Navigating State Drift in Retrieval-Augmented Generation (RAG) Agents: A Diagnostic Benchmark and Structured Graph-Guided Repair Framework

Abstract

Multi-turn Retrieval-Augmented Generation (RAG) agents are increasingly relied upon to integrate conversational history, external knowledge retrieval, tool use, and persistent memory. However, their failures extend beyond isolated hallucinations to a more systemic and gradual degradation we term state drift, a phenomenon where the agent progressively loses alignment with the user's current goal, active constraints, retrieved evidence, tool observations, or updated memory. This paper offers three contributions. First, we formalize state drift and introduce a comprehensive taxonomy covering goal, constraint, evidence, retrieval, tool-state, memory, and abstention drift. Second, we present DRIFTBENCH, a diagnostic benchmark featuring controlled state perturbations and fine-grained state-level annotations to systematically elicit and measure drift in RAG agents. Third, we propose SAGE-R, a model-agnostic state-graph framework that diagnoses drift in real time and triggers typed repair actions including rollback, re-retrieval, tool cross-validation, clarification, and abstention. The current repository provides a deterministic v0 scaffold with generated examples, baseline proxies, repair traces, analysis tables, and a human-validation packet, establishing a validated research pipeline prior to the integration of external datasets, model-backed experiments, and completed human annotation. All quantitative results reported in this paper are therefore scaffold-level pipeline validations rather than model-backed empirical findings. Our work lays the foundation for building more reliable, state-aware RAG agents that resist drift across complex multi-turn interactions.