RAG converts user questions into vectors during the retrieval phase to recall document fragments, then injects the recalled content into the prompt during the generation phase, thereby reducing model hallucination.