User-Anchored Retrieval-Augmented Generation (UA-RAG)

This paper proposes User-Anchored Retrieval-Augmented Generation (UA-RAG), a retrieval architecture designed to reduce ambiguity, retrieval failure, correction overhead, and unnecessary inference branching in large language model systems. The core argument of this work is that many LLM failures originate not from insufficient reasoning capability, but from unstable target determination. Current AI systems frequently attempt to infer user intent using incomplete textual prompts alone, often resulting in retrieval mismatch, outdated responses, repeated correction loops, and excessive computational overhead. UA-RAG introduces a structural separation between target determination and truth determination. Instead of forcing the model to independently guess what the user is referring to, the user directly anchors the retrieval context through integrated search exploration, rendered-screen sharing, browsing interaction, and target confirmation. In this framework, the selected page or rendered context is not treated as truth itself, but as a semantic anchor indicating what the user is actually referring to. The AI system then performs verification, contradiction analysis, cross-source reasoning, and freshness validation on top of that anchor. The paper additionally critiques rigid source-type hierarchy assumptions in retrieval systems and argues for evidence-validation-centered reasoning rather than excessive pre-filtering based on source category alone. UA-RAG may represent a potential second-stage evolution of LLM systems: moving beyond pure model scaling toward human-context-aligned retrieval, ambiguity collapse architectures, and computationally efficient target stabilization.