从 RAG 到 SAGE: 现状与展望

The emergence of large model technologies has significantly enhanced the efficiency with which humans acquire and utilize knowledge. However, in practical applications, they still confront challenges such as constrained knowledge, transfer obstacles, and hallucinations, which impede the construction of trustworthy and reliable artificial intelligence systems. Retrieval-augmented generation (RAG), by leveraging external knowledge bases and query-related retrieval, has effectively strengthened capability of large models and offers strong support for large models to master real-time, industry-specific, and private knowledge, thereby facilitating the rapid promotion and implementation of large model technologies across diverse scenarios. This paper focuses on RAG, detailing its basic principles, current development status, as well as exemplary applications, and analyzing its advantages and the challenges it faces. Based on RAG, we propose the extended framework of search-augmented generation and extension by incorporating the search module and multi-level cache management module, aiming to create a more flexible and efficient knowledge toolchain for large models.