1.  [](/)
2.  [Diagrams](/diagrams)
3.  MUVERA: Two-Tier Retrieval (Cheap FDE Recall, Exact Rerank)

[Back to Diagrams](/diagrams)MUVERA / Multi-Vector

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

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

![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)Full screen

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

MUVERA serves multi-vector quality at single-vector speed by splitting retrieval into two tiers. Tier 1 is cheap and runs over all N documents: the query is FDE-encoded to one fixed vector and a standard single-vector ANN index (HNSW, IVF, or DiskANN) returns a top-K candidate set (K around 100-500) in milliseconds, each document is just one vector here. Tier 2 is expensive but only ever sees those K candidates: it reranks them with exact Chamfer / MaxSim over the ORIGINAL token vectors, producing the final ranking. The tradeoff most vendor posts gloss over is storage: you keep BOTH representations, the FDE vectors in the ANN index (one per document, serving tier 1) and the full ~100 token vectors per document on disk (served only for the K rerank candidates). Drop the second copy and you cannot rerank; keep it and your storage roughly doubles. Recall quality is bought with disk, not magic.

## Where this diagram appears

*   [Guide: Late-Interaction Retrieval (ColBERT, ColPali, MUVERA) →](/guides/late-interaction-retrieval)

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

[

![Reranking is the precision second stage of retrieval: a fast first-stage retriever recalls a broad candidate set optimized for recall, then a cross-encoder reranker reads each query-document pair together and re-scores them so the most relevant results rise to the top before they reach the model. Fast recall first, precise scoring second.](/diagrams/foundations/reranking.webp)

Wide net first, careful read second



](/diagrams/reranking)[

![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)