Summary

gbrain query (hybrid) returns the same top documents with identical scores for completely unrelated queries — including nonsense input. gbrain search (BM25) is query-responsive and correct. The hybrid/vector path appears to ignore the query.

Environment

• gbrain 0.41.11.1 (also reproduced on 0.35.6.0 before upgrading)
• Engine: postgres
• Embedding: ollama:nomic-embed-text, 768d (gbrain doctor: embedding_provider ✓, embedding_width_consistency OK)
• macOS / bun

Repro

gbrain query "wyckoff accumulation"     # top-3 docs A,B,C @ ~0.99/0.98/0.97
gbrain query "banana recipe quantum"    # IDENTICAL top-3 A,B,C @ IDENTICAL scores
gbrain query "..." --no-expand          # still identical
gbrain search "wyckoff accumulation"    # correct, query-responsive
gbrain search "banana recipe quantum"   # different/correct results

Different queries → identical query results and identical cosine scores ⇒ the query vector used in the hybrid search appears fixed/independent of the input.

Ruled out (investigated)

• Embedding model: ollama nomic-embed-text returns distinct vectors for distinct inputs (cosine ≈0.47 between the two example queries, via both /api/embeddings and /v1/embeddings, with and without input_type).
• Semantic query cache: gbrain cache clear --yes, then fresh (cache-miss) queries are still identical.
• Query expansion: --no-expand still identical.
• BM25 (search): correct and query-responsive — data + BM25 path are fine.

Hypothesis

The query-side embedding (embedQuery → gateway embed([text], {inputType:'query'})) or its use in core/search/hybrid.ts may produce/use a fixed query vector. On every query this warning prints: [ai.gateway] recipe "google" declares an embedding touchpoint without max_batch_tokens — possibly related to query-embedding routing, even though the configured embedding model is ollama:nomic-embed-text.

Happy to provide more detail or test a patch.