Disk-Resident Vector Similarity Search: A Survey

Vector similarity search (VSS) has become ubiquitous in modern multimodal data retrieval systems, supporting a wide range of applications such as document retrieval, music recognition, image search, and code assistants. With the rapid growth of vector datasets and the prohibitive cost of main memory, VSS techniques are increasingly shifting from memory-resident to disk-resident settings. Unlike memory-resident designs, which keep both raw vectors and index structures entirely in memory, disk-resident VSS places raw vectors and fine-grained index structures on secondary storage while retaining only lightweight components in memory. Therefore, disk-resident VSS must explicitly account for data access costs (i.e., I/O) and block-level layouts, leading to index organizations and query execution strategies that are fundamentally different from those used in memory-resident systems.

This survey provides the first systematic review of disk-resident VSS, a critical yet challenging task for modern large-scale data systems. We introduce a unified framework that categorizes existing techniques into IVF-based, graph-based, and tree-based paradigms, based on the primary index structure used for filtering. For each category, we further decompose the overall design into key technical components, including index construction, block-aware layouts, query execution strategies, and update mechanisms. Moreover, we summarize commonly used datasets to facilitate reproducible benchmarking, and identify open challenges and promising directions for future research.