UA-RAG: A User-Anchored Framework for Retrieval-Augmented Generation

This record contains the official manuscript of UA-RAG (User-Anchored Retrieval-Augmented Generation). UA-RAG proposes an architectural extension to Retrieval-Augmented Generation (RAG) by introducing an explicit user-anchoring stage before retrieval. Rather than requiring the model to infer the user's intended referent solely from natural-language input, UA-RAG allows users to establish the retrieval target explicitly, separating target determination from truth determination. The paper introduces Anchor Theory, a conflict taxonomy (Meaning, Identity, and Content Conflict), Constraint-First Retrieval, Partial Branch Surgery, Rendered-Screen Retrieval, and a falsifiable evaluation methodology. It argues that stabilizing the retrieval target before retrieval can reduce ambiguity propagation and downstream correction costs while remaining fully compatible with conventional RAG workflows and other complementary retrieval techniques. This record contains the original English manuscript together with the official Korean translation prepared by the author.