Key points are not available for this paper at this time.
The Southwestern US is a water-scarce region experiencing a megadrought more exceptional than any in the past 1200 years. It is also among the most rapidly growing, urbanizing, and diversifying areas in the country. To help people engage with the information they need to assist their communities in making decisions about water and drought preparedness, our interdisciplinary research team developed the Arizona Water Chatbot, an OpenAI-powered chatbot that uses retrieval-augmented generation to deliver information about Arizona's water situation to Arizona residents. The chatbot uses a distinctive architecture that implements guardrails to mitigate malicious content generation, ensuring the delivery of context-sensitive, relevant, accurate, and user-friendly responses. In this paper we discuss how a custom architecture provides fine-grained control of answers and increased ability to run multiple security checks that make a custom bot preferable to an OpenAI-hosted GPT for some requirements and situations. We also discuss how Waterbot is trained to incorporate Indigenous perspectives about water from the 22 American Indian tribal communities in Arizona to provide a more accurate and holistic view of the important historical, spiritual, and ecological role that water plays in the lives of Arizonans. Finally, we deliver insights on what other teams need to consider when building similar bots for public use.
Rajan et al. (Thu,) studied this question.