crates/hnsw

This companion crate gives you approximate-nearest-neighbour search for VectorMemoryStore. It is an HNSW, a Hierarchical Navigable Small World, implementation of VectorIndex.

It is the scale tier sitting between the linear reference index and the external stores, and it publishes after loopctl like the other companions.

Key items

The index gives you add, search, remove, and len over the same object-safe trait the linear index implements.

Graph construction and search live in graph.rs, and the index mapping lives in map.rs.

Behavior notes

The correctness contracts are the linear index’s: cosine-scored top-k sorted descending with the id tiebreak, idempotent remove, and a dimension mismatch surfacing as LoopError::Memory.

The linear index stays the oracle that your faster indexes must match.

This crate is new in 0.4.0 and did not exist at v0.3.2. Reaching release resolved four contract edges: the removal model, ordering, metrics, and upsert failure handling, so search now upholds the crate’s own documented behavior.

For the deep dive, see the semantic tiers section of Memory.