Water information systems (ISs) are essential for sustainable water management but are often constrained by fragmented, heterogeneous, and rapidly evolving data. Advances in generative artificial intelligence (GenAI), particularly large language models (LLMs), have enabled natural language (NL) question-answering (QA) over curated corpora, allowing analysts to interrogate heterogeneous datasets directly. Yet general-purpose LLMs, trained on static public datasets, remain insufficient for operational decision-making. Retrieval-augmented generation (RAG) addresses this gap by grounding outputs in up-to-date, organisation-specific, and authoritative sources, enabling customised QA across private datasets. Together, LLMs and RAG can deliver grounded, auditable, and context-aware responses across diverse modalities such as text, tables, time series, remote-sensing imagery, geographic information system (GIS) layers, and knowledge graphs (KGs). While early applications of RAG in the water sector exist, adoption remains limited, particularly for complex multi-modal and geo-temporal queries where provenance, uncertainty reporting, and principled abstention are critical. To advance sustainable water ISs, this study proposes a three-step framework comprising (i) identification of high-value water-sector decision and governance applications, (ii) integration of authoritative, multi-source datasets under provenance, privacy, and operational constraints, and (iii) NL interaction enabled through LLM–RAG architectures. Unlike general-purpose RAG pipelines, the framework explicitly links retrieval and generation to domain-specific outcomes, multimodal water data, and sustainability considerations related to cost and carbon. A review of recent literature indicates that, despite growing interest in LLMs for water management, only a small subset of studies (fewer than one-quarter of those reviewed) implement RAG-based approaches, and even fewer address multimodal or geo-temporal reasoning. While current applications predominantly rely on text and time-series data, the framework highlights the importance of integrating additional modalities such as remote sensing, GIS layers, and KGs to support more robust and actionable water decision-making. • LLMs and RAG enable smarter, source-grounded water insights. • Water ISs face fragmented, evolving, and siloed data challenges. • Framework: Identify apps, integrate data, enable NL interaction. • Early RAG adoption in water remains limited but promising. • Key challenges: Data readiness, privacy, evaluation, and sustainability.
Muhammad Arslan (Sun,) studied this question.