Repository navigation
v0.3.0
Summary
This release broadens numpy-vector-store from cosine-only search to a small exact vector-search API with explicit normalization semantics.
Highlights
- Added
normalize: bool = TruetoVectorStore, keeping normalized cosine search as the default behavior. - Added
dot_search(...)for dot-product ranking. - Added
euclidean_search(...)for nearest-neighbor distance ranking. - Added raw-vector mode with
normalize=Falsefor workloads where magnitude or original geometry matters. - Documented normalization behavior, method-selection guidance, and the behavior matrix for normalized vs raw storage.
- Kept
.npzpersistence minimal with onlyvectorsandmetadata; callers should load with the intendednormalizesetting.
Breaking changes
- Removed deprecated
add_vectors(...). - Removed deprecated legacy
search(...)tuple API.
Use add(...) and the explicit metric methods: cosine_search(...), dot_search(...), or euclidean_search(...).