1.  [](/)
2.  [Diagrams](/diagrams)
3.  Multi-Vector Search vs. a Single Embedding

[Back to Diagrams](/diagrams)Diagram Series

# Multi-Vector Search vs. a Single Embedding

Multi-Vector Search vs. a Single Embedding

![Multi-vector (late-interaction) search versus a single embedding: instead of compressing a whole document into one vector, the model keeps one vector per token or patch and scores a query against all of them with a MaxSim operator, which recovers detail a single averaged vector loses.](/diagrams/multi-vector-search.webp)Full screen

Multi-Vector Search vs. a Single Embedding

Most vector search stores one embedding per document. That works for text paragraphs. It doesn't work when your document is a 60-second video with speech, faces, scene transitions, and on-screen text.

Multi-vector search stores N vectors per document: one per visual frame, transcript segment, detected face, OCR region. The document isn't a point in vector space. It's a set of points.

The scoring mechanism: late interaction (MaxSim). Every query vector is scored against every document vector. Take the max similarity per query vector. Sum them. That's your relevance score.

Why this matters for multimodal:

→ Cross-modal matching (a text query matches visual, audio, and text vectors independently

→ No information loss) every frame and segment keeps its own embedding

→ Fine-grained relevance, partial matches contribute at the token-to-vector level

Single-vector search collapses your data into one point and hopes for the best. Multi-vector search keeps full fidelity and scores at the granularity that actually matters.

mixpeek.com

Run this on your own data

Mixpeek turns video, images, audio, and documents in your object storage into searchable, timestamped results through one API.

[Search your own data](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=diagram_detail&utm_campaign=mvs_onramp)

## Related diagrams

[

![A variable-size set of token vectors becomes one fixed-size vector via SimHash bucketing, projection, and repetition: no neural encoder required.](/diagrams/muvera-fde-pipeline.webp)

MUVERA: How a Multi-Vector Set Becomes One Fixed Vector



](/diagrams/muvera-fde-pipeline)[

![Documents fill empty buckets with a mean; queries sum and leave empty buckets zero: the asymmetry is what makes the FDE dot product approximate Chamfer / MaxSim.](/diagrams/muvera-query-doc-asymmetry.webp)

MUVERA: Why Query and Document Are Encoded Asymmetrically



](/diagrams/muvera-query-doc-asymmetry)[

![Tier 1 FDE-encodes the query and does fast single-vector ANN over all N to get a top-K; tier 2 reranks only those K with exact Chamfer over the original token vectors.](/diagrams/muvera-two-tier-retrieval.webp)

MUVERA: Two-Tier Retrieval (Cheap FDE Recall, Exact Rerank)



](/diagrams/muvera-two-tier-retrieval)