Module quantize 

Vector quantization for memory-efficient storage

This module provides quantization methods to reduce memory footprint while maintaining acceptable search quality:

• Scalar Quantization (SQ8): f32 → i8 (4x compression, ~95% recall)
• Binary Quantization (BQ): f32 → bit (32x compression, ~85% recall)

Memory Comparison (1M vectors × 384 dims)

Format	Size	Compression
f32	1.5 GB	1x
int8	384 MB	4x
binary	48 MB	32x

Usage

use foxstash_core::vector::quantize::{ScalarQuantizer, BinaryQuantizer, Quantizer};

let vector = vec![0.5, -0.3, 0.8, -0.1];

// Scalar quantization (4x compression)
let sq = ScalarQuantizer::fit(&[vector.clone()]);
let quantized = sq.quantize(&vector);
let reconstructed = sq.dequantize(&quantized);

// Binary quantization (32x compression)
let bq = BinaryQuantizer::new(4);
let binary = bq.quantize(&vector);

Structs

BinaryQuantizedVector

Quantized vector representation for binary quantization

BinaryQuantizer

Binary quantization: f32 → bit

ProductQuantizerConfig

Product Quantization configuration

ScalarQuantizationParams

Scalar quantization parameters for a single dimension

ScalarQuantizedVector

Quantized vector representation for SQ8

ScalarQuantizer

Scalar quantization (SQ8): f32 → u8

Traits

Quantizer

Trait for vector quantization

Functions

binary_dot_product

Compute dot product between f32 query and binary quantized vector

hamming_distance_simd

SIMD-accelerated Hamming distance for binary vectors

sq8_asymmetric_l2_distance_simd

SIMD-accelerated asymmetric L2 distance (f32 query vs SQ8 database)

sq8_l2_distance_simd

SIMD-accelerated L2 distance for SQ8 vectors