基于大模型检索增强生成的水利问答架构设计

打开文本图片集

中图分类号：TP391；TP18 文献标识码：A 文章编号：2096-4706（2026）03-0052-06

Abstract：Toaddressthe challenges ofknowledge hallucination，poor information timelinessandlowlocalized deployment effciency faced by Large Language Models （LLMs）in their applications to water conservancy and other knowledge-intensive industries，this paper proposes a Knowledge-Enhanced Retrieval-Augmented Generation （KE-RAG）system architecture that integrates vector retrieval with Knowledge Graph （KG）constraints.Thisarchitecture takesthedomesticalldevelopedopensource modelQwen3-32B-AWQas the generationcore，constructsamultimodal waterconservancyknowledgebasecontaining lawsandregulations，techncalstandards，engineringcasesandexpert experience，andachieves high-performancelocalized inference throughthevectorizedLarge LanguageModel（vLLM）framework.Experimentalresultsshowthatontheevaluation set， this architecture reaches aprofessional knowledgequestion and answer accuracy of 91.5% ，anaverage end-to-end response latency of less than 500ms and a throughput nearly four times higher than that of standard deployment， with all metrics significantlyoutperforming those ofbaseline models，which providesa reference solution for the applicationofLargeLanguage Models in the water conservancy industry.

Keywords： Large Language Model; retrieval-augmented generation; Knowledge Graph; waterconservancy knowledge questionand answer; high-performance inference; localized deployment

0 引言

水利行业作为关系国计民生的基础产业，其业务运行高度依赖于海量的专业知识，包括复杂的法律法规、精细的技术标准以及丰富的工程案例。（剩余6167字）