FastEmbed
Qdrant’s Python embedding library. Besides dense and sparse models it serves late-interaction multi-vector models (ColBERT, ColModernVBERT) and, from version 0.7.2, ships a MUVERA postprocessor that turns a multi-vector embedding into a fixed-dimensional single vector.
MUVERA Postprocessing
Muvera lives in fastembed.postprocess and is built from a late-interaction model with Muvera.from_multivector_model() — the course example pairs it with LateInteractionMultimodalEmbedding("Qdrant/colmodernvbert") and k_sim=6, dim_proj=32, r_reps=20. process_document() and process_query() produce the asymmetric document/query encodings, and muvera.embedding_size gives the resulting dimension. The Qdrant course uses these encodings as the prefetch stage in Qdrant Vector DB, with MaxSim reranking on the original multi-vectors.
FastEmbed is also how the AWS sample computes MUVERA encodings client-side for Amazon OpenSearch Service, where server-side plugins can’t be installed (Scaling Late Interaction Retrieval in OpenSearch with MUVERA).