RAG
Also written: retrieval-augmented generation, Retrieval-augmented generation
Retrieval-augmented generation: for each question, find the passages most likely to hold the answer, put them in the prompt as named sources, and have the model answer only from them. The model reads instead of remembering.