ColBERT и многовекторный поиск

ColBERT independently encodes query and document tokens into contextual vectors, then uses late interaction to score their match with MaxSim. Precomputing document vectors avoids running a joint transformer for every query-document pair, while query encoding and token-level similarity still occur at search time. ColBERTv2 adds residual compression to reduce the stored index. A multi-vector index stores more information per document than a single-vector representation and can require more storage. These are method and system trade-offs, not claims about market adoption or product-database roadmaps. ColPali is a separate vision-language retrieval model that embeds images of document pages as multi-vector representations and uses late interaction; its paper evaluates visually rich document-retrieval benchmarks. These are paper results, not evidence of production prevalence or RAG answer-quality gains. In a two-stage retrieval pipeline, a first-stage retriever can find a candidate set and ColBERT can rerank those candidates using MaxSim. This is a design pattern for evaluation, not evidence of a particular production deployment.