Struct HnswCodecIndex 

pub struct HnswCodecIndex<C: VectorCodec> {
    pub dim: usize,
    pub m: usize,
    pub m0: usize,
    pub ef_construction: usize,
    pub max_layer: usize,
    pub entry_point: Option<u32>,
    pub level_mult: f32,
    pub codec: C,
    /* private fields */
}

Hierarchical Navigable Small World graph index parameterised on a VectorCodec.

• Upper-layer routing: codec.fast_symmetric_distance
• Base-layer rerank: codec.exact_asymmetric_distance

Fields

dim: usize

m: usize

Max neighbors per node in upper layers.

m0: usize

Max neighbors per node at layer 0 (= m * 2).

ef_construction: usize

max_layer: usize

Current highest layer in the index.

entry_point: Option<u32>

Dense index (insertion order) of the current entry-point node.

level_mult: f32

1.0 / ln(m) — used by random_layer.

codec: C

Implementations

impl<C: VectorCodec> HnswCodecIndex<C>

pub fn insert(&mut self, id: u32, v: &[f32])

Insert a vector with the given caller-supplied id.

Encodes v via codec.encode, assigns a random layer, and runs the standard HNSW neighbour-selection algorithm.

impl<C: VectorCodec> HnswCodecIndex<C>

pub fn new( dim: usize, m: usize, ef_construction: usize, codec: C, seed: u64, ) -> Self

Create a new empty codec index.

pub fn len(&self) -> usize

Number of nodes (including deleted).

pub fn is_empty(&self) -> bool

pub fn dim(&self) -> usize

pub fn random_layer(&mut self) -> usize

Assign a random layer per the HNSW exponential distribution.

Capped at MAX_LAYER_CAP to prevent pathological draws from inflating max_layer.

pub fn quantized_at(&self, idx: u32) -> Option<&C::Quantized>

Return a reference to the quantized payload at idx, if present and not deleted.

pub fn neighbors_at(&self, idx: u32, layer: usize) -> &[u32]

Return the neighbor slice for idx at layer.

impl<C: VectorCodec> HnswCodecIndex<C>

pub fn search( &self, query: &[f32], k: usize, ef_search: usize, ) -> Vec<CodecSearchResult>

K-NN search returning up to k results.

ef_search controls the beam width at layer 0 (must be >= k).

The returned results are sorted ascending by exact_asymmetric_distance.