Why it matters

ColBERT is precision retrieval. Understanding shapes advanced.

Advertisement

The architecture

Embed each token (not pooled).

Score = sum over query tokens of max over doc tokens sim.

ColBERT late interactionToken embeddingsper tokenMax-sim per query tokenover doc tokensSumfinal scoreKhattab + Zaharia 2020; PLAID for scale; more storage but better quality
ColBERT math.
Advertisement

How it works end to end

Khattab + Zaharia 2020.

PLAID for scale.

ColBERTv2 refinements.

More storage than bi-encoder.