Source: Clavié et al. 2024: Token Pooling for Multi-Vector Retrieval

Answer.AI paper showing that clustering and mean-pooling similar document token vectors at index time cuts ColBERT storage by 50% with almost no loss.

Contents

Summary

Clusters each document’s token vectors (hierarchical clustering) and replaces each cluster by its mean, so fewer vectors are stored. It is an index-time step: no model change, no query-time work.

Key claims

  • Halving the vector count (pool factor 2) gives virtually no retrieval degradation.
  • Reducing vector count by 66-75% (pool factors 3-4) keeps degradation under 5% on most datasets.

Feeds

Late-interaction retrieval

Suggest changes →