Awesome Search KG

MUVERA — Qdrant Multi-Vector Search Course (Module 3)

2 min read

MUVERA in Qdrant — Multi-Vector Search Course

A written lesson (with a Colab notebook) in Qdrant’s multi-vector search course, Module 3 “Scalability and Optimization”. It explains why MUVERA is needed and how to use it in Qdrant Vector DB through FastEmbed.

Why HNSW Struggles with Multi-Vectors

The lesson argues that MaxSim is “inherently asymmetric and non-metric”, so HNSW graph navigation works poorly on raw multi-vectors and searches fall back to full scans. MUVERA gives each document one representative vector that HNSW can index, in three stages: SimHash clustering (LSH), Fixed Dimensional Encoding, and random-projection dimensionality reduction.

Implementation

  • FastEmbed 0.7.2+ — LateInteractionMultimodalEmbedding with model Qdrant/colmodernvbert, and Muvera from fastembed.postprocess built via Muvera.from_multivector_model(); process_document() / process_query() produce the encodings and muvera.embedding_size gives the dimension.
  • Parameters — k_sim=6 (2^6 = 64 buckets), dim_proj=32, r_reps=20.
  • Collection — two named vectors: muvera (cosine) and colmodernvbert (cosine, MultiVectorComparator.MAX_SIM).
  • Query — query_points with a Prefetch on muvera (limit 100), then MaxSim Reranking on colmodernvbert to a final top 10. Candidate sets are typically 100–1000 documents.

Cost

Storage roughly doubles, because both the MUVERA vector and the full multi-vector sequence are kept. Offloading the original vectors to disk is the suggested mitigation. The lesson describes accuracy as near-identical to full multi-vector search but gives no benchmark figures.

Tools

Graph View

Backlinks