关键词: 《活计档》档案, 知识图谱, 清代宫廷造物, 大语言模型, 提示词工程

Abstract: By constructing a knowledge graph of the Qing Yongzheng Imperial Workshop archives, this study systematically organizes the artefact-related information recorded in the Huoji Dang, thereby overcoming the fragmented limitations of traditional research and providing new perspectives and methodological support for the study of imperial artefacts in the Qing dynasty. The research employs large language models (LLMs) and prompt engineering techniques. After iteratively optimizing prompts and incorporating manual verification, the extracted data are stored in a Neo4j graph database, enabling the automatic transformation of unstructured data in the Huoji Dang into a structured knowledge graph. Results demonstrate that DeepSeek-V3 outperforms other LLMs across all evaluation metrics, with clear overall advantages. In sum, this method effectively and accurately extracts entities, attributes, and relationships, and the resulting knowledge graph clearly illustrates the organizational structure, production processes, and imperial aesthetic preferences embedded in the archives. Moreover, the proposed “LLM + prompt engineering” approach demonstrates strong transferability in addressing the automatic extraction of premodern texts, offering a valuable reference for similar studies. It realizes the structuring and visualization of textual information, thereby providing a systematic knowledge-sharing platform for research on Qing imperial artefacts.

Key words: The Huoji Dang (Workshops' archives), Knowledge graph, Qing dynasty palace artifacts, Large language model, Prompt engineering