Corrective RAG (CRAG) - Corrective Retrieval Augmented Generation

Basic Information

• Product Name: CRAG (Corrective Retrieval Augmented Generation)
• Proposer: Academic Research Team
• Paper: "Corrective Retrieval Augmented Generation" (January 2024)
• arXiv: https://arxiv.org/abs/2401.15884
• GitHub: https://github.com/HuskyInSalt/CRAG
• Type: RAG Improvement Method / Self-Corrective Retrieval Technique
• Subsequent Developments: A-RAG (February 2026), Higress-RAG (February 2026)

Product Description

Corrective RAG (CRAG) is a method designed to enhance the robustness of RAG generation, addressing the issue of traditional RAG's heavy reliance on the relevance of retrieved documents. CRAG introduces a lightweight retrieval evaluator to assess the overall quality of retrieved documents and triggers different knowledge retrieval strategies based on confidence levels. When retrieval results are suboptimal, CRAG activates alternative strategies such as web search and selectively focuses on key information through a decompose-recompose algorithm.

Core Features/Characteristics

• Retrieval Quality Assessment: Lightweight evaluator assesses the overall quality of retrieved documents
• Three-Level Confidence Judgment:
• Correct: Retrieved results are relevant and used directly
• Incorrect: Retrieved results are irrelevant, triggering alternative strategies like web search
• Ambiguous: Partially relevant, requiring further processing
• Web Search Extension: Automatically extends to large-scale web search when retrieval is suboptimal
• Decompose-Recompose Algorithm: Selectively focuses on key information in retrieved documents, filtering out irrelevant content
• Self-Corrective Capability: Detects and corrects retrieval errors, improving generation quality

CRAG vs Self-RAG

Technical Architecture

User Query
    ↓
Initial Retrieval (Vector Search)
    ↓
Retrieval Quality Evaluator
    ↓
┌───────────────┬───────────────┬───────────────┐
│   Correct     │   Ambiguous   │   Incorrect   │
│   Direct Use  │   Refine Filter│   Web Search  │
└───────┬───────┴───────┬───────┴───────┬───────┘
        └───────────────┼───────────────┘
                        ↓
               Decompose-Recompose: Extract Key Info
                        ↓
                   LLM Generate Answer

Subsequent Developments (2025-2026)

• A-RAG Framework (February 2026): Exposes keyword, semantic, and chunk-level retrieval tools directly to agents, improving QA accuracy by 5-13%
• Higress-RAG Framework (February 2026): Enterprise-level deployment solution based on corrective ideas
• Agentic RAG Evolution: CRAG ideas integrated into agentic RAG, where agents autonomously decide retrieval strategies

Business Model

• Academic Open Source: Research methodology, open-source implementation
• Framework Integration: Tutorials for CRAG implementation in frameworks like LangGraph/LangChain
• Educational Platforms like DataCamp: Provides CRAG implementation tutorials

Target Users

• Developers requiring highly robust RAG systems
• Enterprise applications with high demands for retrieval accuracy
• Quality improvement and optimization of RAG systems
• Self-corrective retrieval in AI agent systems

Competitive Advantages

• Significantly enhances the robustness of RAG systems
• Lightweight evaluator with minimal overhead
• Web search fallback ensures answer quality
• Compatible with existing RAG systems, allowing incremental integration
• Complementary to Self-RAG, can be used in combination

Relationship with OpenClaw Ecosystem

CRAG provides self-corrective capabilities to OpenClaw's RAG system. When OpenClaw agents retrieve insufficiently relevant information from user knowledge bases, CRAG's evaluator can automatically trigger alternative strategies (e.g., web search), ensuring that agents always provide high-quality responses. This self-corrective mechanism is crucial for enhancing the reliability and user trust of OpenClaw agents.