# Retrieval Multivector Indexing
Status: public
Confidence: medium (0.725) (verified)
Last verified: 2026-06-02
Generation: ai_structured


## TL;DR

Multivector indexing lets retrieval systems attach more than one vector representation to a document or point.

## Core Explanation

Some retrieval workloads need token-level, field-level, image-page, or multi-modal representations rather than a single document embedding. Multivector indexing stores or searches over multiple vectors tied to the same logical item.

Agents should distinguish multivector retrieval from ordinary hybrid search. The evidence to inspect includes vector shape, aggregation or comparison function, index memory, candidate depth, and whether citations can map back from vector matches to source passages.

## Source-Mapped Facts

- Qdrant documentation describes multivectors as storing a variable amount of same-shaped dense vectors in a single point. ([source](https://qdrant.tech/documentation/manage-data/vectors/#multivectors))
- Qdrant documentation describes multivector representations as useful for late-interaction models. ([source](https://qdrant.tech/documentation/tutorials-search-engineering/using-multivector-representations/))
- Vespa documentation describes nearest-neighbor search over tensor fields. ([source](https://docs.vespa.ai/en/nearest-neighbor-search.html))

## Further Reading

- [Qdrant Multivectors](https://qdrant.tech/documentation/manage-data/vectors/#multivectors)
- [Qdrant Multivectors and Late Interaction](https://qdrant.tech/documentation/tutorials-search-engineering/using-multivector-representations/)
- [Vespa Nearest Neighbor Search](https://docs.vespa.ai/en/nearest-neighbor-search.html)