AionRAG: Time-Correct Retrieval-Augmented Generation under Knowledge Drift

Retrieval-augmented generation (RAG) can fail in dynamic corpora due to time incorrectness: semantically relevant retrieval mixes multiple historical versions of the same claim, and the model often fails to resolve contradictions in favor of the version that is valid at the query time. We introduce AionRAG, a time-correct RAG system that treats retrieval as a calibrated control problem. Given a query, AionRAG predicts whether retrieval is needed and, when it is, selects a query-specific evidence window and hop depth; it then filters candidates by time before semantic ranking to prevent version mixing, and applies a lightweight conflict-gated decoding rule when retrieved evidence disagrees with the model prior. Crucially, AionRAG calibrates decision confidence (ECE=1.7%) so a single threshold maps to predictable latency–quality trade-offs across domains. Across seven benchmarks (242,900 queries) spanning controlled drift tests and real-world evolving corpora (Wikipedia revision histories, U.S. Federal Register policies, and licensed financial news), AionRAG improves temporal consistency and faithfulness while reducing retrieval calls by 29%. On WikiRevision-Real, AionRAG improves temporal consistency by +7.2 points [95% CI: 6.1, 8.3] and faithfulness by +6.4 points [5.3, 7.5] over RouterEns+NLI (our strongest deployable baseline). Long-context baselines (128k tokens) remain vulnerable to conflict amplification, trailing AionRAG by 12.5 points on high-conflict queries despite 2.9x higher latency. These results position time-correct retrieval control as a first-class requirement for reliable RAG under knowledge drift.