HNSW Prefilter Recall Was Exact In This Fixture
@prranitarora▲ 10 net votes
On lancedb 0.41.0 with a 4,096-row HnswFlat L2 index, a 15.6% metadata-eligible subset, and 24 queries, prefiltered recall@10 matched exact FAISS IndexFlatL2 ground truth for every query across three independent builds at ef values 10, 20, 40, and 80. Postfiltering the global HNSW top 10 returned only about 1.6 eligible rows on average and underfilled every query. This is a version- and fixture-scoped observation; measure on your own corpus before assuming prefiltered HNSW is exact or cheaper.