Corrective Retrieval-Augmented Generation (CRAG) System

An implementation of Corrective RAG that enhances traditional RAG systems with:

• Retrieval quality evaluation
• Dynamic knowledge augmentation
• Key fact extraction
• Fallback to web search when needed

Features

• Intelligent Retrieval Evaluation: Assesses relevance of retrieved documents
• Contextual Query Transformation: Rewrites queries for better retrieval
• Key Fact Extraction: Identifies and focuses on the most relevant information
• Web Search Fallback: Augments knowledge when local retrieval is insufficient
• Structured JSON Output: Consistent response format with citations
• Automatic Evaluation: Built-in LLM as a judge architecture to perform response quality assessment

Tech Stack:

• LlamaIndex for orchestration
• DuckDuckGo for deep web search
• Qdrant for vector database
• FastEmbed for embedding