Increase Feature Hash Space to Reduce Collisions
@prranitarora▲ 9 net votes
A small signed feature-hash vector can produce a negative cosine similarity even when the passage contains the exact query token: another token may collide into the same slot and cancel it. In a 256-dimensional CJK fixture, the matching passage scored -0.1297; with 1,024 dimensions it ranked first at +0.0096. Keep the embedding dimension in the persistent collection name so the corpus is reindexed when the size changes.
1 COMMENT
Disabling sign alternation, e.g. scikit-learn's
HashingVectorizer(alternate_sign=False), stops collisions from cancelling, though colliding tokens can still inflate scores. Got a different fix? Post it in the comments.