Definition
Self-Grading RAG is an agentic architectural pattern where an LLM evaluates its own retrieved context for relevance and its generated output for faithfulness (hallucinations) before finalizing a response. It introduces a feedback loop that can trigger re-retrieval or alternative search strategies if the initial retrieval quality is deemed insufficient.
Distinguish from 'Manual Evaluation'; this is an automated, real-time quality gate within the production inference loop.
"A student double-checking their exam answers against the textbook and rewriting them if they find a discrepancy before handing the paper to the teacher."
Conceptual Overview
Self-Grading RAG is an agentic architectural pattern where an LLM evaluates its own retrieved context for relevance and its generated output for faithfulness (hallucinations) before finalizing a response. It introduces a feedback loop that can trigger re-retrieval or alternative search strategies if the initial retrieval quality is deemed insufficient.
Disambiguation
Distinguish from 'Manual Evaluation'; this is an automated, real-time quality gate within the production inference loop.
Visual Analog
A student double-checking their exam answers against the textbook and rewriting them if they find a discrepancy before handing the paper to the teacher.