The article presents a new approach to maximum inner product search (MIPS), arguing that conventional approximate retrieval methods assume query vectors resemble key vectors, an assumption that breaks down in RAG where short user questions are matched against long document keys. The proposed method targets this geometric mismatch and outperforms prior systems notably when query volume is large.
Termsslug: maximum-inner-product-search
Maximum Inner Product Search
最大内积搜索
The retrieval problem of finding the key vector with the largest inner product to a query, the core operator behind recommendati
Curated Intelligence
Hacker News AIT2·Tools
58 pts
Apple Machine Learning ResearchT1·Tools
62 pts
Targeting the computational bottleneck of maximum inner product search (MIPS) in RAG, this work argues that existing approximate methods model only key-vector geometry and ignore the mismatch between short user queries and long documents. The authors propose a retrieval scheme that exploits query distribution properties to cut cost while preserving accuracy, extending it to private search settings.