Multi-Vector Retrieval
Also called Multi-Representation Retrieval
Representing one document with several vectors instead of one, to capture more of its meaning.
Think of it like
Describing a house with separate photos of each room rather than a single blurry exterior shot.
Example
A long document is stored as one vector per section, so a query can match the exact section it needs, not a mushy whole-doc average.
How it actually works
A single embedding for a long or multi-topic document averages everything into one point, blurring specifics. Multi-vector retrieval keeps several embeddings per item — per chunk, per token, or per generated summary — so matching happens at finer grain. It improves recall and precision on complex content, at the cost of more storage and more vectors to search.
For product teams
Helps when documents cover many topics and single-vector search feels imprecise.
For engineers
Multiple embeddings per document (chunk/token/summary level); finer matching, higher index size and search cost.
Related
- ColBERT — ColBERT is a well-known approach.
- Dense Retrieval — Refines ordinary dense retrieval.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome