A Google method for multi-vector retrieval that compresses many per-token vectors into a single fixed-dimensional encoding for fast search.
MUVERA (Multi-Vector Retrieval via Fixed Dimensional Encodings) is a Google Research method, published at NeurIPS 2024, for making multi-vector retrieval fast and scalable. Multi-vector models represent a document as many per-token vectors, which is expressive but expensive to search; MUVERA compresses that set into a single fixed-dimensional encoding that standard nearest-neighbour search can handle.
The authors — Laxman Dhulipala, Majid Hadian, Jason Lee, Rajesh Jayaram and Vahab Mirrokni — sit in Google's retrieval and large-scale-algorithms groups, whose broader work spans parallel clustering and approximate nearest-neighbour search. That lineage is a strong hint at the kind of retrieval machinery behind Google's systems.
For AI SEO, MUVERA is worth knowing because it points at how modern retrieval keeps the richness of token-level matching without the cost — sitting alongside vector search and vector embeddings in the retrieval stack.
