói tiếng Ả Rập thành văn bản, đọc tài liệu và chat với Gemma 4 trên playground của BroadNet, không cần đăng ký hay thẻ, rồi chuyển sang API. - [Gửi request đầu tiên tới BroadNet Inference bằng curl](https://broadrouter.com/vi/guides/curl-quickstart/) (translation of https://broadrouter.com/guides/curl-quickstart/): Gọi BroadNet Inference lần đầu bằng curl: một chat completion, một phản hồi streaming và một bản chép lời âm thanh, kèm cách đọc phản hồi, usage và lỗi. - [Streaming chat completions và xử lý khi mất kết nối](https://broadrouter.com/vi/guides/streaming-chat-completions/) (translation of https://broadrouter.com/guides/streaming-chat-completions/): Cách streaming hoạt động trên BroadNet Inference: Server-Sent Events, dấu [DONE], thông tin sử dụng token trong chunk cuối và cần làm gì khi stream kết thúc sớm. - [Vì sao model suy luận trả về câu trả lời rỗng](https://broadrouter.com/vi/guides/reasoning-tokens-empty-response/) (translation of https://broadrouter.com/guides/reasoning-tokens-empty-response/): DeepSeek V4.1 Flash, GLM 5.3 và Qwen3.8 27B suy luận trước khi trả lời, và phần suy luận được tính vào max_tokens. Cách để không còn nhận phản hồi rỗng. - [Mã lỗi BroadNet Inference và khi nào nên thử lại](https://broadrouter.com/vi/guides/errors-and-retries/) (translation of https://broadrouter.com/guides/errors-and-retries/): Ý nghĩa từng mã trạng thái HTTP của BroadNet Inference API, lỗi nào nên thử lại với backoff, lỗi nào cần sửa trước và gửi gì cho bộ phận hỗ trợ kèm request ID. - [Khóa API key vào server của bạn bằng danh sách IP cho phép](https://broadrouter.com/vi/guides/api-keys-ip-allow-lists/) (translation of https://broadrouter.com/guides/api-keys-ip-allow-lists/): Bảo vệ key BroadNet Inference bằng danh sách IP cho phép, tách key inference khỏi thông tin xác thực portal và thu hồi key bị lộ trong vài giây, kèm ví dụ curl. - [Gói LLM API không giới hạn cho agent hoạt động thế nào?](https://broadrouter.com/vi/guides/unlimited-llm-api-sessions/) (translation of https://broadrouter.com/guides/unlimited-llm-api-sessions/): Các gói DeepSeek V4.1 Flash không giới hạn không đếm token và không có giới hạn tuần. Đây là cách phiên, agent song song, key và thanh toán thực sự hoạt động. - [Chạy Aider trên DeepSeek V4.1 Flash không giới hạn](https://broadrouter.com/vi/guides/aider-unlimited-deepseek/) (translation of https://broadrouter.com/guides/aider-unlimited-deepseek/): Kết nối coding agent Aider với BroadNet Inference qua thiết lập tương thích OpenAI và chạy trên gói DeepSeek V4.1 Flash không giới hạn với giá cố định. - [Dùng LangChain với BroadNet Inference](https://broadrouter.com/vi/guides/langchain-openai-base-url/) (translation of https://broadrouter.com/guides/langchain-openai-base-url/): Kết nối ChatOpenAI của LangChain với BroadNet Inference bằng base_url và api_key, và biết LangChain giữ lại gì, bỏ đi gì trong phản hồi của model suy luận. - [Dùng Cline với BroadNet Inference (provider OpenAI Compatible)](https://broadrouter.com/vi/guides/cline-openai-compatible/) (translation of https://broadrouter.com/guides/cline-openai-compatible/): Thiết lập extension Cline trong VS Code với BroadNet Inference qua provider OpenAI Compatible: các trường cần điền, model ID, context và cách sửa lỗi thường gặp. - [Dùng Roo Code với BroadNet Inference (OpenAI Compatible)](https://broadrouter.com/vi/guides/roo-code-openai-compatible/) (translation of https://broadrouter.com/guides/roo-code-openai-compatible/): Roo Code đã ngừng hoạt động vào tháng 5/2026. Cấu hình cho bản cài sẵn và các fork, cùng lựa chọn thay thế (Cline hoặc fork Zoo Code) với BroadNet Inference. - [Dùng Continue với BroadNet Inference (config.yaml và apiBase)](https://broadrouter.com/vi/guides/continue-dev-openai-compatible/) (translation of https://broadrouter.com/guides/continue-dev-openai-compatible/): Thêm model BroadNet Inference vào extension Continue cho IDE bằng provider openai và apiBase trong config.yaml, kèm capabilities tool_use và độ dài context. - [Thêm BroadNet Inference vào opencode làm custom provider](https://broadrouter.com/vi/guides/opencode-custom-provider/) (translation of https://broadrouter.com/guides/opencode-custom-provider/): Thêm BroadNet Inference vào coding agent opencode làm custom provider tương thích OpenAI trong opencode.jsonc, kèm alias model, capabilities và limit. - [Kết nối Open WebUI với BroadNet Inference](https://broadrouter.com/vi/guides/open-webui-openai-api/) (translation of https://broadrouter.com/guides/open-webui-openai-api/): Thêm BroadNet Inference vào Open WebUI như một kết nối tương thích OpenAI, thêm model ID vào allowlist và sửa lỗi xác minh, phản hồi rỗng và giới hạn tần suất. - [Thêm BroadNet Inference vào LibreChat dưới dạng custom endpoint](https://broadrouter.com/vi/guides/librechat-custom-endpoint/) (translation of https://broadrouter.com/guides/librechat-custom-endpoint/): Thêm BroadNet Inference vào LibreChat làm custom endpoint tương thích OpenAI trong librechat.yaml: key trong .env, danh sách model, model đặt tiêu đề, sửa lỗi. - [Dùng BroadNet Inference trong n8n (OpenAI credential, Base URL)](https://broadrouter.com/vi/guides/n8n-openai-compatible/) (translation of https://broadrouter.com/guides/n8n-openai-compatible/): Trỏ OpenAI credential của n8n tới BroadNet Inference bằng trường Base URL, rồi dùng node OpenAI Chat Model theo ID với nút Use Responses API đã tắt. - [Dùng Vercel AI SDK với BroadNet Inference](https://broadrouter.com/vi/guides/vercel-ai-sdk-openai-compatible/) (translation of https://broadrouter.com/guides/vercel-ai-sdk-openai-compatible/): Gọi BroadNet Inference từ Vercel AI SDK bằng createOpenAICompatible trong @ai-sdk/openai-compatible: generateText, streamText, tool và reasoning effort. - [API chuyển giọng nói tiếng Ả Rập thành văn bản: từng bước](https://broadrouter.com/vi/guides/arabic-speech-to-text-api/) (translation of https://broadrouter.com/guides/arabic-speech-to-text-api/): Chuyển âm thanh tiếng Ả Rập thành văn bản với BroadNet Transcribe Arabic qua endpoint tương thích OpenAI: định dạng, trường language, ví dụ Python, curl và giá. - [Kết nối LiteLLM tới BroadNet Inference (SDK và proxy)](https://broadrouter.com/vi/guides/litellm-proxy-custom-provider/) (translation of https://broadrouter.com/guides/litellm-proxy-custom-provider/): Gọi BroadNet Inference qua LiteLLM bằng tiền tố model openai/ và api_base, trong Python SDK hoặc config.yaml của proxy, với giới hạn tần suất khớp key. - [Chạy OpenAI Agents SDK trên BroadNet Inference](https://broadrouter.com/vi/guides/openai-agents-sdk-custom-provider/) (translation of https://broadrouter.com/guides/openai-agents-sdk-custom-provider/): Dùng OpenAI Agents SDK cho Python với BroadNet Inference qua AsyncOpenAI và OpenAIChatCompletionsModel, đồng thời tắt tracing vốn cần key OpenAI. - [Chép lời tin nhắn thoại WhatsApp thành văn bản bằng API](https://broadrouter.com/vi/guides/transcribe-whatsapp-voice-notes/) (translation of https://broadrouter.com/guides/transcribe-whatsapp-voice-notes/): Chuyển tin nhắn thoại WhatsApp .ogg hoặc bản ghi âm điện thoại .m4a sang WAV bằng ffmpeg, rồi chép lời bằng tiếng Ả Rập hoặc 13 ngôn ngữ khác với một lệnh gọi API. - [Chép lời tin nhắn thoại WhatsApp trong n8n](https://broadrouter.com/vi/guides/n8n-whatsapp-voice-note-transcription/) (translation of https://broadrouter.com/guides/n8n-whatsapp-voice-note-transcription/): Workflow n8n nhận tin nhắn thoại WhatsApp, tải file từ Meta, chuyển đổi bằng ffmpeg rồi chép lời bằng tiếng Ả Rập hoặc 13 ngôn ngữ khác. - [Đọc hóa đơn và biên lai tiếng Ả Rập bằng model thị giác](https://broadrouter.com/vi/guides/read-arabic-invoices-with-vision/) (translation of https://broadrouter.com/guides/read-arabic-invoices-with-vision/): Gửi ảnh hóa đơn hoặc biên lai tới model thị giác trên BroadNet Inference và nhận lại tổng tiền, ngày và nhà cung cấp dạng JSON, hỏi bằng tiếng Ả Rập hoặc tiếng Anh. - [Dùng Zed với BroadNet Inference làm provider tương thích OpenAI](https://broadrouter.com/vi/guides/zed-openai-compatible-provider/) (translation of https://broadrouter.com/guides/zed-openai-compatible-provider/): Thêm BroadNet Inference vào Zed làm provider tương thích OpenAI trong settings.json, kèm model ID, giới hạn token, capabilities và biến môi trường chứa key. - [Dùng Kilo Code với BroadNet Inference (provider OpenAI Compatible)](https://broadrouter.com/vi/guides/kilo-code-openai-compatible/) (translation of https://broadrouter.com/guides/kilo-code-openai-compatible/): Kết nối Kilo Code với BroadNet Inference làm custom provider OpenAI Compatible, qua hộp thoại VS Code hoặc kilo.json, với model ID, giới hạn và cờ tool. - [Thêm BroadNet Inference vào goose làm custom provider](https://broadrouter.com/vi/guides/goose-custom-provider/) (translation of https://broadrouter.com/guides/goose-custom-provider/): Kết nối AI agent goose với BroadNet Inference dưới dạng custom provider OpenAI Compatible, từ goose Desktop, lệnh goose configure hoặc một file JSON. - [Dùng Pydantic AI với BroadNet Inference (OpenAICompatibleProvider)](https://broadrouter.com/vi/guides/pydantic-ai-openai-compatible/) (translation of https://broadrouter.com/guides/pydantic-ai-openai-compatible/): Chạy agent Pydantic AI trên BroadNet Inference với OpenAIChatModel và OpenAICompatibleProvider, kèm gọi tool, đầu ra có cấu trúc và chỉnh model profile. - [Dùng CrewAI với BroadNet Inference làm endpoint OpenAI tùy chỉnh](https://broadrouter.com/vi/guides/crewai-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/crewai-custom-openai-endpoint/): Trỏ agent CrewAI tới BroadNet Inference bằng LLM(custom_openai=True, base_url=...), rồi chọn model cho tool, context dài và gói giá cố định. - [Dùng LlamaIndex với BroadNet Inference qua OpenAILike](https://broadrouter.com/vi/guides/llamaindex-openailike/) (translation of https://broadrouter.com/guides/llamaindex-openailike/): Kết nối LlamaIndex với BroadNet Inference bằng lớp OpenAILike, đặt đúng api_base, context_window, is_chat_model và is_function_calling_model cho từng model. - [Thêm BroadNet Inference vào Dify bằng plugin OpenAI-API-compatible](https://broadrouter.com/vi/guides/dify-openai-api-compatible/) (translation of https://broadrouter.com/guides/dify-openai-api-compatible/): Kết nối Dify với BroadNet Inference qua plugin model provider OpenAI-API-compatible, với giá trị chính xác cho từng trường, kích thước context và function calling. - [Dùng Jan với BroadNet Inference làm provider tương thích OpenAI](https://broadrouter.com/vi/guides/jan-custom-endpoint/) (translation of https://broadrouter.com/guides/jan-custom-endpoint/): Thêm BroadNet Inference vào ứng dụng desktop Jan làm custom provider tương thích OpenAI, tự thêm model ID và bật tool, vision cho từng model. - [Kết nối SillyTavern với BroadNet Inference (Custom tương thích OpenAI)](https://broadrouter.com/vi/guides/sillytavern-custom-openai-endpoint/) (translation of https://broadrouter.com/guides/sillytavern-custom-openai-endpoint/): Kết nối SillyTavern với BroadNet Inference qua nguồn Chat Completion Custom (OpenAI-compatible): model ID đúng, xử lý hậu kỳ prompt và giới hạn token. - [Dùng Spring AI với BroadNet Inference (OpenAI base-url)](https://broadrouter.com/vi/guides/spring-ai-openai-base-url/) (translation of https://broadrouter.com/guides/spring-ai-openai-base-url/): Trỏ OpenAI chat starter của Spring AI tới BroadNet Inference bằng spring.ai.openai.base-url, một model ID BroadNet và max-tokens, rồi gọi từ ChatClient. - [Gọi BroadNet Inference từ Go bằng SDK openai-go](https://broadrouter.com/vi/guides/openai-go-sdk/) (translation of https://broadrouter.com/guides/openai-go-sdk/): Dùng SDK chính thức openai-go v3 với BroadNet Inference qua option.WithBaseURL: chat completions, streaming kèm usage và xử lý lỗi trong Go. - [Xây agent LangGraph trên BroadNet Inference](https://broadrouter.com/vi/guides/langgraph-openai-compatible/) (translation of https://broadrouter.com/guides/langgraph-openai-compatible/): Chạy StateGraph hoặc create_agent của LangGraph trên BroadNet Inference bằng ChatOpenAI, base_url và model ID BroadNet, kèm tool, streaming và bộ nhớ hội thoại. - [Dùng pipeline và agent Haystack với BroadNet Inference](https://broadrouter.com/vi/guides/haystack-openai-compatible/) (translation of https://broadrouter.com/guides/haystack-openai-compatible/): Trỏ OpenAIChatGenerator của Haystack tới BroadNet Inference bằng api_base_url, một model ID BroadNet và key Secret, cho pipeline, tool, streaming và agent. - [Dùng Semantic Kernel với BroadNet Inference](https://broadrouter.com/vi/guides/semantic-kernel-openai-compatible/) (translation of https://broadrouter.com/guides/semantic-kernel-openai-compatible/): Kết nối Microsoft Semantic Kernel với BroadNet Inference: trong Python qua client AsyncOpenAI, trong C# qua endpoint tùy chỉnh, cho chat, plugin và agent. - [Dùng agent Mastra với BroadNet Inference](https://broadrouter.com/vi/guides/mastra-openai-compatible/) (translation of https://broadrouter.com/guides/mastra-openai-compatible/): Chạy agent Mastra trên BroadNet Inference bằng một đối tượng model đặt id, url và apiKey, kèm tiền tố broadnet/ lặp đôi mà model ID cần có. - [Lấy đầu ra có cấu trúc bằng Instructor trên BroadNet Inference](https://broadrouter.com/vi/guides/instructor-structured-output/) (translation of https://broadrouter.com/guides/instructor-structured-output/): Trích xuất đối tượng Pydantic từ BroadNet Inference bằng Instructor, dùng from_openai hoặc from_provider với base_url tùy chỉnh, và chọn đúng Mode. - [Dùng DSPy với BroadNet Inference (đặt model_type="chat")](https://broadrouter.com/vi/guides/dspy-openai-compatible/) (translation of https://broadrouter.com/guides/dspy-openai-compatible/): Cấu hình dspy.LM cho BroadNet Inference với tiền tố openai/, api_base và model_type="chat", vì DSPy 3.5 gửi model OpenAI tới Responses API. - [Đánh giá các model BroadNet bằng promptfoo](https://broadrouter.com/vi/guides/promptfoo-openai-compatible/) (translation of https://broadrouter.com/guides/promptfoo-openai-compatible/): Chạy eval promptfoo và so sánh model song song trên BroadNet Inference với provider openai:chat, apiBaseUrl và apiKeyEnvar, kèm cấu hình grader. - [Chạy smolagents trên BroadNet Inference với OpenAIModel](https://broadrouter.com/vi/guides/smolagents-openai-compatible/) (translation of https://broadrouter.com/guides/smolagents-openai-compatible/): Kết nối smolagents của Hugging Face với BroadNet Inference qua OpenAIModel và api_base: CodeAgent trên mọi model, ToolCallingAgent trên GLM 5.3 và Qwen3.8 27B. - [Kết nối Flowise với BroadNet Inference (OpenAI Custom Model)](https://broadrouter.com/vi/guides/flowise-openai-compatible/) (translation of https://broadrouter.com/guides/flowise-openai-compatible/): Dùng BroadNet Inference trong chatflow Flowise bằng node OpenAI Custom Model, trường Base Path, một OpenAI API credential và tên model BroadNet tự gõ. - [Kết nối Langflow với BroadNet Inference](https://broadrouter.com/vi/guides/langflow-openai-compatible/) (translation of https://broadrouter.com/guides/langflow-openai-compatible/): Dùng BroadNet Inference trong Langflow qua trường OpenAI API Base của component OpenAI và tên model tự gõ, và vì sao provider OpenAI Compatible không hiện model. - [Glossary entries in Vietnamese](https://broadrouter.com/vi/glossary/) - [AI agent (tác tử AI)](https://broadrouter.com/vi/glossary/agent/) (translation of https://broadrouter.com/glossary/agent/): AI agent là hệ thống trong đó model ngôn ngữ tự điều khiển các bước để đạt mục tiêu: nó quyết định gọi tool nào, đọc kết quả và lặp lại cho tới khi đưa ra câu trả lời cuối cùng hoặc chạm giới hạn. Script cố định gọi model ở các điểm định sẵn thường gọi là workflow. - [API key (khóa API)](https://broadrouter.com/vi/glossary/api-key/) (translation of https://broadrouter.com/glossary/api-key/): API key là một chuỗi bí mật dùng để xác định và cấp quyền cho client gọi API. API LLM thường yêu cầu nó trong header Authorization: Bearer, và bất kỳ ai giữ key đều có thể gửi request được tính phí vào tài khoản sở hữu key đó. - [Dấu phụ tiếng Ả Rập (tashkeel)](https://broadrouter.com/vi/glossary/arabic-diacritics/) (translation of https://broadrouter.com/glossary/arabic-diacritics/): Dấu phụ tiếng Ả Rập, hay tashkeel, là các dấu nhỏ viết trên hoặc dưới chữ cái để biểu thị nguyên âm ngắn, phụ âm kép và việc không có nguyên âm. Chúng không bắt buộc: phần lớn văn bản, kể cả tin tức, bỏ chúng, nên người đọc và phần mềm phải đoán nguyên âm từ ngữ cảnh. - [Nhận dạng phương ngữ tiếng Ả Rập](https://broadrouter.com/vi/glossary/arabic-dialect-identification/) (translation of https://broadrouter.com/glossary/arabic-dialect-identification/): Nhận dạng phương ngữ tiếng Ả Rập là bài toán xác định một văn bản hay bản ghi âm thuộc biến thể nào, như Ai Cập, vùng Vịnh, Levant, Maghreb hay tiếng Ả Rập chuẩn hiện đại, ở cấp vùng, quốc gia hoặc thành phố. Nó dùng để định tuyến tin nhắn, chọn prompt và chia tập đánh giá. - [Chuẩn hóa văn bản tiếng Ả Rập](https://broadrouter.com/vi/glossary/arabic-text-normalization/) (translation of https://broadrouter.com/glossary/arabic-text-normalization/): Chuẩn hóa văn bản tiếng Ả Rập là bước đưa các cách viết khác nhau của cùng một từ về một dạng trước khi so khớp: xóa dấu phụ và tatweel, thống nhất các dạng alef, thường đổi ى thành ي, ة thành ه. Chỉ áp dụng cho khóa tìm kiếm và đánh giá, không cho văn bản hiển thị. - [Arabizi](https://broadrouter.com/vi/glossary/arabizi/) (translation of https://broadrouter.com/glossary/arabizi/): Arabizi là tiếng Ả Rập viết không chính thức bằng chữ Latin, dùng chữ số cho những âm không có trong bảng chữ Latin, như 3 thay cho ع. Nó phổ biến trên chat và mạng xã hội, chủ yếu là phương ngữ, không có chính tả chuẩn và thường trộn với tiếng Anh hoặc tiếng Pháp. - [AWQ (activation-aware weight quantization)](https://broadrouter.com/vi/glossary/awq/) (translation of https://broadrouter.com/glossary/awq/): AWQ (activation-aware weight quantization) là phương pháp lượng tử hóa sau huấn luyện, nén trọng số LLM xuống độ chính xác thấp, thường là số nguyên 4-bit. Nó dùng thống kê activation từ một tập hiệu chuẩn nhỏ để tìm kênh trọng số quan trọng nhất và scale chúng để bớt mất độ chính xác, không cần lan truyền ngược. - [Base URL](https://broadrouter.com/vi/glossary/base-url/) (translation of https://broadrouter.com/glossary/base-url/): Base URL là địa chỉ gốc cố định của một API, như https://api.example.com/v1, mà client đặt trước mọi đường dẫn endpoint, như /chat/completions. Trong SDK tương thích OpenAI, đổi base URL cùng với API key và model ID sẽ trỏ cùng một đoạn code sang nhà cung cấp khác. - [Batch API](https://broadrouter.com/vi/glossary/batch-api/) (translation of https://broadrouter.com/glossary/batch-api/): Batch API là endpoint LLM bất đồng bộ nhận nhiều request cùng lúc, thường dưới dạng file, xử lý chúng trong một khung thời gian định sẵn thay vì ngay lập tức, rồi trả kết quả cùng nhau. Các nhà cung cấp như OpenAI và Anthropic tính giá batch thấp hơn API tiêu chuẩn. - [BM25](https://broadrouter.com/vi/glossary/bm25/) (translation of https://broadrouter.com/glossary/bm25/): BM25 (Okapi BM25) là hàm xếp hạng cho tìm kiếm từ khóa, chấm điểm tài liệu dựa trên tần suất các từ trong truy vấn xuất hiện trong đó, độ hiếm của các từ ấy trong cả bộ tài liệu và độ dài tài liệu. Đây là hàm similarity mặc định của Lucene và Elasticsearch. - [Byte pair encoding (BPE)](https://broadrouter.com/vi/glossary/bpe/) (translation of https://broadrouter.com/glossary/bpe/): Byte pair encoding (BPE) là thuật toán tách token theo subword, bắt đầu từ từng ký tự hoặc byte rồi liên tục gộp cặp liền kề xuất hiện nhiều nhất thành một mục từ vựng mới, tới khi đạt kích thước từ vựng mục tiêu. Phần lớn tokenizer LLM hiện nay dùng BPE hoặc biến thể cấp byte của nó. - [Token đầu vào từ cache (cached input tokens)](https://broadrouter.com/vi/glossary/cached-input-tokens/) (translation of https://broadrouter.com/glossary/cached-input-tokens/): Token đầu vào từ cache là token prompt mà nhà cung cấp LLM vừa xử lý gần đây và có thể dùng lại từ cache thay vì tính toán lại, thường là phần đầu lặp lại như system prompt hoặc tài liệu. Nhà cung cấp tính giá chúng thấp hơn token đầu vào mới. - [Chain-of-thought (CoT)](https://broadrouter.com/vi/glossary/chain-of-thought/) (translation of https://broadrouter.com/glossary/chain-of-thought/): Chain-of-thought (CoT, chuỗi suy nghĩ) là kỹ thuật prompt trong đó model ngôn ngữ được cho xem ví dụ hoặc được yêu cầu viết các bước suy luận trung gian trước câu trả lời cuối. Nó tăng độ chính xác với bài toán nhiều bước như số học và logic, và model suy luận giờ tự sinh ra nó. - [Tỷ lệ lỗi ký tự (CER, character error rate)](https://broadrouter.com/vi/glossary/character-error-rate/) (translation of https://broadrouter.com/glossary/character-error-rate/): Character error rate (CER, tỷ lệ lỗi ký tự) đo độ chính xác của bản chép lời hoặc OCR ở cấp ký tự: số phép thay thế, xóa và chèn cần để biến đầu ra thành văn bản tham chiếu, chia cho số ký tự trong văn bản tham chiếu. Càng thấp càng tốt, 0 là khớp hoàn toàn. - [Chat Completions API](https://broadrouter.com/vi/glossary/chat-completions-api/) (translation of https://broadrouter.com/glossary/chat-completions-api/): Chat Completions API là endpoint HTTP của OpenAI, POST /v1/chat/completions, nhận tên model và danh sách message có gắn vai trò rồi trả về message tiếp theo của model. Cấu trúc request và response của nó đã trở thành định dạng chung mà các nhà cung cấp và SDK tương thích OpenAI dùng lại. - [Chat template](https://broadrouter.com/vi/glossary/chat-template/) (translation of https://broadrouter.com/glossary/chat-template/): Chat template là quy tắc định dạng, thường viết bằng Jinja và đi kèm tokenizer của model, chuyển danh sách message có vai trò thành một chuỗi token duy nhất mà model cần, kể cả các token điều khiển đặc biệt. Dùng sai template có thể làm câu trả lời của model kém đi rõ rệt. - [Chunking (RAG)](https://broadrouter.com/vi/glossary/chunking/) (translation of https://broadrouter.com/glossary/chunking/): Chunking là bước trong pipeline truy xuất chia tài liệu thành các mảnh nhỏ hơn, gọi là chunk, trước khi embed và index. Chunk size quyết định mỗi mảnh chứa bao nhiêu văn bản, còn chunk overlap lặp lại một phần văn bản giữa các chunk liền kề để ý bị cắt ở ranh giới không bị mất. - [Code-switching (chuyển mã ngôn ngữ)](https://broadrouter.com/vi/glossary/code-switching/) (translation of https://broadrouter.com/glossary/code-switching/): Code-switching là việc dùng hai hay nhiều ngôn ngữ, hoặc nhiều biến thể của một ngôn ngữ, trong cùng một cuộc hội thoại hay một câu, khi nói hoặc viết. Ví dụ: người nói tiếng Ả Rập chêm từ tiếng Anh vào phương ngữ, hoặc chuyển qua lại giữa tiếng Ả Rập chuẩn hiện đại và phương ngữ của mình. - [AI coding agent](https://broadrouter.com/vi/glossary/coding-agent/) (translation of https://broadrouter.com/glossary/coding-agent/): AI coding agent là công cụ phần mềm trong đó một model ngôn ngữ làm việc trên codebase theo vòng lặp: đọc file, sửa file, chạy lệnh như test, đọc kết quả và quyết định bước tiếp theo cho tới khi xong task hoặc chạm giới hạn. - [Concurrency (request đồng thời)](https://broadrouter.com/vi/glossary/concurrency/) (translation of https://broadrouter.com/glossary/concurrency/): Concurrency (request đồng thời) là số request mà một client đang xử lý cùng lúc. Các API LLM thường giới hạn nó theo key hoặc theo gói, tách biệt với số request mỗi phút, vì mỗi request đang mở chiếm dụng năng lực GPU cho tới khi response kết thúc. - [Context compaction (nén context)](https://broadrouter.com/vi/glossary/context-compaction/) (translation of https://broadrouter.com/glossary/context-compaction/): Context compaction là cách tóm tắt một cuộc hội thoại sắp chạm giới hạn context window rồi tiếp tục phiên làm việc từ bản tóm tắt đó, để agent giữ lại các quyết định và trạng thái quan trọng trong khi bỏ đi lịch sử thô như kết quả tool cũ. - [Context engineering (kỹ thuật quản lý context)](https://broadrouter.com/vi/glossary/context-engineering/) (translation of https://broadrouter.com/glossary/context-engineering/): Context engineering là công việc chọn và duy trì tập token mà model ngôn ngữ thấy ở mỗi request, gồm system prompt, định nghĩa tool, lịch sử hội thoại, tài liệu truy xuất và ghi chú, để model nhận được tập thông tin nhỏ nhất nhưng giàu tín hiệu nhất cần cho tác vụ. - [Context rot (suy giảm khi context dài)](https://broadrouter.com/vi/glossary/context-rot/) (translation of https://broadrouter.com/glossary/context-rot/): Context rot là sự suy giảm dần về độ chính xác và độ tin cậy của model ngôn ngữ khi input dài ra, ngay cả khi tác vụ không đổi và input vẫn nằm gọn trong context window. Thuật ngữ này phổ biến nhờ một báo cáo kỹ thuật năm 2025 của Chroma thử nghiệm 18 model. - [Context window (cửa sổ ngữ cảnh)](https://broadrouter.com/vi/glossary/context-window/) (translation of https://broadrouter.com/glossary/context-window/): Context window là số token tối đa mà model ngôn ngữ xử lý được trong một request. Nó bao gồm prompt, lịch sử hội thoại, định nghĩa tool, hình ảnh và đầu ra được sinh ra, kể cả token suy luận. Phần không vừa phải được cắt bớt, tóm tắt hoặc chia ra nhiều request. - [Data residency (lưu trú dữ liệu)](https://broadrouter.com/vi/glossary/data-residency/) (translation of https://broadrouter.com/glossary/data-residency/): Data residency (lưu trú dữ liệu) là yêu cầu hoặc cam kết rằng dữ liệu được lưu trữ và xử lý tại một quốc gia hay khu vực cụ thể. Với LLM API, nó gồm nơi xử lý prompt khi inference và nơi đặt nội dung đã lưu, thường không gồm dữ liệu tài khoản và thanh toán. - [Phương ngữ tiếng Ả Rập (dialectal Arabic)](https://broadrouter.com/vi/glossary/dialectal-arabic/) (translation of https://broadrouter.com/glossary/dialectal-arabic/): Phương ngữ tiếng Ả Rập là tiếng Ả Rập nói hằng ngày của một vùng hay thành phố, như tiếng Ả Rập Ai Cập, vùng Vịnh, Levant hay Maghreb. Chúng khác MSA về âm, từ vựng và ngữ pháp, không có chính tả chuẩn và ngày càng được viết trên mạng, nên khó hơn với model ngôn ngữ. - [Speaker diarization (phân tách người nói)](https://broadrouter.com/vi/glossary/diarization/) (translation of https://broadrouter.com/glossary/diarization/): Speaker diarization (phân tách người nói) là tác vụ chia một bản ghi âm thành các đoạn theo người nói, trả lời câu hỏi “ai nói lúc nào”. Mỗi đoạn lời nói được gắn một ID ẩn danh như SPEAKER_00, và thường được kết hợp với bản chép lời để cho thấy ai nói từng dòng. - [Chưng cất tri thức (knowledge distillation)](https://broadrouter.com/vi/glossary/distillation/) (translation of https://broadrouter.com/glossary/distillation/): Distillation, hay chưng cất tri thức (knowledge distillation), là phương pháp huấn luyện trong đó một model trò nhỏ hơn học tái tạo đầu ra của một model thầy lớn hơn, như phân phối xác suất hay câu trả lời sinh ra, để model trò giữ được phần lớn chất lượng của model thầy với chi phí chạy thấp hơn. - [DPA (thỏa thuận xử lý dữ liệu)](https://broadrouter.com/vi/glossary/dpa/) (translation of https://broadrouter.com/glossary/dpa/): DPA (thỏa thuận xử lý dữ liệu) là hợp đồng giữa bên kiểm soát và bên xử lý về cách bên xử lý xử lý dữ liệu cá nhân. Điều 28(3) GDPR bắt buộc DPA và liệt kê nội dung: chỉ thị bằng văn bản, bảo mật, an ninh, quy tắc subprocessor, xóa hoặc trả dữ liệu, kiểm toán. - [Embedding (vector nhúng)](https://broadrouter.com/vi/glossary/embeddings/) (translation of https://broadrouter.com/glossary/embeddings/): Embedding là danh sách số thực (vector) mà một model tạo ra để biểu diễn đoạn văn bản, hình ảnh hay đầu vào khác. Đầu vào có nghĩa tương tự nhận vector gần nhau, nên khoảng cách giữa vector dùng được để xếp hạng kết quả tìm kiếm, gom cụm tài liệu hoặc tìm mục liên quan. - [Eval (đánh giá LLM)](https://broadrouter.com/vi/glossary/eval/) (translation of https://broadrouter.com/glossary/eval/): Eval (viết tắt của evaluation, đánh giá) là một bài kiểm thử lặp lại được cho model ngôn ngữ hoặc prompt: một tập input cố định, mô tả thế nào là đầu ra tốt, và một bộ chấm điểm (code, con người hoặc một model khác) chấm từng đầu ra để có thể so sánh các phiên bản. - [Exponential backoff (lùi thời gian theo cấp số nhân)](https://broadrouter.com/vi/glossary/exponential-backoff/) (translation of https://broadrouter.com/glossary/exponential-backoff/): Exponential backoff là chiến lược thử lại trong đó client chờ lâu hơn sau mỗi lần thất bại trước khi thử tiếp, thường nhân đôi độ trễ cho tới một mức trần. Thêm jitter ngẫu nhiên vào mỗi lần chờ giúp dàn trải các lần thử lại từ nhiều client để chúng không đổ dồn vào server cùng lúc. - [Few-shot prompting (prompt kèm ví dụ)](https://broadrouter.com/vi/glossary/few-shot-prompting/) (translation of https://broadrouter.com/glossary/few-shot-prompting/): Few-shot prompting là kỹ thuật trong đó prompt chứa một số ít ví dụ input-output mẫu trước input thật, để model ngôn ngữ suy ra tác vụ và định dạng từ đó. Trọng số của model không thay đổi; zero-shot không dùng ví dụ nào và one-shot dùng một ví dụ. - [Fine-tuning (tinh chỉnh model)](https://broadrouter.com/vi/glossary/fine-tuning/) (translation of https://broadrouter.com/glossary/fine-tuning/): Fine-tuning (tinh chỉnh) là quá trình tiếp tục huấn luyện một model ngôn ngữ đã được huấn luyện trước trên một tập dữ liệu nhỏ dành riêng cho tác vụ, như hội thoại mẫu, để trọng số thay đổi và model làm tác vụ hay phong cách đó ổn định hơn mà không cần chỉ dẫn dài trong mỗi prompt. - [finish_reason (lý do kết thúc)](https://broadrouter.com/vi/glossary/finish-reason/) (translation of https://broadrouter.com/glossary/finish-reason/): finish_reason là một trường trên mỗi choice trong response chat completion, cho biết vì sao model ngừng sinh. Trong định dạng OpenAI, các giá trị là stop, length, tool_calls, content_filter và function_call đã bị deprecated. Messages API của Anthropic báo cùng ý tưởng qua stop_reason. - [Gói LLM giá cố định (flat-rate LLM plan)](https://broadrouter.com/vi/glossary/flat-rate-llm-plan/) (translation of https://broadrouter.com/glossary/flat-rate-llm-plan/): Gói LLM giá cố định tính một mức giá cố định cho mỗi kỳ, ví dụ một tháng, để dùng API mô hình ngôn ngữ thay vì tính theo token. Gói này thường giới hạn thứ khác, như số request đồng thời, tốc độ, model hay khung giờ sử dụng, nên giới hạn quan trọng không kém giá. - [FP8 (số thực dấu phẩy động 8-bit)](https://broadrouter.com/vi/glossary/fp8/) (translation of https://broadrouter.com/glossary/fp8/): FP8 là định dạng số dấu phẩy động 8-bit dùng trong deep learning, có hai cách mã hóa: E4M3 với 4 bit số mũ và 3 bit phần định trị, và E5M2 với 5 bit số mũ và 2 bit phần định trị. Nó dùng nửa bộ nhớ của định dạng 16-bit và chạy native trên GPU đời mới. - [Frequency penalty và presence penalty](https://broadrouter.com/vi/glossary/frequency-and-presence-penalty/) (translation of https://broadrouter.com/glossary/frequency-and-presence-penalty/): Frequency penalty và presence penalty là các thiết lập lấy mẫu hạ điểm của những token đã có trong đầu ra. Presence penalty áp dụng một lần khi token đã xuất hiện; frequency penalty tăng dần theo mỗi lần lặp. Trong API của OpenAI, cả hai nằm trong khoảng -2.0 đến 2.0, mặc định là 0. - [GDPR (Quy định chung về bảo vệ dữ liệu của EU)](https://broadrouter.com/vi/glossary/gdpr/) (translation of https://broadrouter.com/glossary/gdpr/): GDPR (Quy định (EU) 2016/679) là luật của EU về xử lý dữ liệu cá nhân, áp dụng từ 25/5/2018 cho tổ chức thành lập tại EU và một số tổ chức ngoài EU. Với LLM API, GDPR điều chỉnh hợp đồng với bên xử lý, chuyển dữ liệu quốc tế và tối thiểu hóa dữ liệu. - [GGUF (định dạng file model của llama.cpp)](https://broadrouter.com/vi/glossary/gguf/) (translation of https://broadrouter.com/glossary/gguf/): GGUF là định dạng file nhị phân để lưu model ngôn ngữ phục vụ inference với các runtime dựa trên GGML như llama.cpp. Một file GGUF chứa trọng số, thường đã lượng tử hóa, cùng metadata cần để nạp model, gồm thông tin tokenizer và kiến trúc, và có thể được memory-map để nạp nhanh. - [GPU-hour (giờ GPU)](https://broadrouter.com/vi/glossary/gpu-hour/) (translation of https://broadrouter.com/glossary/gpu-hour/): GPU-hour (giờ GPU) là một bộ xử lý đồ họa được dùng trong một giờ. Nhà cung cấp cloud tính tiền instance GPU thuê theo giờ hoặc theo giây, còn nhà nghiên cứu dùng GPU-hour để nói một job tốn bao nhiêu tính toán. Bạn trả tiền cho thời gian của máy, dù nó có đang làm việc hay không. - [Grounding (neo câu trả lời vào nguồn)](https://broadrouter.com/vi/glossary/grounding/) (translation of https://broadrouter.com/glossary/grounding/): Grounding là cách kết nối đầu ra của model ngôn ngữ với các nguồn thông tin cụ thể, kiểm chứng được, như tài liệu, kết quả tìm kiếm hay bản ghi cơ sở dữ liệu, và yêu cầu model trả lời dựa trên đó. Câu trả lời có grounding truy được về nguồn và ít dữ kiện bịa đặt hơn. - [Guardrails (lớp bảo vệ cho LLM)](https://broadrouter.com/vi/glossary/guardrails/) (translation of https://broadrouter.com/glossary/guardrails/): Guardrails là các bước kiểm tra ứng dụng chạy quanh lời gọi model ngôn ngữ để giữ nó trong giới hạn. Guardrail input sàng lọc thứ đi vào model, guardrail output kiểm tra thứ đi ra, và guardrail tool xác thực hành động trước khi chạy. Kiểm tra thất bại thì request bị chặn, sửa hoặc gắn cờ. - [Ảo giác AI (hallucination)](https://broadrouter.com/vi/glossary/hallucination/) (translation of https://broadrouter.com/glossary/hallucination/): Ảo giác (hallucination) là đầu ra của model ngôn ngữ nghe hợp lý nhưng sai, bịa đặt hoặc không được đầu vào hỗ trợ, như trích dẫn bịa, con số sai hay API không tồn tại. Nó xuất phát từ việc model dự đoán văn bản khả dĩ thay vì kiểm chứng sự thật. - [Hamza (chữ hamza trong tiếng Ả Rập)](https://broadrouter.com/vi/glossary/hamza/) (translation of https://broadrouter.com/glossary/hamza/): Hamza (ء) là chữ cái tiếng Ả Rập biểu thị âm tắc thanh hầu. Nó đứng riêng hoặc nằm trên chữ đỡ: alef (أ إ), waw (ؤ) hoặc yeh (ئ), còn آ là alef kết hợp madda. Vì văn bản gõ tay viết hamza không nhất quán, tìm kiếm và chấm giọng nói thường gộp các dạng alef về ا. - [Hybrid search (tìm kiếm lai)](https://broadrouter.com/vi/glossary/hybrid-search/) (translation of https://broadrouter.com/glossary/hybrid-search/): Hybrid search là phương pháp truy xuất chạy cả tìm kiếm từ khóa như BM25 lẫn tìm kiếm vector trên embedding cho cùng một truy vấn, rồi gộp hai danh sách xếp hạng làm một, thường bằng reciprocal rank fusion (RRF) hoặc tổng có trọng số của điểm đã chuẩn hóa. - [Idempotency key (khóa idempotency)](https://broadrouter.com/vi/glossary/idempotency-key/) (translation of https://broadrouter.com/glossary/idempotency-key/): Idempotency key là một giá trị duy nhất, thường là ID ngẫu nhiên như UUID, mà client gắn vào request làm thay đổi trạng thái, chẳng hạn thanh toán hay mua hàng. Nếu cùng request được gửi lại với cùng key, server trả về kết quả ban đầu thay vì thực hiện hai lần. - [Image token (token hình ảnh)](https://broadrouter.com/vi/glossary/image-tokens/) (translation of https://broadrouter.com/glossary/image-tokens/): Token hình ảnh là các token đầu vào mà model ngôn ngữ thị giác dùng để biểu diễn một ảnh trong prompt. Nhà cung cấp tính chúng từ kích thước ảnh, thường bằng cách resize rồi chia thành patch hoặc tile cố định, và tính tiền như token văn bản, nên ảnh lớn tốn hơn. - [Nhà cung cấp inference (inference provider)](https://broadrouter.com/vi/glossary/inference-provider/) (translation of https://broadrouter.com/glossary/inference-provider/): Nhà cung cấp inference là công ty chạy model học máy, thường là LLM open-weight, trên GPU sở hữu hoặc thuê, rồi bán quyền truy cập qua API, thường tính theo token, request hoặc thời gian. Họ lo phần phục vụ, mở rộng quy mô và phần cứng. - [Token đầu vào và đầu ra (input and output tokens)](https://broadrouter.com/vi/glossary/input-and-output-tokens/) (translation of https://broadrouter.com/glossary/input-and-output-tokens/): Token đầu vào là token mà model ngôn ngữ đọc trong một request, gồm system prompt, lịch sử hội thoại, định nghĩa tool và hình ảnh. Token đầu ra là token nó sinh ra để trả lời, kể cả phần suy luận. API LLM thường tính giá riêng cho hai loại, và đầu ra thường đắt hơn. - [INT4 (lượng tử hóa 4-bit)](https://broadrouter.com/vi/glossary/int4/) (translation of https://broadrouter.com/glossary/int4/): INT4 là định dạng số nguyên 4-bit biểu diễn được 16 mức. Trong inference LLM, nó thường chỉ lượng tử hóa 4-bit chỉ áp dụng cho trọng số (W4A16): trọng số được lưu ở INT4 kèm hệ số scale và giải nén về 16-bit khi tính toán, tốn khoảng một phần tư bộ nhớ so với trọng số 16-bit. - [IP allow list (danh sách IP được phép)](https://broadrouter.com/vi/glossary/ip-allow-list/) (translation of https://broadrouter.com/glossary/ip-allow-list/): IP allow list là cơ chế bảo mật chỉ chấp nhận request từ các địa chỉ IP hoặc dải CIDR trong danh sách và từ chối mọi nguồn khác. Trên API, nó hạn chế thiệt hại khi key bị lộ, vì key không hoạt động từ các mạng không có trong danh sách. - [Jailbreak (LLM)](https://broadrouter.com/vi/glossary/jailbreak/) (translation of https://broadrouter.com/glossary/jailbreak/): Jailbreak là đầu vào được tạo ra để khiến một model ngôn ngữ bỏ qua huấn luyện an toàn hoặc các quy tắc mà ứng dụng triển khai nó đặt ra, nhằm sinh nội dung hay thực hiện hành động mà bình thường nó sẽ từ chối. OWASP coi jailbreak là một dạng prompt injection. - [JSON mode (chế độ JSON)](https://broadrouter.com/vi/glossary/json-mode/) (translation of https://broadrouter.com/glossary/json-mode/): JSON mode là một thiết lập của API LLM, response_format {"type": "json_object"} trong các API kiểu OpenAI, khiến câu trả lời của model là JSON hợp lệ, parse được. Khác với structured output dùng JSON Schema, nó không đảm bảo object chứa những key, kiểu hay giá trị nào. - [JSON Schema](https://broadrouter.com/vi/glossary/json-schema/) (translation of https://broadrouter.com/glossary/json-schema/): JSON Schema là một bộ từ vựng viết bằng JSON để mô tả và kiểm tra cấu trúc dữ liệu JSON: kiểu, trường bắt buộc, giá trị cho phép và cách lồng nhau. Phiên bản hiện tại là 2020-12. Các API LLM dùng nó để định nghĩa tham số tool và cấu trúc chính xác mà structured output phải theo. - [KV cache (bộ nhớ đệm key-value)](https://broadrouter.com/vi/glossary/kv-cache/) (translation of https://broadrouter.com/glossary/kv-cache/): KV cache là vùng bộ nhớ nơi model transformer lưu key và value attention đã tính cho mọi token đã xử lý, để mỗi token mới dùng lại thay vì tính lại cả chuỗi. Nó lớn dần theo độ dài context và thường giới hạn số request một GPU phục vụ được. - [Độ trễ (latency)](https://broadrouter.com/vi/glossary/latency/) (translation of https://broadrouter.com/glossary/latency/): Với API model ngôn ngữ, độ trễ (latency) là thời gian từ lúc gửi request tới khi nhận được token đầu ra cuối cùng. Nó bằng thời gian tới token đầu tiên cộng thời gian sinh, và thường được báo cáo theo percentile như p50 và p95 thay vì giá trị trung bình. - [LLM-as-a-judge (LLM làm giám khảo)](https://broadrouter.com/vi/glossary/llm-as-judge/) (translation of https://broadrouter.com/glossary/llm-as-judge/): LLM-as-a-judge là phương pháp đánh giá trong đó một model ngôn ngữ chấm đầu ra của model khác: chấm một câu trả lời theo rubric hoặc chọn câu tốt hơn trong hai câu. Nó gần với đánh giá của con người, rẻ hơn nhiều, nhưng có thiên lệch như ưu tiên câu đầu tiên hoặc câu dài hơn. - [Benchmark LLM (LLM benchmark)](https://broadrouter.com/vi/glossary/llm-benchmark/) (translation of https://broadrouter.com/glossary/llm-benchmark/): Benchmark LLM là một bộ tác vụ cố định, thường công khai, kèm cách chấm điểm, dùng để đo và so sánh model ngôn ngữ ở một kỹ năng như kiến thức, lập trình hay nhận dạng giọng nói. Điểm benchmark cho thấy năng lực tương đối trên bài kiểm tra đó, không phải hiệu quả trên workload của bạn. - [LLM router](https://broadrouter.com/vi/glossary/llm-router/) (translation of https://broadrouter.com/glossary/llm-router/): LLM router là phần mềm hoặc dịch vụ nằm giữa ứng dụng và một hay nhiều API model ngôn ngữ, quyết định với từng request model hoặc nhà cung cấp nào sẽ xử lý, dựa trên các quy tắc như chi phí, độ trễ, tính sẵn sàng hoặc loại tác vụ. - [Logprobs (log xác suất)](https://broadrouter.com/vi/glossary/logprobs/) (translation of https://broadrouter.com/glossary/logprobs/): Logprobs là log xác suất mà một model ngôn ngữ đã gán cho các token nó sinh ra. API trả về logprobs sẽ cho thấy, với mỗi token đầu ra, model đánh giá token đó khả năng cao đến đâu, và tùy chọn thêm các token thay thế khả năng cao nhất ở vị trí đó. - [LoRA (low-rank adaptation)](https://broadrouter.com/vi/glossary/lora/) (translation of https://broadrouter.com/glossary/lora/): LoRA (low-rank adaptation) là phương pháp fine-tune tiết kiệm tham số: đóng băng trọng số của model đã huấn luyện sẵn và chỉ huấn luyện một cặp ma trận low-rank nhỏ cho các layer được chọn. Phần cập nhật có thể lưu thành adapter nhỏ hoặc gộp vào trọng số, không thêm độ trễ inference. - [Lost in the middle (lạc ở giữa)](https://broadrouter.com/vi/glossary/lost-in-the-middle/) (translation of https://broadrouter.com/glossary/lost-in-the-middle/): Lost in the middle là xu hướng đã được quan sát ở các model ngôn ngữ: chúng dùng thông tin liên quan tốt nhất khi nó nằm gần đầu hoặc cuối một đầu vào dài, và kém nhất khi nó nằm ở giữa. Tên gọi xuất phát từ nghiên cứu năm 2023 của Liu và cộng sự. - [max_tokens (giới hạn token đầu ra)](https://broadrouter.com/vi/glossary/max-tokens/) (translation of https://broadrouter.com/glossary/max-tokens/): max_tokens là tham số của API LLM đặt số token tối đa mà model được sinh trong một phản hồi. Khi chạm giới hạn, việc sinh dừng lại và phản hồi được đánh dấu là bị cắt, ví dụ finish_reason được đặt là "length". - [MCP (Model Context Protocol)](https://broadrouter.com/vi/glossary/mcp/) (translation of https://broadrouter.com/glossary/mcp/): MCP (Model Context Protocol) là giao thức mở, xây trên JSON-RPC 2.0, chuẩn hóa cách ứng dụng LLM kết nối với tool và dữ liệu bên ngoài. MCP server cung cấp tool, resource và prompt; MCP client trong ứng dụng host tìm ra chúng và chuyển lệnh gọi tool của model tới server. - [Mixture of experts (MoE)](https://broadrouter.com/vi/glossary/mixture-of-experts/) (translation of https://broadrouter.com/glossary/mixture-of-experts/): Mixture of experts (MoE) là kiến trúc mạng nơ-ron trong đó một layer chứa nhiều mạng con expert và một router đã học chỉ gửi mỗi token tới vài expert trong số đó. Model ngôn ngữ MoE có tổng số tham số lớn nhưng mỗi token chỉ dùng một tập con active nhỏ. - [Model deprecation (ngừng hỗ trợ model)](https://broadrouter.com/vi/glossary/model-deprecation/) (translation of https://broadrouter.com/glossary/model-deprecation/): Model deprecation là khi nhà cung cấp API thông báo một model sẽ ngừng được cung cấp, thường kèm ngày ngừng hoạt động (retirement hoặc shutdown) và một model thay thế được khuyến nghị. Sau ngày đó, request tới model sẽ lỗi, nên ứng dụng phải chuyển sang ID model khác trước hạn. - [Model fallback (chuyển sang model dự phòng)](https://broadrouter.com/vi/glossary/model-fallback/) (translation of https://broadrouter.com/glossary/model-fallback/): Model fallback là mẫu thiết kế tăng độ tin cậy, trong đó request LLM bị lỗi được thử lại trên một model hoặc nhà cung cấp khác theo danh sách có thứ tự. Nó được kích hoạt bởi lỗi, rate limit, lỗi độ dài context hay từ chối do chính sách nội dung, và chạy ở client hoặc gateway. - [Giấy phép model (model license)](https://broadrouter.com/vi/glossary/model-license/) (translation of https://broadrouter.com/glossary/model-license/): Giấy phép model là điều khoản pháp lý quy định cách dùng, sửa đổi và chia sẻ trọng số, mã nguồn và đầu ra của model. Model trọng số mở dùng giấy phép phần mềm chuẩn như Apache 2.0, MIT, hoặc giấy phép riêng có thể thêm điều kiện, như giới hạn với dịch vụ thương mại lớn. - [Tiếng Ả Rập chuẩn hiện đại (MSA)](https://broadrouter.com/vi/glossary/msa/) (translation of https://broadrouter.com/glossary/msa/): Tiếng Ả Rập chuẩn hiện đại (MSA) là dạng tiếng Ả Rập chuẩn hóa dùng trong văn viết, tin tức, giáo dục và phát biểu trang trọng khắp thế giới Ả Rập. Không ai dùng nó làm tiếng mẹ đẻ hằng ngày: người ta nói phương ngữ như Ai Cập hay vùng Vịnh và chuyển sang MSA khi trang trọng. - [Hệ thống đa agent (multi-agent system)](https://broadrouter.com/vi/glossary/multi-agent-system/) (translation of https://broadrouter.com/glossary/multi-agent-system/): Hệ thống đa agent là ứng dụng LLM trong đó nhiều agent, mỗi agent có prompt, tool và context riêng, cùng xử lý các phần của một tác vụ. Thiết kế phổ biến là orchestrator-worker: agent chính lập kế hoạch, khởi chạy các worker song song và tổng hợp kết quả. - [AI đa phương thức (multimodal AI)](https://broadrouter.com/vi/glossary/multimodal/) (translation of https://broadrouter.com/glossary/multimodal/): AI đa phương thức là model hoặc hệ thống làm việc với nhiều loại dữ liệu, như văn bản, hình ảnh, video và audio. Một model đa phương thức có thể đọc ảnh cùng câu hỏi, hoặc nhận giọng nói và trả lời bằng văn bản. Vision-language model là loại phổ biến nhất. - [Nhận dạng thực thể có tên (NER)](https://broadrouter.com/vi/glossary/named-entity-recognition/) (translation of https://broadrouter.com/glossary/named-entity-recognition/): Nhận dạng thực thể có tên (NER) là tác vụ tìm các tên trong văn bản và gán cho mỗi tên một loại, như người, tổ chức hay địa điểm. Hệ thống cổ điển gán nhãn từng token; với LLM, người ta thường yêu cầu trả về danh sách thực thể dạng JSON rồi đối chiếu với văn bản gốc. - [Nhận dạng ký tự quang học (OCR)](https://broadrouter.com/vi/glossary/ocr/) (translation of https://broadrouter.com/glossary/ocr/): Nhận dạng ký tự quang học (OCR) là công nghệ phát hiện chữ in hoặc chữ viết tay trong ảnh, bản scan hay PDF và chuyển thành ký tự máy đọc được. OCR cổ điển trả về văn bản và vị trí trên trang; vision-language model có thể đọc chữ và trả lời câu hỏi về nó ngay. - [Ogg Opus](https://broadrouter.com/vi/glossary/ogg-opus/) (translation of https://broadrouter.com/glossary/ogg-opus/): Ogg Opus là audio được mã hóa bằng codec Opus (RFC 6716) và lưu trong container Ogg (RFC 7845). Nó được thiết kế cho giọng nói và audio tương tác ở bitrate thấp, và là định dạng WhatsApp dùng cho tin nhắn thoại, thường là tệp .ogg hoặc .opus. - [Model trọng số mở (open-weight model)](https://broadrouter.com/vi/glossary/open-weight-model/) (translation of https://broadrouter.com/glossary/open-weight-model/): Model trọng số mở (open-weight) là model học máy có trọng số đã huấn luyện được công bố để ai cũng tải về và chạy, thường kèm giấy phép có điều kiện. Nó khác AI mã nguồn mở hoàn toàn, vốn công bố thêm code huấn luyện và chi tiết dữ liệu huấn luyện để người khác tái tạo model. - [API tương thích OpenAI (OpenAI-compatible API)](https://broadrouter.com/vi/glossary/openai-compatible-api/) (translation of https://broadrouter.com/glossary/openai-compatible-api/): API tương thích OpenAI là một HTTP API nhận cùng định dạng request và phản hồi như OpenAI API, ví dụ POST /v1/chat/completions, nên SDK và công cụ của OpenAI dùng được nó chỉ bằng cách đổi base URL, API key và tên model. - [Gọi tool song song (parallel tool calls)](https://broadrouter.com/vi/glossary/parallel-tool-calls/) (translation of https://broadrouter.com/glossary/parallel-tool-calls/): Gọi tool song song (parallel tool calls) là khi model trả về nhiều lệnh gọi tool trong một phản hồi, ví dụ đọc ba tệp cùng lúc. Ở định dạng OpenAI, chúng là nhiều phần tử trong tool_calls, và đặt parallel_tool_calls bằng false giới hạn model ở không hoặc một lệnh gọi mỗi lượt. - [Trả theo mức dùng (pay-as-you-go)](https://broadrouter.com/vi/glossary/pay-as-you-go/) (translation of https://broadrouter.com/glossary/pay-as-you-go/): Trả theo mức dùng (pay-as-you-go) là mô hình giá mà bạn chỉ trả cho những gì đã dùng, đo theo đơn vị như token, request hoặc phút âm thanh, không có gói đăng ký hay cam kết tối thiểu. Hầu hết API LLM tính phí theo cách này, trừ vào số dư trả trước hoặc qua hóa đơn định kỳ. - [PDPL (luật bảo vệ dữ liệu Ả Rập Xê Út và UAE)](https://broadrouter.com/vi/glossary/pdpl/) (translation of https://broadrouter.com/glossary/pdpl/): PDPL (Personal Data Protection Law) là tên chung của luật dữ liệu cá nhân Ả Rập Xê Út (Sắc lệnh M/19, SDAIA thực thi) và Sắc luật Liên bang số 45/2021 của UAE. Cả hai điều chỉnh việc xử lý dữ liệu cá nhân và việc chuyển ra nước ngoài, kể cả prompt gửi API LLM nước ngoài. - [PII (thông tin nhận dạng cá nhân)](https://broadrouter.com/vi/glossary/pii/) (translation of https://broadrouter.com/glossary/pii/): PII (thông tin nhận dạng cá nhân) là thông tin giúp phân biệt hoặc truy ra danh tính một người, như tên, số định danh, dữ liệu sinh trắc, cùng thông tin gắn được với người đó. GDPR dùng khái niệm rộng hơn là dữ liệu cá nhân. Với LLM, che PII là gỡ nó khỏi prompt trước khi gửi. - [Prefill và decode](https://broadrouter.com/vi/glossary/prefill-and-decode/) (translation of https://broadrouter.com/glossary/prefill-and-decode/): Prefill và decode là hai giai đoạn của LLM inference. Ở prefill, model xử lý song song toàn bộ prompt và lấp đầy KV cache; ở decode, model sinh câu trả lời từng token một, dùng lại cache đó. Prefill chủ yếu quyết định time to first token; decode quyết định tốc độ đầu ra. - [Tín dụng trả trước (prepaid credit)](https://broadrouter.com/vi/glossary/prepaid-credit/) (translation of https://broadrouter.com/glossary/prepaid-credit/): Tín dụng trả trước là tiền nạp vào tài khoản API trước khi sử dụng. Mỗi request trừ chi phí của nó vào số dư, và khi số dư hết, các request tính theo mức dùng dừng lại cho tới khi bạn nạp thêm. Mức chi tiêu không vượt quá số tiền đã nạp. - [Prompt engineering (kỹ thuật viết prompt)](https://broadrouter.com/vi/glossary/prompt-engineering/) (translation of https://broadrouter.com/glossary/prompt-engineering/): Prompt engineering là quá trình viết, sắp xếp và kiểm thử chỉ dẫn đưa cho model ngôn ngữ để nó luôn tạo ra đầu ra đáp ứng yêu cầu. Nó bao gồm message vai trò, chỉ dẫn rõ ràng, ví dụ, định dạng và đo kết quả bằng các test case. - [Prompt injection (tấn công chèn prompt)](https://broadrouter.com/vi/glossary/prompt-injection/) (translation of https://broadrouter.com/glossary/prompt-injection/): Prompt injection là kiểu tấn công trong đó văn bản đưa vào language model, do người dùng nhập hoặc ẩn trong trang web, email, tệp hay kết quả tool, làm thay đổi hành vi của model trái với ý định của lập trình viên. OWASP xếp nó hạng nhất (LLM01:2025) trong Top 10 cho ứng dụng LLM. - [Lượng tử hóa (quantization)](https://broadrouter.com/vi/glossary/quantization/) (translation of https://broadrouter.com/glossary/quantization/): Lượng tử hóa (quantization) là quá trình lưu các con số của mạng nơ-ron, chủ yếu là trọng số, ở định dạng độ chính xác thấp hơn như 8-bit hoặc 4-bit thay vì 16-bit, để model dùng ít bộ nhớ hơn và thường chạy nhanh hơn. Phương pháp tốt giữ mức suy giảm độ chính xác nhỏ. - [RAG (tạo sinh tăng cường truy xuất)](https://broadrouter.com/vi/glossary/rag/) (translation of https://broadrouter.com/glossary/rag/): RAG (retrieval-augmented generation) là mô hình trong đó ứng dụng trước tiên truy xuất các đoạn liên quan tới câu hỏi từ một kho bên ngoài, như chỉ mục tìm kiếm hay chỉ mục vector, rồi đưa chúng vào prompt để model ngôn ngữ trả lời dựa trên văn bản đó thay vì chỉ dựa vào trí nhớ. - [Giới hạn tần suất (rate limit)](https://broadrouter.com/vi/glossary/rate-limit/) (translation of https://broadrouter.com/glossary/rate-limit/): Giới hạn tần suất (rate limit) là mức trần mà API đặt ra cho lượng sử dụng của một client trong một khoảng thời gian, như số request mỗi phút hoặc số token mỗi phút. Request vượt giới hạn bị từ chối, thường với HTTP 429 Too Many Requests, cho đến khi mức dùng giảm xuống dưới giới hạn. - [ReAct (reason and act)](https://broadrouter.com/vi/glossary/react-agent/) (translation of https://broadrouter.com/glossary/react-agent/): ReAct (reason and act) là mẫu prompting do Yao và cộng sự giới thiệu năm 2022, trong đó language model xen kẽ các bước suy luận với hành động như tìm kiếm hay gọi tool. Model đọc từng kết quả quan sát, cập nhật kế hoạch và lặp lại cho tới khi đưa ra được câu trả lời cuối. - [Token suy luận (reasoning tokens)](https://broadrouter.com/vi/glossary/reasoning-tokens/) (translation of https://broadrouter.com/glossary/reasoning-tokens/): Token suy luận (reasoning tokens) là token mà model suy luận sinh ra để giải quyết vấn đề trước khi viết câu trả lời hiển thị. Chúng thường được tính phí như token đầu ra, tính vào giới hạn đầu ra và context window, và có thể bị API ẩn đi hoặc tóm tắt. - [Request timeout (thời gian chờ request)](https://broadrouter.com/vi/glossary/request-timeout/) (translation of https://broadrouter.com/glossary/request-timeout/): Request timeout là thời gian tối đa một HTTP client chờ server trước khi bỏ request. Client thường chia nó thành connect timeout cho việc mở kết nối và read timeout cho việc chờ dữ liệu. Lệnh gọi LLM cần read timeout rộng rãi vì câu trả lời dài cần thời gian. - [Reranking (xếp hạng lại)](https://broadrouter.com/vi/glossary/reranking/) (translation of https://broadrouter.com/glossary/reranking/): Reranking là giai đoạn truy xuất thứ hai, trong đó một model chính xác hơn nhưng chậm hơn, thường là cross-encoder, chấm điểm từng cặp truy vấn và tài liệu ứng viên từ lượt tìm kiếm đầu rồi sắp xếp lại, để chỉ vài kết quả liên quan nhất được chuyển cho language model. - [Văn bản RTL (viết từ phải sang trái, right-to-left)](https://broadrouter.com/vi/glossary/rtl-text/) (translation of https://broadrouter.com/glossary/rtl-text/): Văn bản RTL (right-to-left) là chữ viết hiển thị từ phải sang trái, như tiếng Ả Rập hay tiếng Do Thái. Máy tính lưu nó theo thứ tự đọc logic; khi RTL trộn với chữ trái sang phải, chữ số, tên model hay code, Unicode Bidirectional Algorithm quyết định thứ tự hiển thị. - [Tần số lấy mẫu (sample rate)](https://broadrouter.com/vi/glossary/sample-rate/) (translation of https://broadrouter.com/glossary/sample-rate/): Tần số lấy mẫu (sample rate) là số mẫu âm thanh ghi được mỗi giây, đo bằng hertz (Hz). Model nhận dạng giọng nói thường cần audio mono 16.000 Hz (16 kHz). Cùng với bit depth và số kênh, sample rate quyết định dung lượng mỗi phút của file audio không nén. - [Tham số seed](https://broadrouter.com/vi/glossary/seed-parameter/) (translation of https://broadrouter.com/glossary/seed-parameter/): Tham số seed là một số nguyên bạn gửi kèm request để API cố gắng lấy mẫu cùng các token khi request giống hệt được lặp lại. OpenAI mô tả đây là best effort (nỗ lực tối đa): tính tất định không được đảm bảo, và trường system_fingerprint cho biết khi cấu hình backend đã thay đổi. - [Serverless inference (suy luận serverless)](https://broadrouter.com/vi/glossary/serverless-inference/) (translation of https://broadrouter.com/glossary/serverless-inference/): Serverless inference là cách chạy model machine learning trong đó nền tảng tự khởi động, co giãn và dừng tài nguyên tính toán sau một endpoint, thường về 0 khi rảnh, và tính phí theo mức dùng thay vì theo máy đặt trước. Đổi lại có thể có độ trễ cold start. - [SLA (thỏa thuận mức dịch vụ, service level agreement)](https://broadrouter.com/vi/glossary/sla/) (translation of https://broadrouter.com/glossary/sla/): SLA (thỏa thuận mức dịch vụ) là cam kết theo hợp đồng của nhà cung cấp về một mức dịch vụ đo được, thường là phần trăm uptime hằng tháng, kèm biện pháp khắc phục cụ thể, thường là service credit, nếu không đạt. SLA định nghĩa downtime, các loại trừ và cách yêu cầu bồi hoàn. - [Speculative decoding (giải mã suy đoán)](https://broadrouter.com/vi/glossary/speculative-decoding/) (translation of https://broadrouter.com/glossary/speculative-decoding/): Speculative decoding là kỹ thuật inference trong đó một model nháp nhỏ, nhanh đề xuất vài token tiếp theo, rồi model đích lớn kiểm tra chúng trong một lượt forward pass và giữ lại các token được chấp nhận. Nó tăng tốc sinh văn bản mà không thay đổi phân phối của output. - [Chuyển giọng nói thành văn bản (speech-to-text)](https://broadrouter.com/vi/glossary/speech-to-text/) (translation of https://broadrouter.com/glossary/speech-to-text/): Chuyển giọng nói thành văn bản, hay nhận dạng giọng nói tự động (ASR), là công nghệ biến lời nói thành chữ viết. API nhận tệp hoặc luồng âm thanh, trả về bản chép lời, thường tính theo phút. Độ chính xác đo bằng tỷ lệ lỗi từ, thay đổi theo giọng, phương ngữ và chất lượng ghi âm. - [Giới hạn chi tiêu (spend limit)](https://broadrouter.com/vi/glossary/spend-limit/) (translation of https://broadrouter.com/glossary/spend-limit/): Giới hạn chi tiêu (spend limit) là mức trần số tiền một tài khoản, project hoặc key API được tiêu trong một kỳ, thường là một tháng. Khi chạm mức này, nhà cung cấp từ chối request tiếp theo, thường bằng lỗi HTTP 429 hoặc 400, cho đến khi sang kỳ mới hoặc giới hạn được nâng. - [Phụ đề SRT (SRT subtitles)](https://broadrouter.com/vi/glossary/srt-subtitles/) (translation of https://broadrouter.com/glossary/srt-subtitles/): SRT (SubRip) là định dạng phụ đề văn bản thuần. Mỗi phụ đề gồm số thứ tự, thời gian bắt đầu và kết thúc viết dạng hh:mm:ss,mmm --> hh:mm:ss,mmm, một hoặc nhiều dòng chữ, và một dòng trống. WebVTT là định dạng web tương tự dùng cho video HTML. - [Server-sent events (SSE)](https://broadrouter.com/vi/glossary/sse/) (translation of https://broadrouter.com/glossary/sse/): Server-sent events (SSE) là một chuẩn web cho phép server đẩy một luồng sự kiện dạng văn bản tới client qua một HTTP response duy nhất, giữ mở lâu, với content type text/event-stream. Mỗi sự kiện gồm một hoặc nhiều dòng trường, như data:, và kết thúc bằng một dòng trống. - [Stop sequences (chuỗi dừng)](https://broadrouter.com/vi/glossary/stop-sequences/) (translation of https://broadrouter.com/glossary/stop-sequences/): Stop sequences (chuỗi dừng) là các chuỗi bạn gửi kèm request để model dừng sinh ngay khi tạo ra một trong số đó. OpenAI gọi trường này là stop và nhận tối đa 4 chuỗi; Anthropic gọi là stop_sequences. Phần văn bản khớp thường bị loại khỏi đầu ra trả về. - [Streaming (trả phản hồi dạng luồng)](https://broadrouter.com/vi/glossary/streaming/) (translation of https://broadrouter.com/glossary/streaming/): Streaming là cách trả phản hồi của model ngôn ngữ thành từng chunk nhỏ ngay trong lúc nó đang được sinh, thay vì một câu trả lời trọn vẹn ở cuối. Các API tương thích OpenAI stream qua server-sent events khi request đặt stream là true. - [Đầu ra có cấu trúc (structured output)](https://broadrouter.com/vi/glossary/structured-output/) (translation of https://broadrouter.com/glossary/structured-output/): Đầu ra có cấu trúc (structured output) là tính năng của API LLM buộc câu trả lời của model theo định dạng do lập trình viên chỉ định, thường là JSON Schema, để code parse được tin cậy. Nó chặt hơn JSON mode, vốn chỉ đảm bảo JSON hợp lệ chứ không đảm bảo một cấu trúc cụ thể. - [Subagent (agent con)](https://broadrouter.com/vi/glossary/subagent/) (translation of https://broadrouter.com/glossary/subagent/): Subagent là một agent riêng biệt mà agent chính khởi chạy để xử lý một việc cụ thể. Nó chạy với context window, system prompt và quyền dùng tool riêng, làm xong việc và chỉ trả kết quả hoặc bản tóm tắt cho agent chính, nhờ đó cuộc hội thoại chính luôn gọn. - [Bên xử lý phụ (subprocessor)](https://broadrouter.com/vi/glossary/subprocessor/) (translation of https://broadrouter.com/glossary/subprocessor/): Bên xử lý phụ (subprocessor) là bên thứ ba được bên xử lý thuê xử lý dữ liệu cá nhân thay bên kiểm soát. Với API LLM, đó có thể là GPU host, đối tác inference và capacity dự phòng. Theo Điều 28 GDPR, họ cần bên kiểm soát cho phép và chịu cùng nghĩa vụ bảo vệ dữ liệu. - [System prompt (lời nhắc hệ thống)](https://broadrouter.com/vi/glossary/system-prompt/) (translation of https://broadrouter.com/glossary/system-prompt/): System prompt là đoạn chỉ dẫn mà ứng dụng gửi cho model ngôn ngữ trước cuộc hội thoại, thường là tin nhắn có vai trò system hoặc developer. Nó đặt vai trò, quy tắc, giọng điệu và định dạng đầu ra, và model được kỳ vọng tuân theo nó hơn các tin nhắn người dùng mâu thuẫn. - [Taa marbuta (chữ ة tiếng Ả Rập)](https://broadrouter.com/vi/glossary/taa-marbuta/) (translation of https://broadrouter.com/glossary/taa-marbuta/): Taa marbuta (ة, U+0629) là chữ cái tiếng Ả Rập chỉ xuất hiện ở cuối từ, thường đánh dấu danh từ và tính từ giống cái. Nó trông như heh (ه, U+0647) có thêm hai chấm, nên văn bản gõ tay hay lẫn hai chữ này, và nhiều bộ chuẩn hóa tìm kiếm đổi ة thành ه. - [Temperature (nhiệt độ)](https://broadrouter.com/vi/glossary/temperature/) (translation of https://broadrouter.com/glossary/temperature/): Temperature là tham số sampling điều chỉnh mức ngẫu nhiên trong đầu ra của model ngôn ngữ. Model chia điểm của từng token cho temperature trước khi đổi thành xác suất: giá trị thấp làm token khả dĩ nhất chiếm ưu thế, giá trị cao làm phân phối phẳng hơn nên token ít khả dĩ được chọn thường hơn. - [Chuyển văn bản thành giọng nói (text-to-speech, TTS)](https://broadrouter.com/vi/glossary/text-to-speech/) (translation of https://broadrouter.com/glossary/text-to-speech/): Text-to-speech (TTS), hay tổng hợp giọng nói, chuyển văn bản viết thành âm thanh lời nói. Một API TTS nhận văn bản, một giọng đọc và định dạng output như MP3 hoặc WAV, rồi trả về giọng nói được sinh ra. Nó là chiều ngược lại của speech-to-text, vốn chuyển âm thanh thành văn bản. - [Throughput (thông lượng)](https://broadrouter.com/vi/glossary/throughput/) (translation of https://broadrouter.com/glossary/throughput/): Throughput (thông lượng), trong việc phục vụ LLM, là tổng khối lượng công việc một hệ thống hoàn thành trong một đơn vị thời gian trên mọi request đồng thời, thường đo bằng token đầu ra mỗi giây hoặc request mỗi giây. Nó khác với tốc độ từng luồng, tức tốc độ một request nhận token của nó. - [Thời gian tới token đầu tiên (TTFT)](https://broadrouter.com/vi/glossary/time-to-first-token/) (translation of https://broadrouter.com/glossary/time-to-first-token/): Thời gian tới token đầu tiên (TTFT) là khoảng thời gian từ lúc gửi request tới model ngôn ngữ cho tới khi token đầu ra đầu tiên đến. Nó gồm thời gian mạng, thời gian xếp hàng trên server và prefill, bước model đọc toàn bộ prompt, nên tăng theo độ dài prompt và tải. - [Timestamp trong speech-to-text (dấu thời gian ASR)](https://broadrouter.com/vi/glossary/timestamps-asr/) (translation of https://broadrouter.com/glossary/timestamps-asr/): Timestamp trong speech-to-text là thời điểm bắt đầu và kết thúc, tính bằng giây, gắn với từng phần của bản phiên âm. Timestamp theo đoạn bao một cụm từ hoặc câu; timestamp theo từ gắn với từng từ. Chúng cần cho phụ đề, tìm kiếm trong bản ghi âm và căn văn bản theo người nói. - [Token](https://broadrouter.com/vi/glossary/token/) (translation of https://broadrouter.com/glossary/token/): Token là đơn vị văn bản mà model ngôn ngữ đọc và viết: một từ, một phần của từ, một dấu câu hoặc một chuỗi byte do tokenizer của model tạo ra. API LLM đếm, giới hạn và thường tính phí theo token. - [Tokenizer (bộ tách token)](https://broadrouter.com/vi/glossary/tokenizer/) (translation of https://broadrouter.com/glossary/tokenizer/): Tokenizer là thành phần chuyển văn bản thành chuỗi token ID số nguyên mà mô hình ngôn ngữ đọc, và chuyển các ID được sinh ra trở lại thành văn bản. Mỗi model có tokenizer riêng, nên cùng một văn bản có thể cho ra số token khác nhau trên các model khác nhau. - [Số token mỗi giây (tokens per second, TPS)](https://broadrouter.com/vi/glossary/tokens-per-second/) (translation of https://broadrouter.com/glossary/tokens-per-second/): Số token mỗi giây (TPS) là tốc độ model ngôn ngữ sinh token đầu ra, thường đo cho một request sau khi token đầu tiên tới. Nó quyết định câu trả lời stream nhanh đến đâu. Cộng trên mọi request đồng thời, nó thành thông lượng hệ thống; thời gian chờ trước token đầu tiên được đo riêng. - [Tool calling (gọi tool, function calling)](https://broadrouter.com/vi/glossary/tool-calling/) (translation of https://broadrouter.com/glossary/tool-calling/): Tool calling, còn gọi là function calling, là tính năng của API LLM trong đó model trả về một yêu cầu có cấu trúc để chạy một hàm do nhà phát triển định nghĩa, kèm đối số JSON. Ứng dụng chạy hàm rồi gửi kết quả lại; model không bao giờ tự chạy code. - [tool_choice (tham số chọn tool)](https://broadrouter.com/vi/glossary/tool-choice/) (translation of https://broadrouter.com/glossary/tool-choice/): tool_choice là tham số request điều khiển việc dùng tool. Trong Chat Completions của OpenAI, giá trị có thể là auto (model tự quyết), required (phải gọi ít nhất một tool), none (không dùng tool), hoặc tên một function cụ thể để ép gọi. Các tùy chọn tương ứng của Anthropic là auto, any, tool và none. - [Top-p (nucleus sampling)](https://broadrouter.com/vi/glossary/top-p/) (translation of https://broadrouter.com/glossary/top-p/): Top-p, còn gọi là nucleus sampling, là phương pháp giải mã trong đó model ngôn ngữ chỉ lấy mẫu token tiếp theo từ tập nhỏ nhất gồm các token khả năng cao nhất có tổng xác suất đạt ngưỡng p, ví dụ 0,9, và bỏ phần đuôi ít khả năng. - [Chuẩn hóa Unicode (Unicode normalization)](https://broadrouter.com/vi/glossary/unicode-normalization/) (translation of https://broadrouter.com/glossary/unicode-normalization/): Chuẩn hóa Unicode chuyển văn bản về một dạng chuẩn để các chuỗi trông giống nhau dùng cùng code point. Có bốn dạng: NFD, NFC, NFKD và NFKC; các dạng K còn gộp ký tự tương thích, như dạng trình bày (presentation form) và chữ ghép tiếng Ả Rập, thành chữ cái thông thường. - [VAD (phát hiện giọng nói)](https://broadrouter.com/vi/glossary/vad/) (translation of https://broadrouter.com/glossary/vad/): Voice activity detection (VAD) là bước xử lý tín hiệu phân loại từng khung audio ngắn là có tiếng nói hay không. Hệ thống giọng nói dùng nó để bỏ qua khoảng lặng, tách bản ghi dài tại chỗ ngắt, nhận biết khi người nói đã dừng và tránh chép lời tiếng ồn nền. - [Vector database (cơ sở dữ liệu vector)](https://broadrouter.com/vi/glossary/vector-database/) (translation of https://broadrouter.com/glossary/vector-database/): Vector database (cơ sở dữ liệu vector) là kho dữ liệu lưu các vector embedding, thường kèm văn bản gốc và metadata, và trả lời truy vấn láng giềng gần nhất: với một vector truy vấn, nó trả về các vector đã lưu giống nhất, thường đo bằng khoảng cách cosine, khoảng cách Euclid hoặc tích vô hướng. - [Vision-language model (VLM, mô hình thị giác-ngôn ngữ)](https://broadrouter.com/vi/glossary/vision-language-model/) (translation of https://broadrouter.com/glossary/vision-language-model/): Vision-language model (VLM) là model nhận đầu vào gồm cả ảnh và văn bản rồi tạo ra văn bản, nhờ đó có thể mô tả hình ảnh, trả lời câu hỏi về ảnh và đọc tài liệu, biểu đồ hay ảnh chụp màn hình. VLM còn được gọi là model image-text-to-text hoặc model đa phương thức. - [vLLM](https://broadrouter.com/vi/glossary/vllm/) (translation of https://broadrouter.com/glossary/vllm/): vLLM là thư viện mã nguồn mở để chạy suy luận và phục vụ (serving) LLM, ban đầu được phát triển tại Sky Computing Lab của UC Berkeley. Nó quản lý bộ nhớ KV cache bằng PagedAttention, gom request liên tục theo continuous batching, hỗ trợ nhiều định dạng lượng tử hóa và cung cấp HTTP server tương thích OpenAI. - [Vocabulary size (kích thước từ vựng)](https://broadrouter.com/vi/glossary/vocabulary-size/) (translation of https://broadrouter.com/glossary/vocabulary-size/): Vocabulary size là số token khác nhau trong tokenizer của model ngôn ngữ, mỗi token có ID và embedding riêng. Từ vựng lớn hơn có thể mã hóa văn bản thông dụng bằng ít token hơn, nhưng làm lớp embedding, lớp đầu ra to hơn, và chỉ giúp với loại văn bản đã dùng huấn luyện từ vựng. - [WAV](https://broadrouter.com/vi/glossary/wav/) (translation of https://broadrouter.com/glossary/wav/): WAV (Waveform Audio File Format, hay WAVE) là một container audio xây trên định dạng RIFF của Microsoft và IBM. Nó thường chứa các mẫu PCM tuyến tính không nén, nên dung lượng dễ dự đoán: mỗi giây bằng tần số lấy mẫu nhân số byte mỗi mẫu nhân số kênh. - [Word error rate (WER, tỷ lệ lỗi từ)](https://broadrouter.com/vi/glossary/word-error-rate/) (translation of https://broadrouter.com/glossary/word-error-rate/): Word error rate (WER, tỷ lệ lỗi từ) là thước đo độ chính xác tiêu chuẩn cho nhận dạng giọng nói: số lần thay thế, xóa và chèn từ cần để biến bản chép lời thành bản tham chiếu đúng, chia cho số từ trong bản tham chiếu. Càng thấp càng tốt. - [Không lưu giữ dữ liệu (zero data retention, ZDR)](https://broadrouter.com/vi/glossary/zero-data-retention/) (translation of https://broadrouter.com/glossary/zero-data-retention/): Không lưu giữ dữ liệu (ZDR) là chính sách mà nhà cung cấp AI chỉ xử lý prompt và đầu ra để phục vụ request, sau đó không lưu nội dung đó vào log, cache hay dữ liệu huấn luyện. Các bản ghi không chứa nội dung như số token và thời gian xử lý thường vẫn được giữ. ## Optional - [Sign up and console](https://inference.broadnet.ai/) - [Official API reference](https://inference.broadnet.ai/docs) - [Official machine-readable prices](https://inference.broadnet.ai/pricing.json) - [About BroadRouter](https://broadrouter.com/about/) - [Editorial policy](https://broadrouter.com/editorial-policy/) - [Fact-check log: every correction with old claim, new claim and source](https://broadrouter.com/fact-check-log/) - [Terms of Use](https://broadrouter.com/terms/) - [Privacy policy](https://broadrouter.com/privacy/) - [Refund policy](https://broadrouter.com/refund/)