Struct QuantizedVectorStore 

pub struct QuantizedVectorStore { /* private fields */ }

Stores f32 vectors quantized to INT8, reducing memory ~4× vs raw f32.

Each vector is independently scaled with per-vector scale and zero_point so that dequantization error is bounded to roughly scale / 2 per element.

Memory layout

Internally uses a single flat Vec<i8> of length count * dim.

Implementations

impl QuantizedVectorStore

pub fn new(dim: usize) -> Self

Create an empty store for vectors of dim dimensions.

pub fn push(&mut self, vector: &[f32]) -> usize

Quantize vector to INT8 and append it to the store.

Returns the assigned integer ID (0-based).

pub fn get(&self, id: usize) -> Option<Vec<f32>>

Dequantize the vector stored at id back to f32.

Returns None if id >= self.len().

pub fn cosine_similarity_q(&self, a_id: usize, b_id: usize) -> f32

Compute approximate cosine similarity between two stored vectors.

Dequantizes both vectors then computes exact cosine similarity in f32. Returns 0.0 if either id is out of range or a vector has zero norm.

pub fn bytes_per_vector(&self) -> f64

Approximate memory usage per stored vector (bytes).

Counts the i8 storage only (not scale/zero_point metadata): dim * 1 byte per vector.

pub fn len(&self) -> usize

Number of stored vectors.

pub fn is_empty(&self) -> bool

Returns true if no vectors have been stored.

pub fn dim(&self) -> usize

Vector dimension.