🚀 RAG Agent — Retrieval-Augmented Generation Pipeline

This project implements a Retrieval-Augmented Generation (RAG) agent using:

LangGraph for workflow orchestration

Ollama (gpt-oss, moondream) for LLM + embeddings

ChromaDB for vector storage

A multi-step pipeline with:

Query rewriting

Document retrieval

Relevance checking

Answer generation

It is designed to test and demonstrate graph-based, controllable RAG systems.

📌 Features ✅ 1. Graph-Orchestrated RAG Pipeline

The pipeline is built using StateGraph and contains these nodes:

agent → decides whether to retrieve or finish

tool → performs vector retrieval

check_relevance → evaluates if retrieved docs are relevant

rewrite → rewrites query to improve retrieval

generate → final answer generation

✅ 2. Automatic Query Rewriting Loop

If retrieved documents are not relevant, the system rewrites the query and retries until relevance is detected.

✅ 3. Custom Retrieval Logic

Uses Ollama Embeddings (moondream)

Stores vectors in ChromaDB

Supports persistent vector store (vector_store/ directory)

✅ 4. Full Logging / Step-by-Step Execution

Every step prints output so you can debug the internal flow.

📂 Project Structure rag_agent/ │ ├── rag_graph.py # Main RAG graph / pipeline ├── vector_store/ # Chroma persistent DB ├── .gitignore └── venv/ # Local environment (ignored)

🛠️ Installation

1. Clone repository git clone https://github.com/Malikabriq/rag_agent.git cd rag_agent

2. Create virtual environment python -m venv venv source venv/bin/activate # Mac/Linux venv\Scripts\activate # Windows

3. Install dependencies pip install -r requirements.txt

4. Install & run Ollama

Download Ollama: https://ollama.com/download

Pull required models:

ollama pull gpt-oss:120b-cloud ollama pull moondream:latest

▶️ Usage

Run the graph:

python rag_graph.py

You will see output showing every node:

STEP: agent decided to retrieve STEP: tool retrieved 4 docs STEP: relevance check = yes STEP: generate produced the answer

🧠 How the RAG Graph Works

The logic matches this workflow:

START ↓ agent → (function_call or finish) ↓ tool (retriever) ↓ check_relevance ├── yes → generate → END └── no → rewrite → tool → (loop)

This creates a self-improving retrieval loop until good results are found.

📦 Requirements

You can generate this file:

langgraph langchain-ollama langchain-chroma chromadb typing-extensions

🤝 Contributing

Pull requests are welcome! Feel free to open issues for:

Improving the agent logic

Adding more graphs or nodes

Enhancing retrieval quality

Integrating third-party APIs

📄 License

MIT License — free to use for personal and commercial purposes.

⭐ Support

If this project helps you, consider giving the repo a star ⭐ on GitHub!