Module 3: MUVERA

Load the ColModernVBERT model and initialize the MUVERA transform. MUVERA converts variable-length multi-vector representations into fixed-size single vectors that simulate MaxSim using standard dot product.

Create a collection with both vector types: a MUVERA single-vector for fast HNSW-based retrieval and the original multi-vector for precise MaxSim reranking.

Generate both MUVERA and multi-vector embeddings for a document image.

Upload both vector representations as a single point.

Encode the query in both formats: multi-vector for MaxSim reranking and MUVERA for fast candidate retrieval.

Run two-stage retrieval: MUVERA prefetch retrieves 100 candidates using fast single-vector search, then ColModernVBERT reranks them with precise MaxSim.

Display the final search results after two-stage retrieval.