Search

A Dedicated Q&A System for University Libraries Based on Multi-agent Collaboration

Author(s)

Linjun Xiao1, Yuting Qing2

Affiliation(s)

1Hong Kong Polytechnic University, Data Science and Artificial Intelligence Institute, Hong Kong, China 2Beijing University of Technology, Beijing-Dublin International College, Beijing, China

Abstract

Aiming at the core issues existing in the implementation of large language models (LLMs) in the scenarios of university library business Q&A and librarian training, namely "defining the working boundary", "the problem of output hallucination", "computational resource limitations" and "adaptability to specific businesses", a dedicated Q&A system for university libraries based on multi-agent collaboration is proposed. This system first completes the security and compliance inspection of the data based on the agents, and then superimposes retrieval-augmented generation technology to answer questions. The experimental results show that the accuracy rate of the compliance inspection of the system reaches 94.4%. The fact consistency (93.6%), answer relevance (60%) and answer semantic similarity (79.6%) during answer generation far exceed those of ERNIE Bot. In a GPU 1-core virtual machine, the time consumption for completing the task is only 1-10 seconds, which proves that it can effectively empower the relevant businesses of university libraries.

Keywords

Large Language Model; Retrieval-Augmented Generation; Agent; University Library

References

[1] Zhang Yan, Hong Guangzong. Continuously Deepening the Construction of Service System to Promote High-quality Resources to Reach the Grassroots: The Strategic Focus of the Innovative Development of Public Libraries during the "15th Five-Year Plan" Period [J/OL]. Library Tribune, 1-13 [2025-02-26]. [2] Yang Feng, Zhu Li, Hu Lin. Constructing an Integrated Model of Artificial Intelligence Literacy Education: Insights from the Responses of Foreign University Libraries [J/OL]. Library Tribune, 1-10 [2025-02-26]. [3] Miao Yunwei. Generative AI Empowers the Classification and Protection of Reader Information in Libraries [J]. Library Tribune, 2024, 44(08): 34-43. [4] Xu Lingyan. Research on the Security Risks and Governance Paths of Large-scale Models [J]. Information Security Research, 2024, 10(10): 975-980. [5] Inan H, Upasani K, Chi J, et al. Llama guard: Llm-based input-output safeguard for human-ai conversations [J]. arXiv preprint arXiv:2312.06674, 2023. [6] Rebedea T, Dinu R, Sreedhar M, et al. Nemo guardrails: A toolkit for controllable and safe llm applications with programmable rails [J]. arXiv preprint arXiv:2310.10501, 2023. [7] Yuan Z, Xiong Z, Zeng Y, et al. Rigorllm: Resilient guardrails for large language models against undesired content [J]. arXiv preprint arXiv:2403.13031, 2024. [8] Rajpal S. Guardrails ai [EB/OL].(2023) [9] Dong Y, Mu R, Zhang Y, et al. Safeguarding large language models: A survey [J]. arXiv preprint arXiv:2406.02622, 2024. [10] Beurer-Kellner L, Fischer M, Vechev M. Prompting is programming: A query language for large language models [J]. Proceedings of the ACM on Programming Languages, 2023, 7(PLDI): 1946-1969. [11] Bellamy R K E, Dey K, Hind M, et al. AI Fairness 360: An extensible toolkit for detecting and mitigating algorithmic bias [J]. IBM Journal of Research and Development, 2019, 63(4/5): 4: 1-4: 15. [12] Bird S, Dudík M, Edgar R, et al. Fairlearn: A toolkit for assessing and improving fairness in AI [J]. Microsoft, Tech. Rep. MSR-TR-2020-32, 2020. [13] Hanu, Laura and {Unitary team}. Detoxify [EB/OL].[2024].https://github.com/unitaryai/detoxify. [14] Shen Si, Feng Shuyang, Wu Na, et al. Research on Retrieval-Augmented Generation of Policy Texts Integrating Large Language Models [J/OL]. Data Analysis and Knowledge Discovery, 1-18 [2025-02-26]. [15] Li Ming, Luo Xiaolan, Zhu Bangxian. Data Mining of Prescriptions in Traditional Chinese Medicine Ancient Books and Construction of Knowledge Question-Answering System [J/OL]. Library Tribune, 1-12 [2025-02-26]. [16] Peng Jia'er, Yu Feifei, Zhao Yuehua. From Patient-generated Health Data to Medication Recommendation Generation: A Patient Medication Information Question-Answering Model Based on Graph Retrieval-Augmented Generation [J/OL]. Modern Information, 1-22 [2025-02-26]. [17] Wang Heqing, Wei Jie, Jing Hongyu, et al. Meta-RAG: A Metadata-Driven Retrieval-Augmented Generation Framework in the Power Field [J/OL]. Computer Engineering, 1-11 [2025-02-26]. [18] Ollama. Ollama [EB/OL]. [2025].https://github.com/ollama/ollama. [19] Guo D, Yang D, Zhang H, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning [J]. arXiv preprint arXiv:2501.12948, 2025. [20] Ma X, Fang G, Wang X. Llm-pruner: On the structural pruning of large language models[J]. Advances in neural information processing systems, 2023, 36: 21702-21720. [21] LangGenius. dify[EB/OL]. [2025].https://github.com/langgenius/dify. [22] Microsoft. Autogen [EB/OL]. [2025].https://github.com/microsoft/autogen. [23] langchain-ai. Langchain [EB/OL]. [2025].https://github.com/langchain-ai/langchain.