Iterate Until Retrieved: Factual Nugget Optimization for Discoverable Continual Corrections in Agentic RAG

ArXi:2605.25641v1 Announce Type: new Agentic retrieval-augmented generation (RAG) systems in complex B2B (business-to-business) settings may often receive free-form response feedback. Rather than generic feedback signals such as style, preference, or overall response quality, we focus on actionable factual corrections. We identify these instances and convert them into compact knowledge-base entries, which we call factual nuggets. We