Anatomic location search returns irrelevant results for multi-word queries

[talkasab]

Summary

The anatomic location hybrid search returns badly wrong results for straightforward multi-word queries:

• "cardiac chamber" → returns eye chambers instead of heart chambers
• "cardiac muscle" → returns shoulder muscles instead of myocardium

The semantic (vector) component appears to be matching on generic terms ("chamber", "muscle") while ignoring the qualifying anatomic context ("cardiac").

Investigation areas

1. Embedding quality — are the stored embeddings capturing anatomic context, or just individual word similarity? Check what vectors are stored for cardiac structures vs. eye/shoulder structures.
2. FTS vs. semantic weighting — the hybrid search blends FTS (BM25) and vector scores. If FTS correctly ranks cardiac results higher but the semantic score overwhelms it, the weights need adjustment.
3. Query preprocessing — is the full query string being embedded, or is it being tokenized/split in a way that loses the multi-word context?
4. Index coverage — verify the anatomic location database actually contains the expected cardiac entries (heart chambers, myocardium, etc.) with appropriate synonyms.
5. Eval suite — evals/anatomic_search.py has 10 test cases. Check whether these cases cover multi-word anatomic queries and whether scores are acceptable. Add failing cases for the examples above.

Reproduction

from anatomic_locations import AnatomicLocationIndex

async with AnatomicLocationIndex() as idx:
    results = await idx.search("cardiac chamber")
    # Expected: heart chambers (left ventricle, right atrium, etc.)
    # Actual: eye chambers (anterior chamber, posterior chamber, etc.)

[talkasab]

Resolved in anatomic-locations by making embedding config respect standard OpenAI env vars and .env:

• Added AliasChoices so ANATOMIC_OPENAI_* falls back to OPENAI_* (API key/model/dimensions).
• Enabled .env loading in AnatomicLocationSettings so OPENAI_API_KEY works without shell export.
• Added eval cases for multiword qualifiers to guard regression.
  This restores semantic search for multiword queries like “cardiac chamber/muscle”.