Self-RAG enables models to decide when retrieval is needed versus generating from internal knowledge. Motivation: Not every query needs retrieval - simple questions answered from memory, complex/factual ones need grounding. Unconditional retrieval adds latency and may introduce noise. Mechanism: Model first predicts "retrieve" or "generate" token, if retrieve: execute RAG pipeline, if generate: answer directly from parameters, model self-evaluates answer quality. Training: Train model (or classifier) on examples of when retrieval helps vs hurts. Reward model for correct retrieve/no-retrieve decisions. Self-critique: Model generates answer, evaluates factuality, decides if retrieval needed to verify or improve. Implementation: Either fine-tune model with retrieval decisions, or use prompted self-evaluation. Benefits: Lower latency (skip retrieval when unnecessary), reduced cost, potentially higher quality (no irrelevant context). Challenges: Model must calibrate uncertainty, may skip retrieval when needed. Related: FLARE (Forward-Looking Active REtrieval), Adaptive RAG. Represents move toward smarter, more efficient retrieval decisions.