MRVQ: One Resident Index for Dimension- and Rate-Elastic Vector Search
MRVQ keeps one vector-search index resident while varying both embedding dimension and code rate.
The paper presents a post-hoc residual quantizer for frozen embeddings whose maximum-rate code can be truncated by residual stage or coordinate count. In the authors’ tests across FiQA, NFCorpus, four embedding families, and 4-, 8-, and 16-byte codes, it was the lowest-RAM design evaluated. The tradeoff is quality: per-rate QINCo2 beat MRVQ by 0.026–0.107 nDCG@10 on FiQA. MRVQ still outperformed PQ, OPQ, and AdANNS-OPQ at matched code size, making it a memory-saving option rather than a universal winner. ArXiv · AI/CL/LG's note
The paper presents a post-hoc residual quantizer for frozen embeddings whose maximum-rate code can be truncated by residual stage or coordinate count. In the authors’ tests across FiQA, NFCorpus, four embedding families, and 4-, 8-, and 16-byte codes, it was the lowest-RAM design evaluated. The tradeoff is quality: per-rate QINCo2 beat MRVQ by 0.026–0.107 nDCG@10 on FiQA. MRVQ still outperformed PQ, OPQ, and AdANNS-OPQ at matched code size, making it a memory-saving option rather than a universal winner. ArXiv · AI/CL/LG's note
score 5