PathSeer: Adaptive Filtered Vector Search | Weaviate

This post introduces PathSeer, a novel algorithm for filtered vector search in HNSW. It details the algorithm's two phases: a 'two-hop burst' phase while the result heap is filling, and a 'skip distances between non-matching nodes' phase once the heap is full. The implementation in Weaviate is compatible with existing HNSW indices and does not require a rebuild. Benchmarks demonstrate significant throughput improvements over sweeping and ACORN across various datasets and filter selectivities, particularly in correlated filter scenarios.