Agentic RAG System with Multi-Agent Orchestration

RAG + Web Search + General Chat + LangGraph Supervisor

This project implements an Agentic Retrieval-Augmented Generation (RAG) system using a three-agent architecture orchestrated by LangGraph.
It provides a clean Streamlit-based chat interface with intelligent routing between:

• A RAG agent (local PDF knowledge)
• A Web agent (internet search using Tavily)
• A General agent (conversation + reasoning)

The supervisor LLM decides dynamically which agent should handle each query.

---

1. Introduction

What this system is

This is an agent-based chatbot capable of:

• Answering from local PDFs using Retrieval-Augmented Generation
• Searching the internet for external information
• Handling casual chat and general reasoning
• Automatically choosing the best agent for each question

It is designed for academic, research, and technical workflows requiring both local knowledge and external information when needed.

---

What is RAG?

RAG (Retrieval-Augmented Generation) retrieves relevant information from an external knowledge base—such as PDFs or documents—and feeds it into an LLM.
This significantly improves accuracy and reduces hallucinations.

---

What is Agentic AI?

Agentic AI consists of multiple specialized agents working together.
A supervisor intelligently routes user queries to the right agent, creating:

• Better accuracy
• Cleaner answers
• More dynamic behavior

This project uses three agents.

---

2. Overview of the Three Agents

1. RAG Agent

• Uses Qdrant vector DB
• Uses BGE embeddings
• Retrieves answers from your local PDFs
• Best for academic and technical questions

2. Web Agent

• Uses Tavily Search API
• Fetches online results
• Uses embeddings + LLM reasoning to pick the most relevant result
• Summarizes reliably
• Best for geography, public figures, world info, etc.

3. General Agent

• Handles greetings, goodbyes, and casual chat
• Handles “Explain this concept” type questions
• Explains the system when explicitly asked
• Does not hallucinate content from PDFs or the web

---

3. How to Use the System

1. Place your PDFs inside the data/ folder.
2. Run the ingestion script to index the documents.
3. Launch the Streamlit UI.
4. Ask questions freely—the supervisor will route automatically:
   • RAG → For document-based answers
   • Web → For real-world queries
   • General → For chat or simple conceptual answers

The system ensures the best agent is always chosen.

---

4. Installation

Step 1 — Clone the Repository

git clone <https://github.com/Shardy2907/AcademicRagSystem.git>
cd <AcademicRagSystem>

---

Step 2 — Install uv (if not installed)

pip install uv

or

pipx install uv

---

Step 3 — Install dependencies

uv sync

This will create and manage a virtual environment using your pyproject.toml.

---

5. Install Ollama and Pull Model

Step 1 — Install Ollama:

https://ollama.com/download

Step 2 — Pull the model used by the system:

ollama pull phi3:mini

Step 3 — Ensure Ollama is running:

ollama serve

---

6. Build the RAG Vector Index (Ingest Data)

Run the ingestion script to process your PDFs:

uv run src/ingest_data.py

This will:

• Load PDFs from /data
• Split them into chunks
• Generate embeddings
• Store vectors in qdrant_local/

7. Set Environment Variables

Create a .env file:

TAVILY_API_KEY=your_api_key_here

Get your key from:

• https://app.tavily.com/

8. Run the Streamlit UI

Launch the chat interface:

uv run streamlit run src/streamlit_app.py

A browser window will open (usually at http://localhost:8501).