Product Quantization (PQ) is a vector compression technique that addresses the memory and latency challenges of large-scale vector similarity search. Instead of storing full 32-bit float vectors, PQ splits each vector into m subvectors, trains a small codebook of k centroids per subspace using k-means, and encodes each vector as a sequence of centroid indices. This achieves dramatic compression (e.g., 64x) while enabling approximate distance computation via precomputed lookup tables — either Symmetric (SDC) or Asymmetric (ADC) Distance Computation. The post walks through the math, Python implementation of training, encoding, decoding, and distance computation, plus real-world integration examples in Milvus, Weaviate, and Qdrant. For RAG pipelines, PQ can reduce a 61 GB embedding index to ~480 MB, enabling billion-scale search on commodity hardware with recall as the primary tradeoff.
