An Augmented Structured Evidence Retrieval Framework for Generating Chemical Accident Emergency-Response Plans Using a Large Language Model

Abstract Retrieval-augmented generation offers a promising route for automatically generating chemical accident emergency-response plans, but its practical use remains limited by fragmented knowledge sources, weak evidence associations, insufficient regulatory constraints, safety prohibitions, and low credibility. To improve the reliability of large language models in this high-risk task, this study proposes HKRE-ARG, a hierarchical key-path retrieval and enhanced-evaluation framework for emergency-response plan generation. The framework integrates heterogeneous knowledge, including hazardous chemical properties, accident scenarios, response measures, and legal constraints, into a unified structured knowledge representation. A key-path retriever is designed to combine relationship-level candidate compression, structural distance encoding, and path-level re-ranking, thereby constructing highly relevant and traceable evidence chains for plan generation. The retrieved evidence subgraphs are then provided to a downstream large language model as external knowledge. In addition, an LLM-as-a-judge-driven multidimensional evaluation and feedback correction mechanism is introduced, where verification results guide query refinement and targeted re-retrieval of key evidence, forming a closed-loop process from evidence organization to result validation. Experiments with six mainstream large language models show that, compared with BM25, VectorSim, RoG, and GNN-RAG, HKRE-ARG achieves superior evidence retrieval quality, generation credibility, and robustness under uncertain inputs, with an average total performance score of 8.94. These results indicate that HKRE-ARG can reduce ungrounded generation, regulatory mismatches, and unverifiable outputs, providing methodological support for credible emergency-response planning in high-risk decision-making scenarios.