Self-RAG and CRAG: When to Retrieve

Naive always retrieves; Self-RAG gates retrieval; CRAG grades and corrects

Self-RAG and CRAG: When to Retrieve Naive always retrieves; Self-RAG gates retrieval; CRAG grades and corrects 01 / Decide 02 / Retrieve + grade 03 / Generate EX / CRAG correct Query · user question · Decide Query user question Retrieve? · Self-RAG reflects · Decide · gate Retrieve? Self-RAG reflects gate Retrieve · vector store · Retrieve + grade Retrieve vector store Grade Docs · CRAG relevance · Retrieve + grade · gate Grade Docs CRAG relevance gate Generate · grounded answer · Generate Generate grounded answer Answer · Generate Answer Corrective · web fallback · CRAG correct Corrective web fallback Legend User UI Agent logic Policy Tool action Context / trace

Naive RAG

  • • Always retrieves, one shot, no gates
  • • Wastes calls on questions the model already knows
  • • Trusts whatever comes back, relevant or not

Self-RAG

  • • A reflection gate decides whether to retrieve at all
  • • Well-known questions answer straight from memory
  • • Retrieval happens only when it adds value

CRAG

  • • Grades retrieved docs for relevance before use
  • • Insufficient results trigger a corrective search
  • • Only re-grounded evidence reaches generation