Madam Afaf：基于 NileChat-3B 与 RAG 的埃及政务服务阿拉伯语问答助手

🚀 Tips Hindawi Internship (August–October) 2026

🎓 This project was built during the Tips Hindawi Internship (August–October) 2026.

👤 Participant

Field	Value
Full Name	Mohamed Younis Mohamed
Project Name	مدام عفاف \| Madam Afaf
GitHub Username	mohamedyounis10
Internship Batch	August–October 2026
Training Program	Large Language Models (LLMs) Program
Organization	Edrak for Ai

📖 Project Overview

مدام عفاف (Madam Afaf) is an AI-powered chatbot that helps Egyptian citizens get instant, accurate information about government services — in Egyptian Arabic dialect.

The system scrapes 165 government services from 3 official Egyptian government websites, builds a RAG (Retrieval-Augmented Generation) pipeline using FAISS vector search and BAAI/bge-m3 embeddings, and generates grounded answers using NileChat-3B — an Arabic/Egyptian-dialect LLM. The frontend is a Streamlit app with a full RTL Arabic UI, while the backend runs on Kaggle GPU (T4) and is exposed via an ngrok tunnel.

"الموظفة اللي مش هتقولك فوت علينا بكرة! 😉"

✨ Features

🤖 Arabic RAG Chatbot — answers questions about Egyptian government services using only verified, official sources (no hallucination)
🗣️ Egyptian Dialect Support — powered by NileChat-3B, understands and responds in Egyptian Arabic
🌐 Real-time Web Scraping — collects 165 services from 3 official gov.eg websites using parallel scraping (ThreadPoolExecutor)
📚 Source Attribution — every answer comes with clickable links to the official government source
🎨 Full RTL Arabic UI — Streamlit app with custom CSS for right-to-left layout, dark theme, and smooth animations
💰 Zero-cost Infrastructure — runs entirely on free-tier tools (Kaggle GPU + ngrok + Streamlit)

🛠️ Technologies Used

Category	Technology
LLM	UBC-NLP/NileChat-3B (Arabic & Egyptian dialect)
Embeddings	BAAI/bge-m3 (multilingual, Arabic-optimized)
Vector Store	FAISS (in-memory, GPU-accelerated)
RAG Framework	LangChain (chunking, retrieval, prompt templates)
Web Scraping	BeautifulSoup4, Requests, ThreadPoolExecutor
Deep Learning	PyTorch, HuggingFace Transformers
Frontend	Streamlit (RTL Arabic UI with custom CSS)
Backend	FastAPI (served from Kaggle Notebook)
Tunnel	ngrok (Kaggle → local Streamlit)
Compute	Kaggle (NVIDIA T4 GPU — free tier)
Language	Python

⚙️ Installation

1 — Clone the Repository

git clone https://github.com/mohamedyounis10/madam-afaf.git
cd madam-afaf

2 — Install Streamlit Dependencies

pip install streamlit requests

3 — Setup the Kaggle Notebook

Upload citizen-services-rag-prototype.ipynb to Kaggle
Enable GPU T4 — Settings → Accelerator → GPU T4
Enable Internet — Settings → Internet → On
Add your HuggingFace token to Kaggle Secrets as HF_TOKEN
Run All Cells (~15 minutes on first run)

4 — Setup ngrok

Add the following at the end of the Kaggle Notebook:

from pyngrok import ngrok
ngrok.set_auth_token("YOUR_NGROK_AUTHTOKEN")
public_url = ngrok.connect(8000).public_url
print(f"API URL: {public_url}")

5 — Update API URL in `app.py`

# Line 242 in app.py
API_URL = "https://YOUR-NGROK-URL.ngrok-free.app"

6 — Run the App

streamlit run app.py

Opens at http://localhost:8501

🚀 Usage

Start the Kaggle Notebook — run all cells to load the model, scrape data, build the vector store, and start the API
Copy the ngrok URL from the notebook output and paste it into app.py
Run streamlit run app.py locally
Ask in Egyptian Arabic! — type any question about Egyptian government services

Example questions:

إزاي أطلّع بطاقة رقم قومي بدل فاقد؟
عايز أجدد جواز سفري، محتاج إيه؟
ما هي المستندات المطلوبة لاكتساب الجنسية المصرية؟
إيه إجراءات استخراج بطاقة تموينية ذكية جديدة؟

📸 Demo

🖥️ App Interface

🎬 Demo Video

Click the thumbnail below to watch the full demo

▶️ Click here to watch the demo video

📈 Results

Metric	Value
Government websites scraped	3 official sources
Total services collected	165 services
RAG chunks created	206 chunks
Web scraping time	~10 seconds (parallel)
Vector store build time	~40 seconds
Retrieval accuracy	Top-4 most relevant chunks
Total infrastructure cost	$0 (fully free-tier)

Quality Highlights:

✅ Accurate answers grounded exclusively in official government sources
✅ Refuses to answer out-of-scope questions — no hallucination
✅ Cites the official source URL with every answer
✅ Understands Egyptian Arabic dialect naturally

🔮 Future Improvements

📅 Scheduled scraping — automatically update government data periodically
🏛️ More sources — add tax authority, commercial registry, and Ministry of Labor websites
💾 Persistent vector store — save FAISS index to disk to avoid rebuilding on every session
🚀 Cloud deployment — migrate backend to Hugging Face Spaces or a cloud GPU
📱 Mobile app — build a mobile-friendly interface for easier access
🔄 Streaming responses — return answers progressively instead of waiting for full generation
📊 Analytics dashboard — track most-asked questions to improve coverage

📚 About the Internship

This project was developed as part of the Tips Hindawi Internship (August–October) 2026), and it will be showcased on the official Tips Hindawi website.

Tips Hindawi is the internships department of Edrak for Ai, and the internship encourages participants to build real-world projects, apply practical skills, and showcase their work through GitHub.

For more information about the internship, training programs, and upcoming batches, visit the official Tips Hindawi website.

📄 License

This project is shared for educational and portfolio purposes.