Halve Your Multi-Vector Index With HierarchicalTokenPooling at 100.6% BEIR

tomaarsen · x · 2026-08-18

Late interaction is the state of the art for visual document retrieval: text queries against page images with charts and tables intact, no OCR — ColPali-family checkpoints run through the exact same two calls, with MaxSim scoring query text tokens against image patches. The post also covers HierarchicalTokenPooling (Clavié, Chaffin & Adams), which clusters each document's token vectors with Ward linkage and keeps 1/poolfactor of them: poolfactor=2 halves the index at 100.6% of unpooled BEIR performance. It can be applied per call, standalone, or baked into the model.

Related event: Sentence Transformers v6.0 Ships Late-Interaction Multi-Vector Models(27 posts)→

Original post →

More from coding & agent

coding & agent channel →