A Study on the Feasibility and Limitations of Generating Design Search Formulas based on Large Language Models(LLMs)

In common data environment (CDE) systems, design information is retrieved using structured query formulas, which require knowledge of data structures and query syntax. This increases user burden and limits effective utilization of accumulated design data. Recent advances in large language models (LLMs) have shown potential for generating structured search formulas from natural language input; however, their applicability in real industrial design environments remains insufficiently examined. This study investigates the feasibility and limitations of natural language–based design query generation using LLMs under a closed design data environment with restricted external network access. A system was implemented to convert natural language queries into XML-based design search formulas using a locally deployed LLM, and experiments were conducted to analyze the structural characteristics of the generated queries. The experimental results show that the LLM can generate the basic structure of XML design queries with reasonable consistency, and that expanding the training data and iterative fine-tuning improve structural stability. However, limitations were identified in handling complex query conditions, suggesting the need for additional validation and refinement mechanisms. This study offers practical insights into the realistic use of LLMs for design data retrieval in CDE environments.