Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection

Type: paper Tier: 2 (Foundational Paper) URL: https://arxiv.org/abs/2310.11511 Accessed: 2026-06-09 Authors: Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, Hannaneh Hajishirzi (UW / AI2 / IBM) Year: 2023 arXiv: 2310.11511 — https://arxiv.org/abs/2310.11511

Core idea

Standard RAG always retrieves a fixed number of passages and stuffs them in — even when the query needs no retrieval, and regardless of whether the passages are any good. Self-RAG makes retrieval adaptive and self-critiqued using special reflection tokens that let the model:

1. Retrieve-or-not — decide whether external knowledge is needed before fetching anything

(skip retrieval on queries that don't need it).

1. Critique relevance — judge whether each retrieved passage is actually relevant.
2. Critique support — judge whether the generated answer is supported by the passages.

Why it beats always-retrieve RAG

Indiscriminately adding a fixed number of passages "diminishes LM versatility or can lead to unhelpful response generation." Selective, critiqued retrieval keeps factuality up and avoids the degradation of forcing irrelevant context into the answer. It also makes behavior controllable at inference (tune how aggressively to retrieve).

Result

7B/13B Self-RAG models outperformed ChatGPT and retrieval-augmented Llama2-chat on multiple benchmarks.

Why This Source Matters

It is the capstone idea of query engineering: the most advanced move is sometimes not to retrieve, and always to critique what you retrieved rather than trust it. It reframes the faithfulness metric (Project 04 / RAGAS) as an inline decision, not just a post-hoc score — "is this supported?" asked during generation. In this elective it is the extension: a retrieve-or-not gate plus a relevance/support check over the transformed pipeline.

Key Claims

• Standard RAG always retrieves a fixed number of passages and stuffs them in; Self-RAG makes retrieval adaptive and self-critiqued via special reflection tokens.
• Three decisions: retrieve-or-not (skip retrieval when unneeded), critique each passage's relevance, and critique whether the answer is supported by the passages.
• Selective, critiqued retrieval keeps factuality up and avoids degradation from forcing irrelevant context; 7B/13B Self-RAG models beat ChatGPT and retrieval-augmented Llama2-chat on several benchmarks.

Relevant To

• Elective 05 — Advanced RAG / Query Engineering (retrieve-or-not gate + self-critique, M5/extension).
• Related: Project 04 / RAGAS faithfulness, reframed as an inline "is this supported?" decision; Elective 03 (the critique is an extra LLM cost).

Known issues / cautions

• Full Self-RAG trains reflection tokens; the elective approximates the behavior (a retrieve-or-

not gate + a relevance/support critique) with prompting, not training.

• The critique is itself an LLM judgment — fallible, and a cost (Elective 03).
• "Don't retrieve" is a real option, but a wrong skip loses grounding — measure on the eval.