IVF-TQ: Codebook-Free Residual Compression for Streaming Vector Search

A new article on arXiv has unveiled IVF-TQ, a novel inverted-file index aimed at approximate nearest neighbor searches, featuring a data-independent residual compression layer. Unlike common methods like PQ, OPQ, and ScaNN, which rely on adjusting a codebook based on a specific dataset and suffer from decreased recall as databases grow, IVF-TQ uses a fixed random rotation along with a pre-calculated Lloyd-Max scalar quantizer that depends only on the bit width b and dimension d. The only part that requires training is the coarse k-means partition of IVF. This method provides a consistent inner-product error bound across the sphere, a significant advantage over learned-codebook approaches. Experiments reveal that product quantization can cause a -3.8pp drop in recall, highlighting the problem that IVF-TQ addresses.