A CRAG pipeline adds a retrieval-quality gate to RAG. It retrieves internal documents, evaluates their confidence, and branches toward internal context, hybrid context, or corrective external search before generation. section 11.2.4
The pipeline separates evaluation from orchestration:
- Score retrieved documents with a lightweight evaluator and classify them as correct, ambiguous, or incorrect. section 11.2.4
- Use high-confidence internal evidence directly, supplement ambiguous results with web search, and replace incorrect retrieval with external results. section 11.2.4
- Tune high and low thresholds on held-out data rather than treating defaults as universal; the gap defines an ambiguity zone. section 11.2.4
Outdated compliance retrieval
For the query about onboarding an EU PEP through expedited KYC, an outdated internal policy scores 0.76 while the 2024 directive scores 0.63. The ambiguous result triggers hybrid handling instead of allowing internal retrieval alone to produce the answer. section 11.2.4
