Struct ProductQuantizer 

pub struct ProductQuantizer {
    pub subspace_count: usize,
    pub codebook_size: usize,
    pub codebooks: Vec<Vec<Vec<f32>>>,
    pub subvec_dim: usize,
}

Product quantizer: divides the embedding into subspaces and learns a codebook per subspace for efficient approximate nearest-neighbor search.

Fields

subspace_count: usize

Number of subspaces.

codebook_size: usize

Number of codes per subspace (e.g. 256 for 8-bit codes).

codebooks: Vec<Vec<Vec<f32>>>

Codebooks: [subspace][code][subvec_dim].

subvec_dim: usize

Dimension of each sub-vector (embedding_dim / subspace_count).

Implementations

impl ProductQuantizer

pub fn new(subspace_count: usize, codebook_size: usize) -> Self

Create an untrained product quantizer.

pub fn train(&mut self, embeddings: &[Vec<f32>])

Train the product quantizer on a set of embeddings.

Uses a simplified k-means: randomly selects codebook_size distinct embeddings as initial centroids, then runs a few iterations.

pub fn encode(&self, embedding: &[f32]) -> Vec<u8>

Encode an embedding as a vector of codebook indices (one per subspace).

pub fn decode(&self, codes: &[u8]) -> Vec<f32>

Decode a code vector back to an approximate embedding.

pub fn approx_distance(&self, codes1: &[u8], codes2: &[u8]) -> f32

Compute approximate L2 distance between two encoded vectors using codebook lookups.

pub fn is_trained(&self) -> bool

Check whether the quantizer has been trained.

Trait Implementations

impl Clone for ProductQuantizer

fn clone(&self) -> ProductQuantizer

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ProductQuantizer

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.