RAG-Ultra

RAG-Ultra is a high-accuracy multimodal RAG (Retrieval-Augmented Generation) chatbot designed to handle complex documents including scanned PDFs, tables, formulas, and visual content.

It leverages a robust tech stack to ensure reliability and detailed reasoning:

• Amazon Textract: For enterprise-grade OCR ingestion, capable of preserving layout and extracting tables/forms from scanned documents.
• Advanced Retrieval: Implements Parent-Document Retrieval to maintain context and reduce hallucinations.
• Reranking: integrated cross-encoder/reranking steps to ensure the most relevant chunks are sent to the LLM.
• Multimodal capabilities: Uses Vision-Language (VL) models for reasoning over chart/diagram extracts.
• LangSmith: Full observability integration for tracing and debugging.

Architecture

1. Ingestion Pipeline:

   • PDFs are processed via boto3 and Amazon Textract.
   • Text, tables, and raw image regions are extracted.
   • Data is chunked and indexed. Text goes to a vector store; images are summarized by a VLM (e.g., GPT-4o, Claude 3.5 Sonnet) and embedded, or indexed purely by their summaries.

2. Retrieval:

   • Parent Document Retriever: Fetches larger context blocks for retrieved small chunks.
   • Reranking: Re-orders results based on relevance score to improve precision.

3. Generation (LangGraph):

   • A directed cyclic graph (DAG) manages the state.
   • Nodes: Retrieve -> Grade Documents -> Rerank -> Generate.
   • Support for "Visual Reasoning" loops if image context is retrieved.

Setup

1. Clone the repository:

   git clone https://github.com/SriviharReddy/RAG-Ultra.git
   cd RAG-Ultra

2. Install dependencies:

   pip install -r requirements.txt

3. Environment Variables: Copy .env.example to .env and fill in your keys:

   • OPENAI_API_KEY (for embeddings/generation)
   • AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY (for Textract)
   • LANGCHAIN_API_KEY (for LangSmith)

4. Run the App:

   streamlit run app.py

Tech Stack

• LangChain / LangGraph
• Streamlit
• Amazon Textract
• ChromaDB (Vector Store)
• LangSmith

License

MIT