Definition
Self-RAG is an architectural framework that trains a Large Language Model to output specialized reflection tokens, allowing it to autonomously critique its own retrieval relevance and generation quality; it trades off increased inference latency for significantly higher factual precision and reduced hallucinations.
Unlike standard RAG, Self-RAG uses internal model 'critique' tokens to filter documents and self-correct during generation rather than relying on external validation scripts.
"An author who meticulously checks every citation against their draft, crossing out and rewriting paragraphs if the evidence doesn't match the claim."
Conceptual Overview
Self-RAG is an architectural framework that trains a Large Language Model to output specialized reflection tokens, allowing it to autonomously critique its own retrieval relevance and generation quality; it trades off increased inference latency for significantly higher factual precision and reduced hallucinations.
Disambiguation
Unlike standard RAG, Self-RAG uses internal model 'critique' tokens to filter documents and self-correct during generation rather than relying on external validation scripts.
Visual Analog
An author who meticulously checks every citation against their draft, crossing out and rewriting paragraphs if the evidence doesn't match the claim.