ColBERT

A neural retrieval architecture that encodes queries and documents separately, preserves one contextual vector per token, and computes relevance through late interaction. In the original MaxSim operator, each query token finds its highest similarity among document tokens, and those maxima are aggregated into a score.

Documents remain precomputable while the comparison retains more granularity than a single-vector bi-encoder. The cost appears in larger indexes and more operations per candidate, while query processing is generally cheaper than running a full cross-encoder for every pair. ColBERT can support retrieval or reranking, and the right placement depends on recall, latency, memory, and domain measurements over the actual corpus.