🛡️ RAGuard

A command-line AI system for Retrieval-Augmented Generation (RAG) with built-in hallucination detection.
Powered by OpenRouter’s free 🔓 DeepSeek-R1 70B model.

---

🚀 What is RAGuard?

RAGuard is a focused, minimalistic AI pipeline that:

1. Retrieves answers from provided documents using semantic relevance.
2. Verifies those answers for hallucinations using an LLM-based verifier.
3. Flags unsupported content, helping you trust what your AI says.

It’s optimized for developers and interviewers who want to see real-world AI reliability in action.

---

🗂️ Project Structure

raguard/
├── guardrag.py           # Main script (RAG + verifier logic)
├── config.py             # Store your OpenRouter API key
├── requirements.sh       # Quick installer
├── requirements.txt      # Python package list
├── sample_input.json     # Sample docs & question
└── README.md             # This file

🔁 You can rename guardrag.py to raguard.py if you prefer — same logic, new name.

---

⚙️ Setup Instructions

✅ Requires Python 3.7+

1. Clone the project

git clone https://github.com/yourusername/raguard.git
cd raguard

2. Install dependencies

bash requirements.sh

3. Get your OpenRouter API key
   • Go to: https://openrouter.ai
   • Sign in, generate a free key
   • Paste it into config.py:

# config.py
OPENROUTER_API_KEY = "your-api-key-here"

---

🧪 How to Use

1. Prepare input file Create or edit sample_input.json like this:

{
  "question": "What is the capital of France and what iconic structure is there?",
  "documents": [
    "The capital of France is Paris.",
    "France is a country in Europe known for its history and cuisine."
  ]
}

2. Run the script

python guardrag.py sample_input.json

---

🧠 How RAGuard Works

Step	What Happens
🧲 Retrieval	Uses sentence embeddings to select top relevant documents
💬 RAG Answer	Sends context + question to DeepSeek-R1
🕵️ Verifier Agent	Sends answer back to DeepSeek and asks it to check for unsupported claims
⚠️ Output	Returns the full answer and flags hallucinations if found

---

✅ Sample Output

Original Answer:
The capital of France is Paris. An iconic structure found there is the Eiffel Tower.

⚠️ Hallucinations flagged:
- "Eiffel Tower" is not supported by the provided context.

💡 If your documents had mentioned "Eiffel Tower", it would pass cleanly.

---

🛠️ Tech Stack

• 🔗 OpenRouter API with DeepSeek-R1 (free tier)
• 🧠 SentenceTransformers for document similarity
• 🐍 Python — no extra framework, pure and fast
• ✅ Runs on CPU — no GPU needed

---

📌 Use Cases

• ✅ AI QA systems with hallucination control
• ✅ Prototypes for enterprise RAG flows
• ✅ Technical interviews and portfolio projects

---

📎 Future Improvements

• Add CLI options: --top-k, --strict, --log-json
• Integrate with n8n for UI-based chaining
• Add citation-style context markup

---

🎯 Why RAGuard Exists

Built to solve flaws in Botminds v25’s RAG flow, RAGuard introduces a simple yet powerful verification layer that catches hallucinations before they cause damage.

If you're building AI systems that make decisions from documents — you need RAGuard.

---

📄 License

MIT — clone it, fork it, deploy it.