Source: Dhulipala et al. 2024: MUVERA

Google Research paper that compresses a multi-vector set into one fixed-dimensional encoding so ordinary single-vector ANN can fetch candidates for MaxSim reranking.

Contents

Summary

Instead of building a special multi-vector index, MUVERA asymmetrically maps the query’s and each document’s vector sets to single Fixed Dimensional Encodings (FDEs) whose inner product approximates multi-vector (Chamfer/MaxSim) similarity. Any off-the-shelf maximum-inner-product search index can then retrieve candidates, which are reranked with exact MaxSim.

Key claims

  • First single-vector proxy for multi-vector similarity with theoretical approximation guarantees.
  • Reaches the same recall as prior heuristics while retrieving 2-5x fewer candidates.
  • On BEIR, averages 10% higher recall with 90% lower latency than the prior state-of-the-art implementation.

Feeds

Late-interaction retrieval

Suggest changes →