NLP & Language · advanced · concept 151 of 311 · Updated Nov 4, 2026

🧩ColBERT and Late Interaction Retrieval

Late interaction keeps one vector per token for the query and for each document, then scores relevance at query time by matching every query token to its best document token (MaxSim) and adding the scores. ColBERT introduced it in 2020 and ColBERTv2 shrank the index. It keeps most of a cross-encoder's accuracy at close to single-vector speed. Example: ColPali applies the same scoring to page images, so PDF search needs no OCR.