Skip to content
Citadel docs

Citadel docs

Overview

  • Introduction — Base URL, products, statuses and request conventions.
  • Authentication — API keys, the Authorization header and session routes.
  • Errors — Error formats, status codes and retry rules.
  • Rate limits — Request, token and concurrency limits.
  • Contributing — Page format, front matter and coverage tests.

LLM

  • LLM API — OpenAI-compatible chat completions on open-weight Qwen and DeepSeek models.
  • Chat completions — Request, response and errors for POST /v1/chat/completions and GET /v1/models.
  • Streaming — Server-sent events for chat completions.
  • Tools and structured output — Function calling and JSON output with response_format.
  • Thinking — Qwen reasoning on and off with chat_template_kwargs.enable_thinking.
  • Model switching — Model switches, cold starts, X-LLM-Wait and 503 model_loading.
  • Queueing — GPU queue, 503 queued, 503 gpu-busy and throughput under concurrency.
  • Quantization and abliteration — Weight precision and refusal removal per model, with the accuracy cost.
  • Prefix caching — Prompt prefix reuse across requests, cache size and measured speedup.
  • Known issues — Open issues with dates and workarounds.
  • Examples — Runnable curl, Python and JavaScript code, parallel requests, 503 retries.