TM-RAG: A Tree-Mapped Retrieval-Augmented generation framework for construction claim report generation

• A novel Tree-Mapped RAG framework for automated claim report generation. • Uses a timeline ontology and time tree to structurally model claim processes. • Hybrid retrieval combines semantic and structural signals for precise localization. • Achieves superior performance with a locally deployable, cost-efficient LLM. • Demonstrates strong adaptability across projects and formats with minimal data. To improve project administration and ensure efficient claim management, it is essential to accurately document all claim-related information. To improve claim recording efficiency, this study proposes an automatic claim report generation method: Tree-Mapped Retrieval-Augmented Generation (TM-RAG). TM-RAG maintains a timeline ontology that represents the workflow of claim events and establishes structured mappings between ontology nodes and corresponding entries in claim reports based on historical data. For new claim cases, TM-RAG can rapidly instantiate the ontology structure, construct a hierarchical time tree (a graph-based structure), and accurately extract information for claim report generation, leveraging mapped relationships. Experimental results show that TM-RAG with a locally deployed Qwen2.5-7B model achieves the best overall score (0.937), substantially improving EM over vector-based (+21–23 %) and graph-based methods (+15–17 %), while delivering on average + 33 % semantic quality gains, at only ∼ 20 % of GraphRAG’s runtime. Moreover, TM-RAG is adaptable to different claim report formats and cross-format knowledge transfer. Beyond claim management, TM-RAG offers a new paradigm for graph-augmented RAG methods to handle large-scale, structurally repetitive information extraction tasks, thereby improving graph construction and retrieval efficiency by leveraging historical data.