Definition
Product Quantization (PQ) is a lossy compression technique for high-dimensional vector embeddings that decomposes a vector space into a Cartesian product of smaller subspaces, quantizing each independently to drastically reduce memory footprint and accelerate Approximate Nearest Neighbor (ANN) searches.
Not to be confused with Scalar Quantization; PQ compresses groups of dimensions (sub-vectors) rather than individual values.
"A LEGO mosaic: instead of storing the exact position and color of every atom, you approximate sections of the image using a limited set of pre-defined plastic bricks from a kit."
- Vector Quantization(Prerequisite)
- IVFPQ (Inverted File Product Quantization)(Component/Hybrid Architecture)
- Codebook(Component)
- Recall vs. Latency Trade-off(Performance Metric)
Conceptual Overview
Product Quantization (PQ) is a lossy compression technique for high-dimensional vector embeddings that decomposes a vector space into a Cartesian product of smaller subspaces, quantizing each independently to drastically reduce memory footprint and accelerate Approximate Nearest Neighbor (ANN) searches.
Disambiguation
Not to be confused with Scalar Quantization; PQ compresses groups of dimensions (sub-vectors) rather than individual values.
Visual Analog
A LEGO mosaic: instead of storing the exact position and color of every atom, you approximate sections of the image using a limited set of pre-defined plastic bricks from a kit.