A Self-Hosted Multi-Agent RAG System for Sustainable Waste and Water Management

This work presents a self-hosted multi-agent Retrieval-Augmented Generation (RAG) system designed to enhance sustainable waste and water management infrastructures. Modern smart waste bins and aquatic monitoring devices generate large volumes of data, but facility managers and cleaners are often overwhelmed by fragmented dashboards and raw logs. To address this bottleneck, the proposed EcoBin-RAG system enables natural language queries, allowing users to ask questions such as “Which bins are nearing overflow?” or “Where were water contamination alerts this week?” and receive immediate, grounded responses. The architecture integrates open-source tools including Marker for document ingestion, Ollama for local LLM hosting, LangGraph for multi-agent orchestration, and Weaviate for hybrid retrieval. By keeping all components fully local, the system preserves privacy, improves modularity, and adapts flexibly across heterogeneous domains such as land-based waste bins and low-connectivity water monitoring units. Experimental evaluation demonstrates high retrieval accuracy (Recall@3 = 90%) and improved faithfulness in generated responses, while highlighting factual correctness as the primary area for future improvement. This study demonstrates how conversational AI can bridge the gap between IoT complexity and actionable insights in sustainability operations. Beyond waste and water management, the approach illustrates potential extensions to smart city deployments and biomedical imaging pipelines.