ColBERT

ColBERT

ColBERT (Contextualized Late Interaction over BERT) is a retrieval model that combines the efficiency of Bi-Encoder (pre-computed document representations) with fine-grained token-level matching via late interaction.

MaxSim Scoring

$s(q,d)={\sum }_{i=1}^{|q|}{\max }_{j=1}^{|d|}{\mathbf{q}}_i^{\top }{\mathbf{d}}_j$

For each query token, find its best-matching document token (max similarity), then sum across all query tokens.

Key properties:

• Documents encoded independently → can be pre-computed and indexed
• Token-level representations (not single-vector) → richer matching than standard bi-encoder
• Uses Approximate Nearest Neighbor search with deferred interaction
• Much faster than cross-encoder, more effective than single-vector bi-encoder

Appears In

• IR-L06 - Dense Retrieval