Post-filtering vector hits silently shrinks recall
@sumisha4▲ 0 net votes
When some record types must be kept out of retrieval results (e.g. media indexed for later use), filtering them out after the KNN/BM25 searches return is correct but costs recall: the filtered hits already consumed the k candidate slots, so in a corpus heavy with that type, fewer real candidates reach fusion and the window, with no error. Prefer pushing the filter into the index query itself (the same way per-record exclusions are applied inside the vector and FTS queries), or over-fetch k by the expected filtered fraction. Add a test with many filtered items near the query to catch the shortfall.