Product Quantization with K-means++ Initialization for Embedded Vector Search

We present an optimized implementation of Product Quantization (PQ) for high-dimensional vector similarity search, incorporating K-means++ initialization for improved codebook quality and Asymmetric Distance Computation (ADC) for efficient search. Our implementation achieves 8-16x memory compression with 95-98% search accuracy (recall@10) while enabling 100-1000x search speedup through distance table precomputation. The system is designed for embedded and resource-constrained environments, making billion-scale vector search practical on edge devices and mobile platforms.