Chat Completions Plugins

Use plugins in /v1/chat/completions when one request needs an explicit PDF engine override, server-side response healing for structured output, or secret sanitization before model execution. The web plugin is also listed here as the planned built-in web search path, but it is not production-ready yet.

Add a plugin to the request

Add one entry to plugins. Keep the rest of the request as a normal chat-completions call.

{
  "model": "qwen/qwen3-235b-a22b-instruct-2507-fp8",
  "messages": [
    {
      "role": "user",
      "content": "Summarize this incident note, but hide apiKey=sk-live-demo-123456."
    }
  ],
  "plugins": [
    {
      "id": "privacy-sanitization",
      "mode": "redact"
    }
  ]
}

• web: add { "id": "web" }. This plugin is still coming soon and is not production-ready yet.
• file-parser: add pdf.engine only when the request must force pdf-text or native.
• response-healing: also send stream: false and a structured response_format.
• privacy-sanitization: also send mode as redact, tokenize, or block.

Available plugins

How activation works

• plugins in the request is the normal activation path for /v1/chat/completions.
• Authenticated API requests may also inherit saved plugin defaults when plugin settings policy is enabled.
• web still requires explicit opt-in per request.
• file-parser can still be applied automatically for PDF inputs when server-managed engine selection is enough.

Limits and compatibility

• web and privacy-sanitization cannot run in the same request.
• response-healing only works with stream: false and response_format.type set to json_object or json_schema.
• web supports streaming and non-streaming requests.
• privacy-sanitization supports streaming and non-streaming requests.

See also

• Web Search for the planned built-in web grounding path. It is still coming soon.
• PDF Inputs for file-parser, pdf-text vs native, and PDF-specific limits.
• Response Healing for non-stream structured-output repair.
• Privacy Sanitization for request-level secret sanitization.
• Structured Outputs if the real job is the JSON contract rather than plugin selection.
• Tool Calling if your application should own retrieval or tool execution.
• Chat Completions API reference for the exact plugins field schema.