Search for a command to run...
Civic information is often dispersed across portals and PDFs, while users increasingly expect grounded answers with verifiable sources. Retrieval-Augmented Generation (RAG) can provide such grounding, but most interfaces operate as black boxes, revealing little about which passages influenced an answer or how sensitive that answer is to specific evidence. This lack of transparency limits trust, hinders debugging, and complicates audit requirements in real-world deployments. CityCopilot-X is a real-time glass-box framework for RAG that exposes evidence use across the entire pipeline. The system provides (1) visual attributions for query rewriting, retrieval, and reranking; (2) token-level coverage showing how generated text aligns with cited passages; and (3) an interactive counterfactual mode that recomputes answers with selected evidence removed, quantifying changes in confidence and coverage. CityCopilot-X supports multilingual civic queries and updates responses within seconds of source edits while maintaining stable latency and cost. In a pilot study with six analysts across 24 civic tasks, the panel reduced post-hoc correction time by 31% and increased citation coverage by 18 points compared with a black-box baseline. Synthetic case studies further highlight the system’s ability to reveal ranking failures, over-dominant documents, and evidence drift. CityCopilot-X demonstrates that RAG systems can be made transparent, auditable, and diagnosable without sacrificing responsiveness, offering a practical path toward trustworthy retrieval-grounded generation.