Developing a question-answering system for crane-lifting regulatory compliance checking through retrieval-augmented generation and fine-tuning an integrated large language model

Purpose This study aims to develop an innovative question-answering system for efficient and effective crane-lifting regulatory compliance checking by integrating retrieval-augmented generation (RAG) and fine-tuning into a domain-specific large language model. Design/methodology/approach A technical framework was developed comprising automated dataset generation, supervised fine-tuning of Llama 3 and advanced RAG. The system, in the form of a chatbot (HeLPer), was benchmarked against base, fine-tuned and RAG-enhanced Llama 3 and GPT-5.2 and also evaluated by conducting an RAG ablation study. Findings HeLPer achieved the highest accuracy and relevance among the models, with accuracy improved by 60% over base Llama 3. The RAG ablation study showed that reranking improved retrieval performance, whereas benefits from query expansion depended on retrieval strategies and query categories. Consistency and response latency remain areas for improvement. Research limitations/implications This paper demonstrates the potential of integrating model adaptation and external knowledge retrieval for developing domain-specific LLM systems and provides an evaluation basis for assessing both system-level performance and individual RAG components in construction applications. Practical implications HeLPer enables construction practitioners to retrieve regulatory information from codes of practice, guidebooks and project-specific documents for conducting lifting planning, safety training and routine inspections. Its updatable knowledge base and automated workflow support efficient implementation in constantly changing regulations and project requirements. Originality/value The paper contributes to an RAG and fine-tuning integrated framework for developing domain-specific QA systems and provides a chatbot that supports human–AI collaboration in crane-lifting regulatory compliance checking.