Documentation
¶
Overview ¶
Package hnsw implements an in-memory Hierarchical Navigable Small World graph (Malkov & Yashunin) for approximate nearest neighbor search over float32 vectors. Scores use the same arithmetic as vec.Score, so a document found approximately carries exactly the score exact search would give it.
Index ¶
- Constants
- type Graph
- func (g *Graph) Add(id uint64, vector []float32) error
- func (g *Graph) Len() int
- func (g *Graph) Search(query []float32, k, ef int, accept func(uint64) bool) ([]vec.Match, error)
- func (g *Graph) SearchContext(ctx context.Context, query []float32, k, ef int, accept func(uint64) bool) ([]vec.Match, error)
- type Params
Constants ¶
const ( DefaultM = 16 DefaultEfConstruction = 200 DefaultEfSearch = 100 )
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type Graph ¶
type Graph struct {
// contains filtered or unexported fields
}
Graph holds vectors and their navigable layers. Add is not safe for concurrent use; once building is done, Search is safe for concurrent use.
func (*Graph) Add ¶
Add inserts vector for document id. Every vector must share one dimension; the same id may be added more than once.
func (*Graph) Search ¶
Search returns up to k documents, best first with ties by ascending id; repeated ids keep their best score. ef bounds the candidate list (0 uses the graph's EfSearch; k raises it). accept optionally filters results with the traversal semantics documented by SearchContext.
func (*Graph) SearchContext ¶
func (g *Graph) SearchContext(ctx context.Context, query []float32, k, ef int, accept func(uint64) bool) ([]vec.Match, error)
SearchContext is Search with cancellation during descent and layer expansion. Discovered rejected nodes are always expanded as bridges, even with a full accepted beam. Accepted nodes remain score-pruned, so search is approximate; selective filters may traverse an entire connected rejected region.
type Params ¶
type Params struct {
// M is the maximum number of neighbors per node above level 0; level 0
// keeps 2*M. Larger values raise recall and memory. Minimum 2.
M int
// EfConstruction is the candidate list size while inserting.
EfConstruction int
// EfSearch is the candidate list size while searching; k raises it when larger.
EfSearch int
}
Params tunes graph construction and search. Zero fields take the defaults.
Source Files