arXiv 2405.19504
MUVERA: Multi-Vector Retrieval via Fixed Dimensional Encodings
By Laxman Dhulipala, Majid Hadian, et al.
Published 2024-05-29
Citation lineage
Review the prior work and downstream research connected to this paper.
Neural embedding models have become a fundamental component of modern information retrieval (IR) pipelines. These models produce a single embedding per data-point, allowing for fast retrieval via highly optimized maximum inner product search (MIPS) algorithms. Recently, beginning with the landmark ColBERT paper, multi-vector models, which produce a set of embedding per data point, have achieved markedly superior per…