基于检索增强生成的化工领域大模型智能问答

Intelligent Question-Answering for Chemical Engineering Based on Large Language Models with Retrieval-Augmented Generation Technology

摘要

化工设备设计需要严格依照标准规范．然而标准规范数量多、内容上相互引用,设计人员面对非常规的设计要求或设计问题时很难准确、全面地查找到所有涉及的标准规范条目．利用检索增强生成(RAG)技术结合大语言模型(LLM)可以对设计要求或设计问题进行准确回答的同时分析并提供相应标准规范内容,从而避免遗漏相关的标准规范．然而,由于化工设备设计领域知识库中具有大量公式、图表等复杂数据,如何构建相应的结构化RAG数据库实现LLM在化工设备设计领域的智能问答尚不明确．针对上述问题,本文提出了一种垂直领域的复杂数据智能问答系统构建一体化框架,该框架结合提示工程方法与多个视觉语言模型以实现RAG数据库的构建,采用语义检索与重排序技术,并选取嵌入模型与大语言模型分别作为检索器与生成器,以实现基于 RAG 的智能问答．基于该框架,本文构建了化工设备设计领域的智能问答系统,并使用Qwen2.5-72b和Qwen2.5-7b模型在以GB/T 150—2011规范为主的压力容器设计问答数据集上进行实验．结果表明,本文所提出的框架在复杂数据提取的准确性上优于现有技术,并通过RAG 技术显著提升了问答系统的性能．相比于未结合RAG的技术,Qwen2.5-72b和Qwen2.5-7b模型的准确率分别提高了19.3%和17.7%．此外还对生成器接受的文档块数量对问答系统准确性的影响与设备设计领域数据的泛化性能进行了研究．

Abstract

Chemical equipment design requires strict adherence to standards and specifications. However,the large number of standards and their cross-references make it difficult for designers to accurately and comprehensively locate all relevant standard entries when dealing with unconventional design requirements or issues. In this situation,design-related questions can be answered accurately by combining retrieval-augmented generation(RAG)technology with large language model(LLM)when analyzing and providing the corresponding standard content,thus preventing omissions of pertinent standards. As the chemical equipment design knowledge base includes numerous formulas,charts,and other complex data,methods to construct a structured RAG database to enable LLM to perform intelligent question-answering remain unclear. To address this issue,this study proposes an integrated framework for building an intelligent question-answering system tailored to vertical domains with complex data. This framework combines prompt engineering and multiple vision-language models to build an RAG database. It adopts semantic retrieval and reranking techniques and specifically selects embedding models and LLM as the retriever and generator,respectively,to implement intelligent question-answering based on RAG. Based on this framework,an intelligent question-answering system is constructed for the field of chemical equipment design. Experiments are conducted using Qwen2.5-72b and Qwen2.5-7b models on a pressure vessel design question-answering dataset primarily based on GB/T 150—2011. The results show that the proposed framework outperforms existing technologies in extracting complex data accurately and significantly improves the performance of the question-answering system via the RAG. Compared with the non-RAG setting,the accuracy of the Qwen2.5-72b and Qwen2.5-7b models is higher by 19.3% and 17.7%,respectively. In addition,the impact of the number of input document blocks on the accuracy of the question-answering system and the data generalization performance in the equipment design domain are examined.

关键词

大语言模型 / 检索增强生成 / 化工设备设计 / 智能问答 / 复杂数据信息提取

Key words

large language model / retrieval-augmented generation / chemical equipment design / intelligent question-answering / complex data extraction

参考文献

[1] Devlin J, Chang M W, Lee K, et al. Bert：Pre—training of deep bidirectional transformers for language understanding[C]// Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics：Human Language Technologies. Minneapolis,USA, 2019:4171-4186.

[2] Raffel C, Shazeer N, Roberts A, et al. Exploring the limits of transfer learning with a unified text—to—text transformer[J]. Journal of Machine Learning Research, 2020, 21:1-67.

[3] Radford A, Wu J, Child R, et al. Language models are unsupervised multitask learners[J]. OpenAI Blog, 2019, 1(8):9.

[4] Brown T, Mann B, Ryder N, et al. Language models are few—shot learners[C]// Advances in Neural Information Processing Systems33. 2020:1877-1901.

[5] Wang X, Yue Q R, Liu X G. Crack image classification and information extraction in steel bridges using multimodal large language models[J]. Automation in Construction, 2025, 171:105995.

[6] Lewis P, Perez E, Piktus A, et al. Retrieval—augmented generation for knowledge—intensive nlp tasks[C]// Advances in Neural Information Processing Systems33. 2020:9459-9474.

[7] 张玉铭, 李红岩, 郎许锋, 等. 基于检索增强生成技术的中医药问答大语言模型的构建[J]. 南京中医药大学学报, 2024, 40(12):1375-1382.

[8] Zhang Yuming, Li Hongyan, Lang Xufeng, et al. Construction of traditional Chinese medicine question—answering large language model based on retrieval—augmented generation technology[J]. Journal of Nanjing University of Traditional Chinese Medicine, 2024, 40(12):1375-1382(in Chinese).

[9] 江双五, 张嘉玮, 华连生, 等. 基于大模型检索增强生成的气象数据库问答模型实现[J]. 计算机工程与应用, 2025, 61(5):113-121.

[10] Jiang Shuangwu, Zhang Jiawei, Hua Liansheng, et al. Implementation of meteorological database question—answering based on large—scale model retrieval—augmentation generation[J]. Computer Engineering and Applications, 2025, 61(5):113-121(in Chinese).

[11] 张艳萍, 陈梅芳, 田昌海, 等. 面向军事领域知识问答系统的多策略检索增强生成方法[J]. 计算机应用, 2025, 45(3):746-754.

[12] Zhang Yanping, Chen Meifang, Tian Changhai, et al. Multi—strategy retrieval—augmented generation method for military domain knowledge question answering systems[J]. Journal of Computer Applications, 2025, 45(3):746-754(in Chinese).

[13] Zheng S W, Pan K, Liu J, et al. Empirical study on fine—tuning pre—trained large language models for fault diagnosis of complex systems[J]. Reliability Engineering & System Safety, 2024, 252:110382.

[14] Tao L F, Liu H F, Ning G A, et al. LLM—based framework for bearing fault diagnosis[J]. Mechanical Systems and Signal Processing, 2025, 224:112127.

[15] Sanh V, Webson A, Raffel C, et al. Multitask prompted training enables zero—shot task generalization[C]// The 10th International Conference on Learning Representations. 2022:186704.

[16] Radford A, Kim J W, Hallacy C, et al. Learning transferable visual models from natural language supervision[C]// Proceedings of the 38th International Conference on Machine Learning. 2021:8748-8763.

[17] Zhang J Y, Huang J X, Jin S, et al. Vision—language models for vision tasks：A survey[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(8):5625-5644.

[18] Zhao R H, Tang J Y, Zeng W X, et al. Towards human—like questioning：Knowledge base question generation with bias—corrected reinforcement learning from human feedback[J]. Information Processing & Management, 2025, 62(3):104044.

[19] Rafailov R, Sharma A, Mitchell E, et al. Direct preference optimization：Your language model is secretly a reward model[C]// Advances in Neural Information Processing Systems36. New Orleans,USA, 2023:53728-53741.

[20] Wei J, Wang X Z, Schuurmans D, et al. Chain—of—thought prompting elicits reasoning in large language models[C]// Advances in Neural Information Processing Systems35. New Orleans,USA, 2022:24824-24837.

[21] Gilbert S, Kather J N, Hogan A. Augmented non—hallucinating large language models as medical information curators[J]. NPJ Digital Medicine, 2024, 7(1):100.

[22] Fan W Q, Ding Y J, Ning L B, et al. A survey on rag meeting LLMs：Towards retrieval—augmented large language models[C]// The 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Barcelona,Spain, 2024:6491-6501.

[23] 赵闪闪. 基于 OCR 技术的图像表格数据识别系统的设计与实现[D]. 南京：东南大学计算机科学与工程学院, 2020.

[24] Zhao Shanshan. Design and Implementation of Image from Data Recognition System Based on OCR Technology[D]. Nanjing： School of Computer Science and Engineering,Southeast University, 2020(in Chinese).

[25] Lin S C, Li M, Lin J. Aggretriever：A simple approach to aggregate textual representations for robust dense passage retrieval[J]. Transactions of the Association for Computational Linguistics, 2023, 11:436-452.

基金资助

天津市重点研发计划资助项目(23YFZCSN00080)