We introduce iRAG, a decision‑support system that integrates Retrieval‑Augmented Generation (RAG) with a hybrid decision‑making framework for human‑in‑the‑loop scenarios. Conventional RAG relies on a pre‑populated knowledge base (KB); iRAG assumes an initially empty KB that is gradually filled by human decisions. Each decision is stored as a retrievable precedent, creating an ever‑growing case library.
The iRAG workflow consists of three stages:
- Silent observation – the system only records inputs without producing output;
- Evidence‑grounded contestation – using retrieved precedents, the system generates explanatory answers for human review and correction;
- Autonomous handling with optional abstention – when sufficient evidence is available, the system answers on its own; otherwise it may abstain.
Compared with an independent baseline, iRAG matches the baseline’s final‑decision error rate when knowledge drift is under 10%, and reduces error by up to 40% when drift reaches 40% across all configurations. Introducing retrieval decay improves performance on rapidly changing knowledge but increases error on stable knowledge, revealing a stability‑plasticity trade‑off rather than a universal benefit.
Human oversight is built into the architecture, yet the system’s usability with real users remains to be validated.
Review