RAG gives an LLM access to specific knowledge at query time by retrieving relevant documents and passing them as context — without retraining the model.