Struct ResidualCodec 

pub struct ResidualCodec { /* private fields */ }

Residual quantization codec for compressing token embeddings.

The codec learns centroids via k-means clustering, then quantizes the residuals (v - centroid) to a small number of bits per dimension.

Compression Process

1. Find nearest centroid for input vector
2. Compute residual = vector - centroid
3. Quantize each dimension to nbits using learned bucket boundaries
4. Pack quantized values into bytes

Scoring

Score computation avoids full decompression:

q · v ≈ q · c + Σ_d q[d] × bucket_weight[d, code[d]]

Implementations

impl ResidualCodec

pub fn train( embeddings: &[f32], dim: usize, num_centroids: usize, nbits: u8, iterations: usize, ) -> Result<Self>

Train a codec from sample embeddings.

Arguments

• embeddings - Flattened sample embeddings [n × dim]
• dim - Embedding dimension
• num_centroids - Number of k-means centroids
• nbits - Bits per dimension (2 or 4)
• iterations - K-means iterations

Errors

Returns an error if training data is insufficient or parameters invalid.

pub fn with_params( centroids: Vec<f32>, num_centroids: usize, dim: usize, bucket_cutoffs: Vec<f32>, bucket_weights: Vec<f32>, nbits: u8, ) -> Self

Create a codec with pre-trained parameters.

Panics

Panics if dim == 0 (poka-yoke: division-by-zero guard).

pub fn num_centroids(&self) -> usize

Get the number of centroids.

pub fn dim(&self) -> usize

Get the embedding dimension.

pub fn nbits(&self) -> u8

Get bits per dimension.

pub fn packed_size(&self) -> usize

Get the packed residual size in bytes.

pub fn centroid(&self, id: usize) -> &[f32]

Get centroid slice by ID.

pub fn centroids(&self) -> &[f32]

Get all centroids as a flat slice.

pub fn find_nearest_centroid(&self, embedding: &[f32]) -> usize

Find the nearest centroid for a vector.

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

Compress an embedding to (centroid_id, packed_residual).

pub fn decompress_score( &self, query_token: &[f32], centroid_id: usize, centroid_score: f32, packed_residual: &[u8], ) -> f32

Compute score between query token and compressed document token.

score ≈ q · d = q · c + q · r

Arguments

• query_token - Query embedding
• centroid_id - Assigned centroid
• centroid_score - Precomputed q · c
• packed_residual - Packed quantized residual

pub fn centroid_score(&self, query_token: &[f32], centroid_id: usize) -> f32

Compute dot product between query and centroid.

Trait Implementations

impl Clone for ResidualCodec

fn clone(&self) -> ResidualCodec

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ResidualCodec

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for ResidualCodec

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for ResidualCodec

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.