vector_quantization

Vector Quantization Feature

Status: ✅ Implemented
Version: v1.3.0
Feature ID: #7

Overview

Vector Quantization provides memory compression for high-dimensional vectors using Product Quantization (PQ), reducing storage requirements by up to 97% while maintaining acceptable search accuracy.

Key Features

• Product Quantization (PQ): Compress vectors using 8-bit codes
• Memory Compression: Reduce 1536D float32 vectors from 6KB to 192 bytes
• K-means Training: Automatic codebook generation from training data
• Asymmetric Distance: Fast distance computation directly from quantized codes
• Configurable Subquantizers: Adjust compression ratio vs. accuracy trade-off

Quick Start

#include "index/vector_index.h"

VectorIndexManager vim(db);
vim.init("documents", 1536);

// Enable quantization
vim.enableQuantization(true, 8);

// Train quantizer
vim.trainQuantizer();

// Vectors are now automatically quantized
vim.addEntity(entity, "embedding");

// Search works with quantized codes
auto [status, results] = vim.searchKnn(query, 10);

Performance

• Memory Reduction: 32x compression (6KB → 192 bytes for 1536D)
• Speed Improvement: 2-4x faster search
• Accuracy: 95-98% recall@10

Documentation

See full documentation at docs/features/vector_quantization.md

References

• Paper: "Product Quantization for Nearest Neighbor Search" (PAMI 2011)
• Implementation: include/index/product_quantizer.h, src/index/product_quantizer.cpp
• Tests: tests/test_product_quantizer.cpp