LLM-based processing of design inspection reports for measuring building design quality

Design inspection reports contain valuable knowledge about building design quality issues, yet they remain underutilized due to their unstructured format, lack of standardized classification systems, and the need for labour-intensive manual interpretation. This paper proposes a methodology for converting unstructured textual design quality issues into a structured and standardized format. First, an information model is developed, defining the target information entities, their corresponding taxonomies, and semantic relationships. A procedure is then introduced to automate Information Extraction (IE) and Information Classification (IC) using Large Language Models (LLMs), Few-Shot Learning (FSL), and Retrieval-Augmented Generation (RAG). The methodology was evaluated on a dataset collected from real-world design inspection reports, assessing IE and IC accuracy, computational cost, and processing time. Results demonstrate the feasibility of the proposed approach to structure information about design quality issues, supporting a data-driven design quality measure, defect pattern analysis, and proactive defect prevention in early design stages.