LLM-based Processing of Design Inspection Reports as a Measure of Building Design Quality

Abstract: Extracting shareable knowledge concerning design-related issues from construction project documentation enables the assessment of design quality and, more broadly, improvement across the industry. This study proposes a solution to automate information extraction from inspection reports in construction industry processes, with a specific focus on those produced by design reviews often conducted before the tendering phase. The approach is based on LLMs, prompt engineering, and few-shot learning. We evaluated three LLMs (GPT-4o, Mistral, and Llama 3) across four-shot scenarios, assessing their performance, computational cost, and time. Results show that GPT-4o achieved the highest performance while ranking second in computational cost and time.