Hierarchical token pooling halves multi-vector indexes at 100.6% of BEIR performance

tomaarsen · x · 2026-08-18

The pain point of multi-vector retrieval is index size: 4,874 NQ passages become 608,414 token vectors — 311.5 MB vs 20 MB for a 1024d dense index, about 16x. HierarchicalTokenPooling (Clavié, Chaffin & Adams) clusters each document's token vectors with Ward linkage and keeps 1/poolfactor of them; at poolfactor=2 the index halves while retaining 100.6% of unpooled BEIR performance, usable per call, standalone, or baked into the model. The other escape routes: a real late-interaction index, or use it as a reranker.

Related event: Sentence Transformers v6.0 ships with first-class late interaction models(33 posts)→

Original post →

More from coding & agent

coding & agent channel →