rag_advanced

Advanced RAG Example

This example demonstrates an advanced RAG pipeline with document chunking, reranking, and citation support.

Overview

The advanced RAG pattern includes:

1. Document Chunking: Split large documents into smaller, manageable chunks
2. Retrieve: Find relevant chunks from a vector store
3. Rerank: Re-score retrieved documents for better relevance
4. Generate: Generate answers with the LLM
5. Format Citations: Add source citations for transparency

This implementation is suitable for:

• Production RAG systems
• Applications requiring high accuracy
• Systems needing source attribution
• Large document collections

Features

• Text Splitting: Automatic document chunking with overlap
• Reranking: Improve retrieval quality with relevance scoring
• Citations: Automatic citation generation
• Higher Quality: Better embeddings and more sophisticated pipeline
• Metadata Tracking: Preserve document metadata through the pipeline

Running the Example

cd examples/rag_advanced
go run main.go

Key Components

• TextSplitter: Splits documents into chunks with configurable size and overlap
• Reranker: Re-scores documents based on query-document relevance
• Vector Store: Stores document chunks with embeddings
• Retriever: Retrieves top-k most relevant chunks
• LLM: DeepSeek-v3 for answer generation
• Pipeline: Advanced RAG (Retrieve → Rerank → Generate → Format Citations)

Pipeline Flow

Query → Retrieve Chunks → Rerank by Relevance → Generate Answer → Add Citations → Result

Example Output

The example demonstrates:

• Document chunking statistics
• Retrieved and reranked documents with scores
• Relevance scores for top documents
• Generated answers with citations
• Source attribution

Best Practices

1. Chunk Size: Balance between context and precision (200-500 tokens)
2. Chunk Overlap: 10-20% overlap maintains context between chunks
3. Reranking: Improves precision by re-scoring with query-document interaction
4. Citations: Always include sources for factual accuracy and transparency

Customization

You can customize:

• Chunk size and overlap
• Number of documents to retrieve and rerank
• Reranking algorithm
• Citation format
• System prompt and generation parameters