Module quantization 

Product Quantization for vector compression.

This module provides Product Quantization (PQ) for compressing high-dimensional vectors into compact codes while preserving approximate distance computation.

Overview

Product Quantization works by:

1. Splitting vectors into M subspaces (segments)
2. Training K centroids (codebook entries) per subspace using k-means
3. Encoding each vector as M indices into the codebooks

For example, a 128-dimensional vector with M=8 segments and K=256 centroids can be compressed from 512 bytes (128 × 4 bytes) to just 8 bytes (8 × 1 byte).

Distance Computation

Two modes are supported:

• Symmetric Distance Computation (SDC): Approximate distances between two compressed vectors. Fast but less accurate.
• Asymmetric Distance Computation (ADC): Exact query vector, compressed database vectors. Better accuracy for search.

Example

use manifoldb_vector::quantization::{ProductQuantizer, PQConfig};

// Create a PQ with 8 segments and 256 centroids per segment
let config = PQConfig::new(128, 8).with_num_centroids(256);
let training_data: Vec<Vec<f32>> = /* your training vectors */;
let pq = ProductQuantizer::train(&config, &training_data)?;

// Encode vectors
let vector = vec![0.1f32; 128];
let code = pq.encode(&vector);

// Compute distances with ADC
let query = vec![0.2f32; 128];
let distance_table = pq.compute_distance_table(&query);
let distance = pq.asymmetric_distance(&distance_table, &code);

Structs

KMeans

K-means clustering result.

KMeansConfig

Configuration for k-means clustering.

PQCode

A compressed vector code from Product Quantization.

PQConfig

Configuration for Product Quantization.

ProductQuantizer

Product Quantizer for vector compression.