A Knowledge-based RAG System for Intelligent Fire Investigation Reporting

作者：Jianyuan Tao, Ping Wang, Hanfeng Jiang, Yu Han · 发表于：AIACT · 年份：2026 · DOI：10.1145/3795496.3795712 · 研究领域：Computer Science

Fire investigation reporting requires transforming fragmented and often incomplete incident descriptions into structured multi-field documentation, but large language models frequently produce hallucinations, inconsistent reasoning, and weak evidence grounding in this safety-critical domain. This study presents a knowledge-based retrieval-augmented generation system tailored for standardized fire investigation reports. A fire-investigation knowledge graph, constructed from 500 official cases and aligned with a 25-label schema, provides explicit domain constraints that guide retrieval and reduce speculative generation. The system integrates lexical, semantic, and graph-assisted retrieval into a hybrid pipeline and employs multi-stage structured prompting to generate each report field sequentially while maintaining cross-section coherence. Evaluations on three representative 8b-scale models DeepSeek-r1-8b, Qwen3-8b, and Llama3.1-8b show substantial gains in Answer Relevancy, Faithfulness, and expert ratings compared with direct model generation. Results demonstrate that injecting structured domain knowledge significantly improves factual accuracy, internal consistency, and practical usability, offering a viable path toward reliable AI-assisted fire investigation documentation.