Efficient Multivector Retrieval with Token-Aware Clustering and Hierarchical Indexing
2026-04-30 • Information Retrieval
Information RetrievalMachine Learning
AI summaryⓘ
The authors present TACHIOM, a new system for retrieving information that improves both speed and accuracy by using token-level data more efficiently. Unlike traditional methods that use k-means clustering, which can be slow and biased towards frequent words, TACHIOM adjusts for token frequency when organizing data clusters, allowing it to handle millions of groups quickly. This system uses a graph-based index and a special quantization method to quickly score documents without expensive detailed computations. Their experiments show TACHIOM is much faster at clustering and retrieval while maintaining or improving accuracy compared to existing systems.
multivector retrievaltoken-level representationk-means clusteringcentroid allocationproduct quantizationgraph-based indexingdocument scoringMS-MARCOLoTTE datasetretrieval speedup
Authors
Silvio Martinico, Franco Maria Nardini, Cosimo Rulli, Rossano Venturini
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.