WaterRAG: A Multiagent Retrieval-Augmented Generation Framework to Support Water Industry Transitions to Net-Zero.

Area of Science:

• Environmental Engineering
• Artificial Intelligence
• Information Science

Background:

• Achieving net-zero emissions in wastewater treatment necessitates integrated information across disciplines.
• Large language models (LLMs) show promise but have limitations in specialized domains like wastewater treatment.
• Existing LLMs struggle to access and synthesize the vast, specific knowledge required for advanced wastewater management.

Purpose of the Study:

• To introduce WaterRAG, a multiagent retrieval-augmented generation (RAG) framework.
• To couple LLM reasoning with verifiable wastewater treatment knowledge.
• To support evidence-based decision-making for sustainable wastewater management.

Main Methods:

• Developed a multiagent RAG framework (WaterRAG) integrating LLM reasoning with a curated database of 7637 peer-reviewed studies and 11 engineering references.
• Implemented iterative collaboration among retrieval, review, and evaluation agents.
• Benchmarked WaterRAG's performance on 370 technical questions and literature review tasks.

Main Results:

• WaterRAG achieved an 80.5% answer correctness rate on professional wastewater treatment questions, surpassing standalone GPT-4.1 (64.9%).
• Generated more comprehensive and citation-supported literature reviews, with quality improving through refinement.
• Ablation experiments confirmed the synergistic benefits of optimized retrieval and the multiagent framework.

Conclusions:

• WaterRAG demonstrates the potential of retrieval-grounded LLM systems in specialized domains.
• The multiagent RAG approach effectively complements professional expertise in wastewater treatment.
• This framework supports evidence-based decision-making for sustainable wastewater management and net-zero emissions goals.