LEMUR: Learned Multi-Vector Retrieval
Researchers have introduced two new methods to improve the efficiency and effectiveness of dense vector retrieval, a core component in modern machine learning systems. The first, VRSD, addresses the challenge of balancing similarity and diversity in search results by proposing a novel optimization problem and a parameter-free heuristic, demonstrating superior performance over existing baselines. The second, LEMUR, tackles the latency issue in multi-vector retrieval by formulating it as a supervised learning problem and reducing inference to single-vector search, achieving significant speedups. AI
IMPACT These advancements in vector retrieval could lead to more efficient and accurate semantic search and retrieval-augmented generation systems.