MRVQ: One Resident Index for Dimension- and Rate-Elastic Vector Search
arXiv cs.AI · 3 Oct 2026, 03:33 AEST
MRVQ introduces a post-hoc residual quantizer for frozen embeddings, allowing for efficient dimension and rate adjustments in dense-retrieval services. It significantly reduces memory usage compared to traditional methods while maintaining competitive performance.
Key facts
- Tuning a quantizer separately for each rate gives the best quality, but the retrieval tier then holds several code streams and quantizer states at once.
- The saving is not free: per-rate QINCo2 is 0.026-0.107 nDCG@10 better on FiQA.
- MRVQ beats PQ, OPQ, and AdANNS-OPQ at matched code size.
- MRVQ is therefore a low-memory operating point for elastic retrieval, not a universal quality winner.
Read aloud
MRVQ introduces a post-hoc residual quantizer for frozen embeddings, allowing for efficient dimension and rate adjustments in dense-retrieval services. It significantly reduces memory usage compared to traditional methods while maintaining competitive performance. Tuning a quantizer separately for each rate gives the best quality, but the retrieval tier then holds several code streams and quantizer states at once. The saving is not free: per-rate QINCo2 is zero point zero two six-zero point one zero seven nDCG@ten better on FiQA. MRVQ beats PQ, OPQ, and AdANNS-OPQ at matched code size. MRVQ is therefore a low-memory operating point for elastic retrieval, not a universal quality winner.
Sources in this cluster
- MRVQ: One Resident Index for Dimension- and Rate-Elastic Vector Search
arXiv cs.AI · 3 Oct 2026, 03:33 AEST