Graph-driven embedding reinforcement and traceable LLM agent for reliable element alignment in construction report generation

Engineering report generation from construction-site Internet of Things (IoT) data using large language models (LLMs) remains challenging due to hallucinations. Ensuring traceability and reliability in information retrieval and multi-step reasoning is essential within retrieval-augmented generation (RAG) for LLM. This paper formalizes the RAG-LLM pipeline and proposes a dual-stream enhancement combining knowledge graph (KG) construction with reinforcement learning (RL)-based retriever tuning. The graph-guided module extracts structured engineering elements, while RL improves semantic alignment and tokenization of critical terms. Leveraging this dual-stream RAG, a traceable reporting agent is developed, providing end-to-end traceability of retrieval and reasoning, along with inter-step similarity measures. When collaborating with existing on-site IoT systems, the agent can extend automated monitoring to decision-making support. This paper presents a reliable approach for construction report generation and advances human-AI collaboration in construction management.