LangGraph Agentic RAG Workflow

A simple self-correcting and hallucination-checking RAG (Retrieval-Augmented Generation) workflow built using LangGraph, running LLMs locally via Ollama.

Features

• Retrieval-Augmented Generation (RAG): Leverages external knowledge sources (web search via Tavily and a local vector store) to enhance generation quality.
• Local LLM Execution: Uses Ollama (qwen3:30b by default) to run language models locally, avoiding reliance on external APIs for core generation.
• LangGraph: Implements the workflow as a state graph for clear logic and execution flow.
• Vector Store: Uses SKLearnVectorStore with NomicEmbeddings (local inference) for efficient document retrieval.
• Self-Correction: Includes mechanisms to review and refine generated answers based on retrieved documents.
• Hallucination Checking: Incorporates steps to verify the factual consistency of the generated output against source documents.
• Optional Caching: Supports caching the vector store embeddings locally (.parquet format) to speed up subsequent runs (requires pandas and pyarrow).

How it Works

This project uses LangGraph to define a cyclical process where:

1. A question is received.
2. The question is routed to either a web search (using Tavily) or the local vector store.
3. Relevant documents are retrieved (from web or vector store).
4. Retrieved documents are graded for relevance to the question.
5. An initial answer is generated by a local LLM (via Ollama) based on the relevant documents.
6. The answer is checked against the source documents for factual consistency (hallucination check) and relevance to the original question.
7. If inconsistencies or irrelevance are found, the process attempts to self-correct (e.g., by performing a web search if not already done, or regenerating the answer) up to a maximum number of retries.
8. A final, verified answer is provided.

Setup

1. Clone the repository:

   git clone https://github.com/safzanpirani/langgraph-agentic-workflow
   cd langgraph-agentic-workflow

2. Install Ollama: Follow the instructions at https://ollama.com/ to install Ollama on your system.

3. Pull the LLM Model: Download the model used in the script (or choose another compatible one):

   ollama pull qwen3:30b

4. Install Python dependencies:

   # Make sure you have pip or a similar package manager (like uv)
   pip install -r requirements.txt
   
   # OPTIONAL: For vector store caching
   # pip install pandas pyarrow 

5. Configure environment variables: Create a .env file in the root directory and add your Tavily API key:

   # Required for the web search tool
   TAVILY_API_KEY="your_tavily_api_key_here"

   Get your key by signing up at Tavily.com

Usage

1. Ensure Ollama is running: Start the Ollama application/server if it's not already running in the background.
2. Run the main workflow script:

   python main.py

Follow the terminal output to see the agent's reasoning and the final answer.

Live Demo

Watch a live demonstration of the agentic RAG workflow in action: Click here to watch the demo.