Self-RAG — AI Glossary

Reference details

Related terms

• Corrective RAG (CRAG)
• Agentic RAG
• Retrieval Pipeline

How it works

1. The generator is fine-tuned to emit special reflection tokens alongside ordinary text, so the decisions are produced by the same forward pass rather than by a separate call.
2. A retrieval token decides whether the current span needs external evidence at all, which makes retrieval conditional rather than unconditional.
3. Relevance tokens judge each retrieved passage, so an unhelpful chunk can be discarded before it influences the answer.
4. Support tokens assess whether the generated claim is actually entailed by the passage, which is a check on grounding rather than on retrieval.
5. Because the tokens are part of the vocabulary, they can be used to score and rank candidate continuations at decode time.

Example: Why training it in costs less per request than prompting it

A prompted critique loop needs separate calls: one to decide whether to retrieve, one to judge relevance, one to check support, plus the answer — four calls where a fixed pipeline makes one. Reflection tokens fold those judgements into the generation pass, so the same decisions arrive at roughly the cost of a single call. The saving is per request and recurring; the cost is a fine-tune, paid once, and a model you now have to maintain.

Commonly confused with

A specific trained mechanism, not a prompting pattern: the model emits reflection tokens deciding whether retrieval is needed and critiquing what came back. That distinguishes it from agentic RAG, which describes the same behaviour achieved with any model and a scaffold, and from Corrective RAG, which grades documents and falls back to another source rather than reflecting on its own output.

When to use it

Reach for it when:

• Retrieval should be conditional and the per-request cost of a prompted critique loop is prohibitive.
• Grounding needs checking at claim level rather than at document level.
• A fine-tune is acceptable and the resulting model can be maintained.

Reach for something else when:

• No capacity to fine-tune or host a custom model — a prompted agentic loop achieves similar control.
• Using a frontier API model, where the vocabulary cannot be extended.
• Traffic is too low for a per-request saving to repay the training cost.

Primary source

• Asai et al., Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (arXiv)

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Self-RAG — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path