Self-RAG

Self-RAG is a pattern in which the language model emits special reflection tokens that control its own retrieval and generation decisions. At inference time, the model decides whether a retrieval call is needed for the current step, whether each retrieved passage is relevant and supportive, and whether the draft generation is faithful to the retrieved evidence. These reflection tokens are introduced during fine-tuning on a dataset labeled with retrieval and critique signals, so the behavior is learned rather than prompted. The result is a model that can skip retrieval on easy prompts, pull in multiple passages only when needed, and self-flag when its own output is under-supported. It differs from Corrective RAG in that the corrective logic is inside the model weights rather than implemented as an external control loop over a generic LLM.