Instead of relying only on what a model memorized in training, RAG fetches fresh, specific data at query time (from a vector database, search index or API) and feeds it into the prompt. The model then answers from that retrieved context.
RAG is how assistants stay current and cite sources. MCP servers are increasingly the retrieval layer — exposing live data an agent pulls in on demand.
It reduces them by grounding answers in retrieved sources, but doesn't eliminate them — retrieval quality and how the model uses the context still matter.