Trait Quantizer

Source

pub trait Quantizer: Send + Sync {
    type Code: Clone + Send + Sync;

    // Required methods
    fn encode(&self, vector: &[f32]) -> Self::Code;
    fn distance_approx(&self, a: &Self::Code, b: &Self::Code) -> f32;
    fn distance_exact(&self, a: &[f32], b: &[f32]) -> f32;
    fn is_accelerated(&self) -> bool;
    fn name(&self) -> &str;

    // Provided methods
    fn needs_training(&self) -> bool { ... }
    fn train(&mut self, _sample: &[&[f32]]) { ... }
}

Expand description

Acceleration strategy for distance computations.

Implementations encode vectors into compact codes and provide fast approximate distance between codes.

The graph uses distance_approx for candidate exploration (hot path) and distance_exact for final neighbor selection (quality-critical).

Required Associated Types§

Source

type Code: Clone + Send + Sync

Compact representation of a vector.

NoQuantizer: () (zero overhead)
Scalar: Vec<u8> (D bytes)
Product: Vec<u8> (M bytes, where M = D/subvector_dim)
Binary: Vec<u64> (D/64 words)

Required Methods§

Source

fn encode(&self, vector: &[f32]) -> Self::Code

Encode a full-precision vector into a compact code.

Called once per insert. The code is stored alongside the vector.

Source

fn distance_approx(&self, a: &Self::Code, b: &Self::Code) -> f32

Fast approximate distance between two codes.

This is the hot path: called O(ef_construction × log N) times per insert. Must be significantly faster than distance_exact for the acceleration to be worthwhile.

Source