LLM

See AI → LLM chat gateway for the provider/config side of this endpoint.

POST /llm/chat — any authenticated record

{
  "messages": [ { "role": "user", "content": "Summarize the onboarding doc." } ],
  "collection": "messages",
  "clientId": "abc123..."
}

The HTTP response is the final result, not the streamed body. If clientId names an already-open GET /realtime connection, incremental tokens are pushed over that connection as llm_chunk ({requestId, delta}) events, then a final llm_done ({requestId, text}) — or llm_error ({requestId, message}) in place of llm_done if the provider fails mid-stream. Omitting clientId still returns the same final JSON, it just never sees incremental deltas.

{ "reply": "...", "promptTokens": 128, "completionTokens": 64, "record": { "id": "...", "prompt": "...", "response": "...", "model": "gpt-4o-mini" } }

record is present only when collection was set and the write succeeded — subject to that collection’s own createRule evaluated against the calling record, since a platform-level feature doesn’t imply superuser-only write access to wherever it persists.

This endpoint spends the operator’s own provider quota on every call — add an explicit rate limit on this exact path in production.