A Survey on Query Processing in Vector Databases

Abstract

High-dimensional vectors have become a fundamental data representation in modern applications, such as information retrieval and large language model systems, making vector databases and their query processing an essential research area. While approximate nearest neighbor search has long been the central primitive, modern vector workloads increasingly involve richer query types, including filtered similarity search, multi-vector similarity search, and similarity join. These developments substantially expand the design space of vector query processing and make it harder to obtain a clear and structured view of existing techniques. This survey presents a comprehensive review of query processing in vector databases. We first formalize four query types: similarity search, filtered similarity search, multi-vector similarity search, and similarity join. We then organize existing studies under a unified taxonomy. In particular, we review proximity graphs and quantizations, the two state-of-the-art approaches for similarity search, together with related directions such as distance computation, hard-query processing, and secure search. We further summarize universal and dedicated approaches for filtered similarity search, different methods for multi-vector similarity search, and both exact and approximate algorithms for similarity join. Through this survey, we provide a structured view of current approaches, highlight their connections and differences, and discuss open challenges and future directions.