RAG addresses three major pain points of large models: outdated training data, inability to access private data, and susceptibility to hallucination.