Module optimized

Advanced HNSW optimizations.

Heuristic Neighbor Selection (Malkov §4.2)

Instead of keeping the M closest neighbors, selects neighbors that provide good graph connectivity by preferring diverse directions over raw proximity.

Time-Decay Edge Weights

Edge weights decay exponentially with age: w(e, t) = w0 * exp(-λ * age). During search, decayed weights penalize stale connections.

Backup Neighbors

Each node maintains a secondary neighbor list used when primary neighbors are removed (e.g., node expiration in streaming scenarios).

Index Persistence

Serialize/deserialize the HNSW graph for crash recovery and snapshots.

Structs

BackupNeighbors

Backup neighbor storage for handling node expiration.

SerializedGraph

Serialized HNSW graph for persistence.

Traits

NodeVectors

Select neighbors using the heuristic from Malkov & Yashunin (2018) §4.2.

Functions

decay_adjusted_distance

Apply time-decay to a distance score.

select_neighbors_heuristic

Select neighbors using the heuristic from Malkov & Yashunin (2018) §4.2.

serialize_graph

Serialize an HNSW graph to a portable format.

time_decay_weight

Time-decay weight for an edge.