摘要
检索增强生成(Retrieval-Augmented Generation, RAG)作为一种融合信息检索与文本生成的范式,通过引入外部知识源,显著提升了模型的知识覆盖率、事实准确性与时效性。系统梳理了RAG技术的演进与研究现状,从基本原理、系统架构出发,深入分析检索器与生成器的协同建模机制及典型变体,包括FiD、REALM、REPLUG、GraphRAG等。总结了RAG在开放域问答、领域知识问答、智能客服、企业知识管理等应用场景的实践效果,并分析了多模态融合、外部工具调用和小样本学习等前沿方向。在此基础上,指出了当前RAG研究所面临的评估体系不统一、训练开销高、可控性不足等挑战,并对未来发展趋势进行了展望。
Retrieval-Augmented Generation(RAG),as a paradigm that integrates information retrieval and text generation,significantly enhances the knowledge coverage,factual accuracy,and timeliness of models by incorporating external knowledge sources.This paper systematically reviews the technical evolution and research status of RAG,starting from its basic principles and system architecture.It delves into the co-modeling mechanisms of the retriever and generator and their typical variants,including FiD,REALM,REPLUG,and GraphRAG.Furthermore,it summarizes the practical effects of RAG in application scenarios such as open-domain question answering,domain knowledge question answering,intelligent customer service,and enterprise knowledge management,and explores cutting-edge directions such as multimodal fusion,external tool invocation,and few-shot learning.Based on this,the paper points out the challenges currently faced by RAG research,such as the lack of a unified evaluation system,high training costs,and insufficient controllability,and looks forward to future development trends.This paper aims to provide a systematic theoretical framework and practical reference for subsequent research.
作者
李子骏
肖辉
李雪峰
LI Zijun;XIAO Hui;LI Xuefeng(College of Electronic and Information Engineering,Tongji University,Shanghai 201804,China)
出处
《微电子学与计算机》
2025年第10期48-65,共18页
Microelectronics & Computer
基金
国家自然科学基金(62175187)。
关键词
检索增强生成
预训练语言模型
信息检索
文本生成
知识密集型任务
retrieval-augmented generation
pre-trained language models
information retrieval
text generation
knowledge-intensive tasks
