Multi-Vector Retrieval

A single vector compresses an item into one representation. Multiple vectors can retain separate facets, such as passages, summaries or visual regions. One system may retrieve these views and map hits back to a parent document, while another computes token-level interactions and aggregates their scores. ColBERT is a specific late-interaction approach that compares query-token vectors with document-token vectors. These designs share representational multiplicity but differ in scoring, indexing and storage requirements. Multi-vector retrieval is therefore an umbrella practice, not an automatic synonym for ColBERT or for splitting a document into independently returned chunks.

The practitioner states what each vector represents and how vector hits are deduplicated, combined or scored at source-item level. Evaluation compares with a single-vector baseline and checks whether items with more vectors receive an unintended advantage. Source mappings and update procedures must keep all representations synchronized. Useful artifacts include the representation schema, aggregation function, storage measurements and relevance results. For late interaction, token or patch handling and supported index mechanisms require explicit configuration. Latency and memory are assessed alongside quality gains.

More representations can increase storage, indexing and scoring cost without improving relevant retrieval. Poor aggregation can overcount duplicate evidence or bias ranking toward large documents. Generated views may also contain unsupported summaries. Quality checks need original source traceability and a clearly defined scoring rule. Any claimed improvement should identify the multi-vector design and baseline, since gains from one token-level method do not establish the value of every chunk or view aggregation strategy.