RAG-Tales

Learning LangChain, Agentic AI & RAGs because me want full control over these clankers.

LangChain

It covers different topics in LangChain such as Introduction, Prompts, Runnables, Models, Document Processing, Retrievers, Chains, Output Parsers and Callbacks.

• Notes
• Code

LangGraph

It covers different topics in LangGraph such as State Management, Nodes & Edges, Conditional Routing, Checkpointing, Human-in-the-Loop, Multi-Agent Systems, Agent Architectures, Subgraphs, Streaming and Error Handling.

• Notes
• Code

CRAG

CRAG (Corrective Retrieval Augmented Generation) is an advanced RAG technique that improves the reliability and accuracy of retrieval-augmented language models by adding a correction mechanism to evaluate and refine retrieved documents before generation.

Traditional RAG systems retrieve documents and use them directly for generation, which can lead to:

• Irrelevant retrieved documents being used in generation
• Inaccurate or outdated information affecting output quality
• No mechanism to verify retrieval quality
• Hallucinations when retrieved context is poor

How CRAG Works

CRAG introduces a retrieval evaluator that assesses the quality of retrieved documents and triggers corrective actions when needed. Architecture Components

Query → Retrieval → Evaluation → Correction (if needed) → Generation

• Retrieval: Standard document retrieval from knowledge base
• Evaluation: Lightweight evaluator model assesses retrieval quality
• Correction: Triggered based on evaluation confidence
• Generation: Final response using corrected/verified context

Based on the retrieval evaluator's confidence assessment:

1. Correct (High Confidence)

• Retrieved documents are relevant and accurate
• Proceed directly to generation
• Use retrieved knowledge as-is

2. Incorrect (Low Confidence)

• Retrieved documents are irrelevant or poor quality
• Discard retrieved documents
• Use web search or alternative knowledge sources
• Prevents hallucination from bad retrieval

3. Ambiguous (Medium Confidence)

• Retrieved documents are partially relevant
• Apply knowledge refinement
• Decompose documents into knowledge strips
• Filter and rerank at granular level
• Use only most relevant segments

# Pseudocode
def crag_pipeline(query):
    # 1. Retrieve documents
    documents = retriever.retrieve(query)
    
    # 2. Evaluate retrieval quality
    confidence = evaluator.evaluate(query, documents)
    
    # 3. Corrective action
    if confidence == "CORRECT":
        context = documents
    elif confidence == "INCORRECT":
        context = web_search(query)
    else:  # AMBIGUOUS
        strips = decompose_to_strips(documents)
        relevant_strips = filter_and_rerank(query, strips)
        context = relevant_strips
    
    # 4. Generate response
    response = generator.generate(query, context)
    return response

---