From Noisy Feedback to Trustworthy IssueSpecifications: An Agent-GovernedRetrieval-Augmented Generation Approach

Post-release user feedback is a major control signal for maintenance and evolutionin modern software development, yet it is noisy, fragmented, and difficult to trans-late into developer-ready issue specifications. Large Language Models (LLMs)can assist this transformation, but they often hallucinate or over-commit whenevidence is weak, conflicting, or incomplete, limiting their reliability in automatedsoftware engineering workflows. We propose AGR (Agent-Governed Retrieval-Augmented Generation), a framework that regulates evidence acquisition andgeneration decisions via agentic control. AGR first applies an agentic triage stepto filter low-signal or off-topic feedback, then retrieves evidence from a three-category hierarchy comprising official documentation, historical bug reports,and targeted web sources. It further performs confidence-weighted fusion acrossauthoritative categories and uses an agentic decision module to verify relevanceand sufficiency, trigger additional retrieval or online search when needed, reuseprior reports via memory, and abstain when trustworthy grounding cannot beestablished. We evaluate AGR on the Firefox ecosystem using community forumposts, the full official support documentation, and the most recent 50,000 Bugzillareports. Results show that AGR achieves high decision accuracy in triage andevidence verification, and produces significantly more actionable, well-organized,and engineering-useful issue specifications than both raw feedback and a strongLLM baseline, while reducing unsupported details.