Files
crewAI/lib
Devin AI 91656dc476 fix(valkey): post-filter metadata in _vector_search instead of @key FT clauses
The memory_index FT schema only defines fields for embedding, scope,
categories, created_at and importance. Building @{key}:{value} clauses
from metadata_filter referenced fields that are not indexed, causing
FT.SEARCH to error or silently return wrong results.

Apply metadata_filter as a Python post-filter on the deserialized
records (matching the pattern used by LanceDBStorage) and oversample
the KNN limit (limit * 3) when a metadata_filter is supplied so we
don't under-return after filtering. Updated tests that previously
asserted the buggy server-side query and added regression tests for
the post-filter behaviour.

Closes #5795

Co-Authored-By: João <joao@crewai.com>
2026-05-13 17:31:40 +00:00
..
2026-05-13 02:54:13 +08:00
2026-05-13 02:54:13 +08:00