crag

**CRAG** is the **Corrective Retrieval-Augmented Generation framework that evaluates retrieval quality and applies corrective actions when evidence is weak** - it aims to prevent low-quality retrieval from propagating into poor final answers. **What Is CRAG?** - **Definition**: RAG architecture with explicit retrieval quality assessment and correction paths. - **Correction Actions**: Can trigger web fallback, query refinement, filtering, or answer abstention. - **Quality Estimation**: Uses confidence signals to judge whether retrieved evidence is sufficient. - **Pipeline Goal**: Improve robustness when initial retriever results are incomplete or noisy. **Why CRAG Matters** - **Failure Containment**: Stops weak retrieval sets from driving confident but wrong answers. - **Robustness**: Adds resilience against domain drift and sparse-corpus edge cases. - **Safety Benefit**: Supports abstain-or-retry behavior when evidence quality is low. - **Answer Reliability**: Corrective loops increase chance of evidence-backed final outputs. - **Operational Visibility**: Quality scores provide diagnostics for retriever health monitoring. **How It Is Used in Practice** - **Quality Classifier**: Score retrieval bundles before generation proceeds. - **Correction Policy**: Route low-confidence cases into refinement or fallback pipelines. - **Outcome Logging**: Track correction triggers and downstream answer accuracy for tuning. CRAG is **a robust control pattern for handling retrieval uncertainty** - CRAG improves reliability by adding explicit quality checks and corrective actions.