NexaAI Self-RAG Assistant

A Self-RAG (Retrieval-Augmented Generation) chatbot built using LangGraph + Ollama + Streamlit. It intelligently decides when to retrieve company documents and ensures answers are grounded, useful, and verifiable.

Features

• Smart Retrieval Decision
  • Uses keyword rules + LLM to decide whether to retrieve docs or answer directly
• Document RAG Pipeline
  • Loads PDFs → chunks → embeddings → FAISS vector store
• Self-RAG Loops
  • Relevance filtering (keeps only useful chunks)
  • IsSUP → verifies grounding (with revise loop)
  • IsUSE → checks usefulness (with query rewrite loop)
• Query Rewriting
  • Automatically improves retrieval queries when answers are weak
• Trace Debug Panel (Streamlit)
  • Step-by-step reasoning
  • Grounding + usefulness badges
  • Sources + evidence

Tech Stack

• LangGraph → pipeline orchestration
• Ollama → local LLM + embeddings
• FAISS → vector database
• Streamlit → frontend UI

Project Structure

.
├── backend.py          # Self-RAG pipeline (LangGraph)
├── app.py              # Streamlit frontend
├── documents/          # PDF knowledge base
│   ├── Company_Policies.pdf
│   ├── Company_Profile.pdf
│   └── Product_and_pricing.pdf
├── .env
└── README.md