🤖 Smart Arabic Article Summarizer API

End-to-End Arabic NLP Application Deployed on OCI

Developed by: Abdulrahman Awadh Asban Internship @ Innovation Team — Riyadh, Saudi Arabia February–March 2026

---

📌 Project Overview

This project is a complete Arabic NLP pipeline that:

1. Accepts Arabic text input (Modern Standard Arabic or dialects)
2. Cleans and normalizes the text (removes diacritics, normalizes characters, strips punctuation)
3. Sends the cleaned text to a Hugging Face-hosted LLM via the Inference API
4. Returns a structured JSON response with the original text, cleaned text, word counts, and an AI-generated summary
5. Exposes the pipeline as a REST API via FastAPI
6. Provides an HTML interface for browser-based testing
7. Deployed on Oracle Cloud Infrastructure (OCI)

---

🧱 System Architecture

User (HTML Page)
      ↓
FastAPI Backend (main.py)
      ↓
Arabic Text Preprocessing (text_processing.py)
      ↓
Hugging Face Inference API (model.py)
      ↓
Structured JSON Response
      ↓
Display in HTML

Why Hugging Face Inference API instead of loading the model locally? Loading Arabic LLMs locally requires 8GB+ of RAM. Using the Inference API offloads model execution to Hugging Face servers, keeping the OCI VM lightweight and the solution 100% free.

---

📁 Project Structure

arabic_summarizer/
│
├── main.py              → FastAPI app & /summarize endpoint
├── text_processing.py   → Arabic text cleaning functions
├── model.py             → Hugging Face Inference API integration
├── templates/
│   └── index.html       → HTML testing interface
├── .env                 → API keys (Hidden by .gitignore file)
├── requirements.txt     → Project dependencies
└── README.md            → Project documentation

---

⚙️ Technical Stack

Component	Technology
Language	Python 3
API Framework	FastAPI + Uvicorn
AI Integration	Hugging Face Inference API
Models	Qwen/Qwen2.5-72B-Instruct, meta-llama/Llama-3.1-8B-Instruct
Frontend	HTML + JavaScript (Fetch API)
Cloud Platform	Oracle Cloud Infrastructure (OCI)
Secret Management	python-dotenv

---

🔧 Arabic Text Processing

Implemented in text_processing.py using Python's re (Regular Expressions) module.

Processing Steps

Step	What it does	Example
1. Remove diacritics	Strips all harakat (tashkeel) using Unicode range \u064B–\u0652	كَتَبَ → كتب
2. Normalize Alef	Converts أ إ آ → ا	انسان → إنسان
3. Remove punctuation	Strips non-Arabic, non-space characters	removes ، . ! etc.
4. Remove numbers	Removes Arabic (٠-٩) and Western (0-9) digits
5. Collapse whitespace	Replaces multiple spaces with single space

Why Regular Expressions?

re is Python's built-in library for pattern matching. It was chosen over specialized Arabic NLP libraries (like camel-tools or farasa) because:

• No installation required beyond standard Python
• Sufficient for structural text cleaning tasks
• Lightweight and fast for deployment on a small OCI VM

---

🤖 AI Model Integration

Implemented in model.py using the Hugging Face Inference API via requests.

Models Used & Why

Qwen/Qwen2.5-72B-Instruct

• Developed by Alibaba Cloud
• Strong multilingual capabilities with excellent Arabic support
• Handles both Modern Standard Arabic and dialects well
• Available on Hugging Face Inference API free tier

meta-llama/Llama-3.1-8B-Instruct

• Developed by Meta
• Lightweight (8B parameters), fast response
• Included for A/B testing and model comparison

Why not ALLaM (علام)?

ALLaM-7B was tested but returned an error: model too large to be loaded automatically. Hugging Face restricts models larger than 10GB from the free Serverless Inference API unless individual permission is granted. This is a known limitation documented in the Hugging Face community forums.

Model Evaluation Results

After testing both models on 6 Arabic texts (Formal Arabic, Riyadh dialect, Jeddah dialect):

	Formal Arabic	Riyadh Dialect	Jeddah Dialect
Qwen 2.5	✅ Excellent	✅ Excellent	✅ Excellent
LLaMA 3.1	⚠️ Code-switching (mixed Polish/English words)	⚠️ Hallucination (_FOLLOW)	❌ Complete breakdown (Chinese characters appeared)

Conclusion: Qwen 2.5 significantly outperforms LLaMA 3.1 on Arabic text, especially dialectal Arabic. This is likely due to larger Arabic training data in Qwen's pretraining corpus.

---

🚀 FastAPI Backend

Implemented in main.py.

Endpoint

POST /summarize

Request Body (JSON):

{
  "text": "النص العربي هنا",
  "model_choice": "qwen"
}

model_choice options: "qwen" (default) or "llama"

Response (JSON):

{
  "original_text": "...",
  "cleaned_text": "...",
  "original_word_count": 95,
  "cleaned_word_count": 94,
  "summary": "...",
  "model_used": "qwen"
}

Why FastAPI?

• Automatic documentation at /docs (Swagger UI)
• Built-in data validation via Pydantic
• High performance (async-capable)
• Industry standard for Python APIs

CORS Middleware

CORSMiddleware is added to allow the HTML frontend (served from a different origin) to communicate with the API without being blocked by the browser's same-origin policy.

---

🌐 HTML Interface

Located in templates/index.html.

• Arabic text input (RTL layout)
• Model selector (A/B Testing between Qwen and LLaMA)
• Displays: summary, word counts before/after cleaning, cleaned text
• Communicates with FastAPI via JavaScript fetch() POST requests

---

☁️ Deployment on OCI

VM Details

• Provider: Oracle Cloud Infrastructure (OCI)
• Region: Saudi Arabia Central (Riyadh)
• OS: Ubuntu (Linux)
• Instance Name: Abdulrahman test

Deployment Steps

1. Connect to VM:

ssh -i ssh-key-2026-01-29.key opc@<PUBLIC_IP>

2. Install dependencies:

sudo apt update
sudo apt install python3-pip python3-venv -y
python3 -m venv venv
source venv/bin/activate
pip install fastapi uvicorn requests python-dotenv

3. Upload project files and create .env:

echo "HF_API_KEY=your_key_here" > .env

4. Run the server:

uvicorn main:app --host 0.0.0.0 --port 8000

5. Configure OCI Security List: Open port 8000 in the OCI Console → VCN → Security List → Add Ingress Rule.

6. Access the app:

http://<PUBLIC_IP>:8000/docs

---

🔒 Security Notes

• API keys are stored in .env and never hardcoded in source files
• .env is excluded from version control (add to .gitignore)
• SSH private key permissions are restricted using icacls on Windows to comply with OpenSSH requirements

---

📦 Installation (Local)

# Clone the project
git clone <repo-url>
cd arabic_summarizer

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

# Install dependencies
pip install -r requirements.txt

# Create .env file
echo HF_API_KEY=your_key_here > .env

# Run
uvicorn main:app --reload

---

📚 References & Resources

Official Documentation

Resource	URL
Hugging Face Inference Providers Docs	https://huggingface.co/docs/inference-providers/providers/hf-inference
Python re Module Documentation	https://docs.python.org/3/library/re.html
Requests Library — POST Requests	https://requests.readthedocs.io/en/latest/user/quickstart/#more-complicated-post-requests
HTTP Status Codes — MDN	https://developer.mozilla.org/en-US/docs/Web/HTTP/Reference/Status
Unicode Arabic Block (U+0600–U+06FF)	https://unicode.org/charts/PDF/U0600.pdf

Arabic NLP References

Resource	URL
Removing Arabic Diacritics using Python — Stack Overflow	https://stackoverflow.com/questions/66988153/removing-arabic-diacritics-using-python
Unicode Explorer (Arabic character lookup)	https://unicode-explorer.com/search/
Arabic NLP Survey — MDPI Computers Journal	https://www.mdpi.com/2073-431X/14/11/497

Tutorials & Guides

Resource	URL
FastAPI Official Tutorial	https://fastapi.tiangolo.com/tutorial/
FastAPI — Request Body with Pydantic	https://fastapi.tiangolo.com/tutorial/body/
Hugging Face — Text Generation Task	https://huggingface.co/tasks/text-generation
OCI — Connect to Linux Instance via SSH	https://docs.oracle.com/en-us/iaas/Content/Compute/Tasks/accessinginstance.htm
OCI — Security Lists and Ingress Rules	https://docs.oracle.com/en-us/iaas/Content/Network/Concepts/securitylists.htm
python-dotenv Documentation	https://pypi.org/project/python-dotenv/
Pydantic V2 Documentation	https://docs.pydantic.dev/latest/
Arabic NLP Challenges — Towards Data Science	https://towardsdatascience.com/arabic-nlp-unique-challenges-and-their-solutions
Qwen2.5 Model Card — Hugging Face	https://huggingface.co/Qwen/Qwen2.5-72B-Instruct
LLaMA 3.1 Model Card — Hugging Face	https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct

---

This project was developed as part of an AI internship at Innovation Team, Riyadh, Saudi Arabia.