Answers are grounded in this paper and cite their sources.
River water quality monitoring generates large, heterogeneous datasets that can be difficult to translate into actionable insights. This study proposes a multi-agent Retrieval-Augmented Generation Large Language Model (RAG-LLM) decision-support framework that integrates specialised agents for query understanding, retrieval, data analysis, visualisation, and grounded response generation. The framework enables users to analyse environmental monitoring data through natural language (NL) interaction and supports evidence-based assessment of water quality trends, pollution hotspots, monitoring coverage, and station status. The framework was evaluated using UK Environment Agency water quality data from Greater Manchester, comprising 291 river monitoring stations and 45,342 observations recorded during 2024–2026, with data for 2026 covering only part of the year. Evaluation using 50 representative NL queries achieved 88.6% accuracy, 86.9% completeness, and 89.1% groundedness. The proposed framework also achieved a 10.9% hallucination rate compared with 17.0% for a standard text-only RAG baseline, representing a 6.1 percentage-point reduction in unsupported claims. These results demonstrate the framework’s ability to generate grounded and interpretable analytical responses. By integrating retrieval, structured environmental analysis, automated visualisation, and NL interaction, the framework provides an accessible and evidence-grounded approach to environmental monitoring and decision support.
No takes yet. Share an insight, caveat, or question.
Arslan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: