Sparcent: faster multi-vector dense retrieval with sparse centroid indices

Abstract

Multi-vector dense retrieval methods have set the state-of-the-art retrieval quality on ranking tasks. However, the superior effectiveness of such methods comes with substantial computational overhead, which makes their deployment in large-scale search engines infeasible. As a remedy, we propose a novel approach, Sparcent, that leverages learned sparse retrieval (LSR) to accelerate the query-centroid matching stage in multi-vector dense retrieval. Specifically, we first derive sparse centroid representations for multi-vector term-embedding clusters and then prune them, producing compact centroid vectors that remain both efficient and effective for the indexing via LSR methods. Next, we integrate Sparcent with two well-known multi-vector retrieval systems, PLAID and WARP, already optimized for single-core CPU search. Our experimental results show that Sparcent provides up to 4.6×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} speed-up over the original PLAID and 1.3×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} speed-up over WARP, respectively, with virtually zero additional memory, while maintaining result effectiveness on both in-domain and out-of-domain retrieval tasks.