Struct Hnsw 

pub struct Hnsw { /* private fields */ }

One shard’s ANN graph for one index.

Implementations

impl Hnsw

pub fn vectors(&self) -> u64

Live (non-tombstoned) vectors — already tracked, so O(1). Self::stats walks every node and every link to estimate bytes; a caller that only wants the count should not trigger that walk.

pub fn stats(&self) -> VectorStats

Counters — O(1): links_total and tombstones are maintained at the three mutation sites (link push, shrink, tombstoning) instead of walking every node per call (this ran on every tiering tick). [Self::recompute_stats] is the walking reference the tests hold them to.

impl Hnsw

pub fn new(dim: usize, params: HnswParams) -> Self

Empty graph for dim-dimensional vectors.

pub fn dim(&self) -> usize

Declared dimensionality.

pub fn apply(&mut self, key: &[u8], vector: Option<Vec<f32>>)

Insert or replace key’s vector (None = remove). Replace = detach old key (tombstone the node once keyless) + insert new. Keys sharing one exact vector share one graph node.

pub fn knn(&self, query: &[f32], k: usize, ef: usize) -> Vec<(Vec<u8>, f32)>

k nearest LIVING vectors to query (raw form; prepared here). ef = query beam width (0 → the max(4k, 100) default); larger beams trade latency for recall — the canonical HNSW knob.

pub fn contains(&self, key: &[u8]) -> bool

Membership (living only).

pub fn rebuild(&mut self)

Bounded rebuild: re-insert every living (key, vector) pair into a fresh graph (drops tombstones and their edges). Vectors are already prepared; add_key re-collapses duplicates.