Vector Quantization
A compression method that maps a high-dimensional vector to one or more representative codebook entries, reducing storage and search cost at the price of approximate distance information.
Keeping every embedding in full floating-point form can make vector search memory- and bandwidth-intensive. Vector quantization replaces the original vector with compact codebook references or subspace codes, allowing approximate distances to be computed from a much smaller representation.
The saving trades against recall. A poorly matched codebook, aggressive compression, or distribution shift can disturb nearest-neighbor ordering. FAISS index architecture for vector similarity search shows where this technique fits in an actual retrieval index.