Hydropower and water conservancy engineering construction (HWCEC) involves high-risk operations and complex safety standards, which challenge traditional general-purpose large language models (LLMs) in reasoning with specialized terminology. To address this, this study proposes a construction safety knowledge reasoning-capable large language model (CSKR-LLM-HWCEC), integrating retrieval-augmented generation (RAG) and knowledge graphs (KG) with LLMs. The CSKR-LLM-HWCEC enhances domain-specific reasoning by combining RAG for knowledge retrieval with LLMs’ generative power and incorporating a safety-specific KG to improve logical reasoning and interpretability. Key contributions include: (1) advancing the domain knowledge in HWCEC through RAG-enhanced LLMs; (2) introducing a specialized safety KG to boost reasoning capabilities and answer transparency; and (3) offering a flexible and scalable framework for future improvements in safety management. In both expert evaluations and automatic metrics, CSKR-LLM-HWCEC demonstrated an overall performance improvement of approximately 20%–30% over ChatGPT-4.0 and Qwen, while achieving more than a 50% relative advantage on Rouge and Bleu scores. This study not only advances innovation in the knowledge system of construction safety but also provides intelligent solutions for future safety management and engineering decision-making.
Chen et al. (Wed,) studied this question.