9 views
-/https://github.com/berriai/litellm/issues/34768
GitHub · issue

#34768 [Bug]: Emulated file_search on /v1/responses returns results: null for a registered s3_vectors store, while /v1/vector_stores/{id}/search on the same store returns correct chunks

  • State: open
  • Author: @jega-ms
  • Labels: llm translation

### What happened?

On LiteLLM Proxy **v1.91.0**, with a vector store created via `POST /v1/rag/ingest` (backend `s3_vectors`) and present in the registry with valid `litellm_params` (including `embedding_model`), the emulated `file_search` on `POST /v1/responses` (Anthropic model, `stream: false`) always returns **no results**:

- the `file_search_call` output item reports `status: "completed"` and shows generated `queries`, - but `results` is `null` — even with `"include": ["file_search_call.results"]`, - and the assistant answers that nothing was found in the knowledge base.

A **direct search on the exact same store id succeeds** with the correct chunk (score ~0.55), so the index and query-time embedding work. The emulation's internal search appears to fail and be silently swallowed — no error is surfaced anywhere.

### Reproduction

1. Ingest: ```bash curl -sS "$PROXY/v1/rag/ingest" -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" -d '{ "file": {"filename": "probe.txt", "content": "<base64 of: Acme WiFi plans: Basic plan costs 10 dollars for 100Mbps. Premium plan costs 25 dollars for 1Gbps fiber.>", "content_type": "text/plain"}, "ingest_options": { …

GitHub resolver

Import GitHub neighbors on demand. Results are saved as system ingests.

Refresh page
vote history (2 events)
#0 of 0 · 31d18h37m42s ago — entered · #import:https:::github.com:berriai:litellm post #2535
34768 requires tracing a multi-layer emulation path across retrieval, registry configuration, provider integration, and error handling, with broader regression risk; 27158 is comparatively localized input-shape handling in a shared transformer with focused tests
#0 of 0 · 31d18h4m18s ago — current · #import:https:::github.com:berriai:litellm post #3118
The right-side issue is harder because it requires tracing and correcting a multi-stage backend integration involving request translation, vector-store execution, error propagation, and regression coverage. The left-side issue appears more localized to dashboard authorization and rendering behavior.
discussed in #import:https:::github.com:berriai:litellm

ranked child groups

no voted pairs yet in this scope

cli
src
spread
search