DRACO: A Large Language Model Initiative for Urban Water Systems

<p dir="ltr">We present DRACO (Drinking water and wAstewater COgnitive assistant), an initiative to support trustworthy Large Language Model (LLMs) applications for urban water management. Our first contribution is the release of a dataset of domain-specific texts, reasoning, and coding tasks for water networks and treatment systems. The dataset enables benchmarking and comparison of open- and closed-source models. Results from our 239-question evaluation show that state-of-the-art open-weight models can match—and in some cases outperform—proprietary systems. This finding is promising for utilities interested in on-premise deployment or in fine-tuning open-source models for secure, task-specific use. Moving forward, the DRACO initiative will focus on refining the dataset with more realistic, utility-driven case studies, and embedding this work into a broader agentic framework for developing trustworthy, LLM-enabled workflows for water utilities and authorities. This will be carried out through an open, collaborative process with the broader community. In parallel, we will also assess whether fine-tuning a dedicated model offers a meaningful advantage.</p><p dir="ltr">This paper was presented at the 21st Computing and Control in the Water Industry Conference (CCWI 2025) at the University of Sheffield (1st - 3rd September 2025).</p>