大规模向量存储系统研究综述

摘要：大规模向量存储系统作为面向海量非结构化数据的高维向量表示载体，是提供高效存储与检索能力的核心基础设施。针对数据规模激增所带来的内存容量瓶颈与成本压力，聚焦于大规模向量索引与检索系统的演进路径。首先，概述内存图索引与聚类索引的基本机制；随后，重点分析十亿级乃至更大规模场景下的向量量化压缩、基于固态盘（SSD）的索引架构以及异构存算硬件加速技术，并从精度、延迟、吞吐量、构建更新效率及资源开销等维度对比各方案的差异；进一步，探讨向量检索与检索增强生成（RAG）的端到端协同优化；最后，总结不同技术路线的适用场景、主要瓶颈与未来研究方向。

关键词：大规模向量存储；近似最近邻检索；SSD；RAG

Abstract: Large-scale vector storage systems serve as the core infrastructure for efficiently storing and retrieving high-dimensional representations of massive unstructured datasets. To address the memory capacity bottlenecks and cost pressures arising from the explosive growth of data volumes, the evolution of large-scale vector indexing and retrieval systems is reviewed. In-memory graph-based and clustering-based indexing methods are first outlined. Then, vector quantization, solid-state drive (SSD)-based index architectures, and heterogeneous storage‑computing hardware acceleration for billion‑scale and even larger datasets are examined. These approaches are compared in terms of accuracy, latency, throughput, construction/update costs, and resource overhead. End‑to‑end collaborative optimization between vector retrieval and Retrieval‑Augmented Generation (RAG) is further discussed. Finally, applicable workloads, major bottlenecks, and future research directions are summarized.

Keywords: large-scale vector storage; approximate nearest neighbor search; SSD; RAG