基于RAG的企业级智能知识库系统设计与实现
Journal: Engineering and Management Science DOI: 10.32629/ems.v8i6.20585
Abstract
针对企业技术文档分散、传统关键词检索语义理解能力不足、知识获取效率低的问题，本文设计并实现了一套基于检索增强生成（Retrieval-Augmented Generation，RAG）的企业级智能知识库系统。研究方法上，系统采用数据层、检索层、生成层的三层架构，以本地化部署的DeepSeek大语言模型为生成核心，结合Milvus向量数据库与Elasticsearch关键词检索引擎构建混合检索策略，并引入交叉编码器Rerank对候选文档精排。在2347份企业技术文档、200组评估问答对上的对比实验表明，混合检索结合Rerank方案的Recall@10达到92.3%，较纯向量检索基线提升18.7个百分点，MRR提升0.211，端到端平均响应时间控制在2.1秒以内。研究结论表明，所提方案可有效提升企业私域知识的检索精度与问答准确性，已在企业内部稳定上线，为技术密集型企业的知识资产数字化管理提供了可行的工程实践范式。
Keywords
检索增强生成；大语言模型；混合检索；Rerank；知识库
Full Text
PDF - Viewed/Downloaded: 8 TimesReferences
[1] Lewis P, Perez E, Piktus A, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks[C]//Advances in Neural Information Processing Systems, 2020, 33: 9459-9474.
[2] Gao Y, Xiong Y, Gao X, et al. Retrieval-augmented generation for large language models: A survey[EB/OL]. (2023-12-18)[2025-04-10]. https://arxiv.org/abs/2312.10997.
[3] Xiao S, Liu Z, Zhang P, et al. C-Pack: Packaged resources to advance general Chinese embedding[EB/OL]. (2023-09-14)[2025-04-10]. https://arxiv.org/abs/2309.07597.
[4] Wang J, Yi X, Guo R, et al. Milvus: A purpose-built vector data management system[C]// Proceedings of the 2021 International Conference on Management of Data (SIGMOD), 2021: 2614-2627.
[5] Cormack G V, Clarke C L A, Buettcher S. Reciprocal rank fusion outperforms condorcet and individual rank learning methods[C]//Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, 2009: 758-759.
[2] Gao Y, Xiong Y, Gao X, et al. Retrieval-augmented generation for large language models: A survey[EB/OL]. (2023-12-18)[2025-04-10]. https://arxiv.org/abs/2312.10997.
[3] Xiao S, Liu Z, Zhang P, et al. C-Pack: Packaged resources to advance general Chinese embedding[EB/OL]. (2023-09-14)[2025-04-10]. https://arxiv.org/abs/2309.07597.
[4] Wang J, Yi X, Guo R, et al. Milvus: A purpose-built vector data management system[C]// Proceedings of the 2021 International Conference on Management of Data (SIGMOD), 2021: 2614-2627.
[5] Cormack G V, Clarke C L A, Buettcher S. Reciprocal rank fusion outperforms condorcet and individual rank learning methods[C]//Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, 2009: 758-759.
Copyright © 2026 何子晗
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License