Performance comparison of retrieval-augmented generation and fine-tuned large language models for construction safety management knowledge retrieval

Construction safety standards are in unstructured formats like text and images, complicating their effective use in daily tasks. This paper compares the performance of Retrieval-Augmented Generation (RAG) and fine-tuned Large Language Model (LLM) for the construction safety knowledge retrieval. The RAG model was created by integrating GPT-4 with a knowledge graph derived from construction safety guidelines, while the fine-tuned LLM was fine-tuned using a question-answering dataset derived from the same guidelines. These models' performance is tested through case studies, using accident synopses as a query to generate preventive measurements. The responses were assessed using metrics, including cosine similarity, Euclidean distance, BLEU, and ROUGE scores. It was found that both models outperformed GPT-4, with the RAG model improving by 21.5 % and the fine-tuned LLM by 26 %. The findings highlight the relative strengths and weaknesses of the RAG and fine-tuned LLM approaches in terms of applicability and reliability for safety management.