DisastRAG: A Multi-Source Disaster Information Integration and Access System Based on Retrieval-Augmented Large Language Models

DisastRAG introduces a multi-path architecture that goes beyond traditional single-path RAG. It supports document retrieval over a curated hazard corpus, structured access to relational disaster records, and external web fallback for out-of-corpus requests. The system also incorporates query understanding, strategy routing, response generation, and contextual memory—all unified into one framework. This design allows disaster managers to pull information from operational records, institutional documents, and live external sources simultaneously.

The team evaluated DisastRAG using four open-source LLMs across multiple retrieval configurations on both multiple-choice and open-ended disaster information tasks. Retrieval augmentation consistently outperformed no-retrieval baselines: multiple-choice accuracy improved by 12–23 percentage points, and open-ended keypoint coverage increased by up to 10.5 points. Larger candidate pools helped weaker models most, while stronger models were more sensitive to retrieval noise. Hybrid retrieval proved best for open-ended coverage, whereas vector retrieval with shallower reranking favored closed-form factual answers. Case studies further demonstrated that structured access and web fallback extend the framework beyond document-only RAG, enabling richer context for time-sensitive decisions.