New RAG paradigm with iterative retrieval at multiple reasoning steps achieves 15-20% accuracy gain on HotpotQA

Positioned as go-to technique for dynamic, fact-heavy applications with frequently changing information

Research exposed a critical vulnerability where just 5 poisoned documents can corrupt RAG systems.

Proposed framework that dynamically decides when to retrieve external data for LLMs to reduce computational overhead

Clarification article published explaining distinction between RAG and fine-tuning for LLM applications

Publication of a framework moving RAG systems from proof-of-concept to production, outlining anti-patterns and a five-pillar architecture.

Ethan Mollick declared the end of the 'RAG era' as dominant paradigm for AI agents