Efficient Multivector Retrieval with Token-Aware Clustering and Hierarchical Indexing

2026-04-30

AI Summary

This work addresses the high computational and memory overhead of multi-vector retrieval models caused by token-level representations, as well as the poor scalability and bias toward frequent tokens exhibited by traditional k-means clustering in large-scale settings. To overcome these limitations, the authors propose TACHIOM, a novel system that integrates token-aware mechanisms into the clustering process, dynamically allocating cluster centroids based on token distributions. By combining hierarchical graph indexing with an optimized product quantization layout, TACHIOM enables efficient and accurate document scoring using only centroid representations, eliminating costly token-level computations. The approach scales to millions of clusters while maintaining high retrieval quality. Experiments on MS-MARCOv1 and LoTTE demonstrate a 247× speedup in clustering over k-means and a 9.8× retrieval acceleration, achieving performance on par with or superior to state-of-the-art systems.

Abstract

Multivector retrieval models achieve state-of-the-art effectiveness through fine-grained token-level representations, but their deployment incurs substantial computational and memory costs. Current solutions, based on the well-known k-means clustering algorithm, group similar vectors together to enable both effective compression and efficient retrieval. However, standard k-means scales poorly with the number of clusters and dataset size, and favours frequent tokens during training while underrepresenting rare, discriminative ones. In this work, we introduce TACHIOM, a multivector retrieval system that exploits token-level structure to significantly accelerate both clustering and retrieval. By accounting for tokens' distribution during centroid allocation, TACHIOM easily scales to millions of centroids, enabling highly accurate document scoring using only centroids, avoiding expensive token-level computation. TACHIOM combines a graph-based index over centroids with an optimized Product Quantization layout for efficient final scoring. Experiments on MS-MARCOv1 and LoTTE show that TACHIOM achieves up to $247\times$ faster clustering than k-means and up to $9.8\times$ retrieval speedup over state-of-the-art systems while maintaining comparable or superior effectiveness.

Problem

• multivector retrieval
• k-means clustering
• token-level representation
• computational efficiency
• rare token underrepresentation

Innovation

• multivector retrieval
• token-aware clustering
• hierarchical indexing
• product quantization
• centroid-based scoring