Representing a document with several vectors instead of one, so that different parts of it can match different queries. A single vector for a long passage averages everything together and loses the details; several vectors keep them.
Technical view
Late-interaction models such as ColBERT store per-token embeddings and score by summing maximum similarities across query tokens, giving finer-grained matching than single-vector approaches at substantially higher storage and compute cost. Compression techniques reduce the footprint. It sits between single-vector retrieval and full cross-encoder reranking in both cost and quality, and suits cases where precise term-level matching matters.
Related
Part of AI Terminology Index, section: Retrieval and Knowledge.