TACHIOM: Token-Aware Multivector Retrieval Achieves 247x Faster Clustering and 9.8x Search Speedup (SIGIR 2026)

arXiv (SIGIR 2026)·medium signalXBlueskyLinkedInCopy link

Accepted at SIGIR 2026, TACHIOM introduces token-aware centroid allocation that accounts for token distribution rather than treating all tokens equally, combined with graph-based indexing over centroids and product quantization for final scoring. The system achieves up to 247x faster clustering than k-means and up to 9.8x retrieval speedup over state-of-the-art systems while maintaining comparable or superior effectiveness on MS-MARCO and LoTTE benchmarks. For builders running RAG systems with ColBERT-style multivector models, this addresses the primary scaling bottleneck.