Expand description
KNN query: approximate nearest neighbor search via the index-wide global HNSW per [[global-vector-indices]] Alternative B.
The global graph is consulted once at bind time, producing a single
Vec<(SegmentId, DocId, distance)>. Hits are bucketed by segment
and exposed to the per-segment scorer dispatch the rest of the
engine uses. Per-segment kNN fan-out and proportional-k oversampling
are gone — there is one graph, one search, one ranking.
Structs§
- KnnQuery
- Approximate k-nearest-neighbor query on a dense_vector field.