[AliZakaee]

Summary

• validate each stored knowledge-base embedding before adding it to the NumPy matrix
• skip non-numeric, nested, non-finite, and float32-overflowing vectors
• strengthen regression coverage for mixed and entirely unusable stored vectors

Why this is meaningful

The loader already promises to ignore malformed embedding rows, but a correctly sized JSON list containing a non-numeric value raised during the final NumPy conversion. That exception escaped the retrieval layer and could abort both the incoming customer-response pipeline and the owner's knowledge-base search. Validating per row preserves healthy vectors and lets an all-invalid set use the existing keyword fallback.

Validation

• python -m pytest tests/unit/services/test_embedding_service.py -q — 13 passed
• python -m pytest — 283 passed
• python -m pytest --cov=. --cov-report=term-missing --cov-report=xml --cov-report=html --junitxml=junit.xml — 283 passed, 61% total coverage
• python -m compileall -q config.py database.py handlers services security.py telegram_api.py utils.py — passed
• git diff --check — passed

Risk and rollback

Risk is low: valid finite embeddings retain the same float32 representation, while only unusable rows are filtered earlier. Revert commit 5ac50d2 to roll back.

Prepared with automated assistance by EQ for AliZakaee.