Struct QuantizedVectors 

pub struct QuantizedVectors {
    pub dims: usize,
    pub num_vectors: usize,
    pub data: Vec<u8>,
    pub mins: Vec<f32>,
    pub scales: Vec<f32>,
    pub norms: Vec<f32>,
    pub metric: DistanceMetric,
}

Int8 quantized vectors with per-dimension calibration data.

Fields

dims: usize

num_vectors: usize

data: Vec<u8>

Flat quantized data: data[i * dims .. (i+1) * dims] = vector i.

mins: Vec<f32>

Per-dimension minimum values.

scales: Vec<f32>

Per-dimension scale: (max - min) / 255.0.

norms: Vec<f32>

Per-vector precomputed norms (approximate, for cosine distance).

metric: DistanceMetric

Distance metric.

Implementations

impl QuantizedVectors

pub fn quantize(vectors: &[Vec<f32>], metric: DistanceMetric) -> Self

Quantize a set of float32 vectors to int8.

pub fn get(&self, idx: usize) -> &[u8]

Get the quantized vector for index idx.

pub fn asymmetric_distance(&self, idx: usize, query: &[f32]) -> f32

Compute asymmetric distance between quantized stored vector and float32 query vector.

For cosine, the query is expected to be unit-length (caller normalizes at entry per [[optimize-cosine-norm-precompute]]). stored_norm is the norm of the dequantized stored vector, which approximates 1.0 but drifts due to int8 quantization rounding — keeping it cancels that drift in the score.

pub fn to_bytes(&self) -> Vec<u8>

Serialize to bytes.

pub fn from_bytes(data: &[u8]) -> Self

Deserialize from bytes.