Why it matters

Self-RAG adds control. Understanding shapes design.

Advertisement

The architecture

LLM emits reflection tokens: [Retrieve?] + [Relevant] + [IsSupport] + [IsUse].

Guides retrieval + generation.

Self-RAG token flowLLM reflection tokensmeta signalsRetrieve when neededon demandFilter + usequality signalsAsai et al. 2023; trained model emits reflection tokens; specialized fine-tuning
Self-RAG.
Advertisement

How it works end to end

Fine-tune model to emit tokens.

Retrieval on demand.

Self-critique of retrieved docs.

Confidence gating.