VQ-bench: a Composable Vector Quantization Framework

This post introduces VQ-bench, an open-source framework for evaluating vector quantization methods. It formalizes quantizers and primitives, enabling the composition of complex quantization pipelines. The framework includes a website with benchmarks, a GitHub repository for contributions, and a research paper. The post details experimental results for reconstruction error, recall@10, and encode time across various datasets and quantizers, highlighting trends and discussing the trade-offs between different methods like PQ, OPQ, EDEN, and E-RaBitQ. It also outlines contribution guidelines for new quantizers and primitives.