Late Interaction
Also called Late-Interaction Scoring
Comparing query and document at the token level only at the end, keeping detail a single vector loses.
Think of it like
Two teams working separately, then meeting at the last minute to match up their notes point by point.
Example
Instead of collapsing a passage into one vector, late interaction keeps its token vectors and matches them against the query’s at scoring time.
How it actually works
"Early interaction" (cross-encoders) feeds query and document together for max accuracy but can’t precompute anything. "No interaction" (bi-encoders) precomputes single vectors, fast but coarse. Late interaction splits the difference: encode separately and precompute token vectors, then do a lightweight fine-grained match at query time — much of the accuracy, much of the speed, at the cost of storage.
For product teams
The design that buys near cross-encoder quality without cross-encoder latency.
For engineers
Per-token precomputed vectors combined via MaxSim at query time; a precision/latency compromise between bi- and cross-encoders.
Related
- ColBERT — ColBERT is its best-known implementation.
- Bi-Encoder — A middle ground between bi- and cross-encoders.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome