Struct ScalarQuantizer

pub struct ScalarQuantizer<D: DistanceMetric> {
    metric: D,
    min_val: f32,
    scale: f32,
}

Scalar Quantization: compress each float32 dimension to uint8.

For each dimension, maps the value range [min, max] → [0, 255]. Distances are computed on uint8 values using integer arithmetic.

Pros: Simple, no training needed, ~4× distance speedup, <1% recall loss. Cons: Requires knowing the value range (uses [-1, 1] for normalized vectors).

This is what Qdrant uses by default for HNSW construction.

Fields

metric: D

min_val: f32

Min value per dimension (for denormalization). Default: -1.0

scale: f32

Scale factor: 255.0 / (max_val - min_val)

Implementations

impl<D: DistanceMetric> ScalarQuantizer<D>

pub fn new(metric: D) -> Self

Create a scalar quantizer for L2-normalized vectors (range [-1, 1]).

pub fn with_range(metric: D, min_val: f32, max_val: f32) -> Self

Create a scalar quantizer with custom value range.

Trait Implementations

impl<D: Clone + DistanceMetric> Clone for ScalarQuantizer<D>

fn clone(&self) -> ScalarQuantizer<D>

Returns a copy of the value.

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

Performs copy-assignment from source.

impl<D: DistanceMetric> Quantizer for ScalarQuantizer<D>

type Code = Vec<u8>

Compact representation of a vector.

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

Encode a full-precision vector into a compact code.

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

Fast approximate distance between two codes.

fn distance_exact(&self, a: &[f32], b: &[f32]) -> f32

Exact distance between full-precision vectors.

fn is_accelerated(&self) -> bool

Whether this quantizer provides actual acceleration.

fn name(&self) -> &str

Human-readable name of this strategy.

fn needs_training(&self) -> bool

Whether this quantizer needs training on a data sample before use.

fn train(&mut self, _sample: &[&[f32]])

Train the quantizer on a sample of vectors.