MaxSim is soft alignment: 'live' matches 'inhabit' at 0.94, zero shared characters

tomaarsen · x · 2026-08-18

The key to understanding multi-vector retrieval is reading MaxSim as soft alignment: every query token points at the document token that best explains it, and the alignment isn't lexical — between "Where do penguins live?" and "Penguins inhabit Antarctica", live finds inhabit at 0.94 despite sharing no characters. A dense model compresses whole texts into one vector, averaging away token-level matching; a multi-vector model keeps a vector per token and sums each query token's best match in the document.

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

Original post →

More from coding & agent

coding & agent channel →