README.md

metadata

title: Multilingual Twi Health Information System
emoji: 🌍
colorFrom: green
colorTo: yellow
sdk: gradio
sdk_version: 4.3.0
python_version: '3.10'
app_file: app.py

🌍 Multilingual Twi Health Information System

A semantic search system that allows users to ask health-related questions in any language (Twi, Ga, Ewe, Hausa, English, etc.) and receive relevant answers in Twi.

🎯 Key Features

• 🌐 Multilingual Input: Ask questions in any language
• 🇬🇭 Twi Answers: All responses are provided in Twi
• 🔍 Semantic Search: Uses E5-Multilingual-Large embeddings for accurate matching
• ⚡ Fast Retrieval: FAISS-powered search across millions of Q&A pairs
• 📝 Three-Paragraph Format: Presents top 3 most relevant answers

💡 How It Works

1. User Input: Enter a question in any supported language
2. Embedding: Question is encoded using E5-Multilingual-Large model
3. Search: FAISS finds the top 3 most similar English questions
4. Response: Corresponding Twi answers are presented as 3 paragraphs

🚀 Quick Start

For Users

Simply visit the Space and:

1. Type your question in any language
2. Click "Hwehwɛ | Search"
3. Receive 3 relevant answers in Twi
4. Optional: Enable "Show matched questions" to see source questions

For Developers

Step 1: Create Embeddings (Google Colab)

# Upload your CSV with columns: question, answer
# Run the embedding creation script
# Download: faiss_index.bin, metadata.json, config.json

Step 2: Deploy to Hugging Face Spaces

1. Create a new Gradio Space
2. Upload required files:
   • app.py
   • requirements.txt
   • faiss_index.bin
   • metadata.json
   • config.json
3. Space will automatically build and deploy

📂 File Structure

.
├── app.py                 # Main Gradio application
├── requirements.txt       # Python dependencies
├── faiss_index.bin       # FAISS index (English questions)
├── metadata.json         # Question-answer pairs
├── config.json           # Model configuration
└── README.md            # This file

🔧 Technical Details

Model Architecture

• Embedding Model: intfloat/multilingual-e5-large
• Embedding Dimension: 1024
• Supported Languages: 100+ languages including:
  • Twi, Ga, Ewe, Hausa
  • English, French, Arabic
  • And many more

Search Configuration

• Similarity Metric: Cosine Similarity (Inner Product)
• Index Type: FAISS Flat Index
• Top-K Results: 3 (configurable)
• Encoding Strategy:
  • Questions: passage: prefix
  • User queries: query: prefix

Performance

• Search Speed: < 100ms for millions of records
• Accuracy: State-of-the-art multilingual semantic matching
• Scalability: Handles millions of Q&A pairs

📊 Data Format

Input CSV

question,answer
"Alcohol while breastfeeding?","Nufu yɛ papa ma wo ba..."
"Ambulance service number?","Ɛfa awoɔ mu ahohiahia..."

Requirements

• Columns: question (English), answer (Twi)
• Format: UTF-8 encoded CSV
• Size: Tested with 3M+ rows

🌟 Example Queries

Language	Question	Result
English	"What should I do about alcohol while breastfeeding?"	✅ Finds relevant answer
Twi	"Dɛn na menyɛ fa alcohol ho wɔ nufunom bere mu?"	✅ Finds relevant answer
Ga	"Mɛnya ambulance service frɛ nɔma no?"	✅ Finds relevant answer

🔐 Privacy & Security

• ❌ No user data is stored
• ✅ All processing happens in real-time
• ✅ No tracking or analytics
• ✅ Open-source and transparent

🛠️ Customization

Adjust Number of Results

# In app.py, modify:
results = search_answers(query, top_k=5)  # Change from 3 to 5

Change Response Format

# Modify format_response() function in app.py
# Customize how answers are presented

Add New Languages

The system automatically supports 100+ languages through E5-Multilingual-Large. No configuration needed!

📈 Performance Optimization

For large datasets:

• Use Git LFS for faiss_index.bin (>100MB)
• Consider quantized FAISS indices for faster search
• Enable caching for frequently asked questions

🤝 Contributing

Contributions are welcome! Areas for improvement:

• Add more Q&A pairs
• Improve answer formatting
• Add audio input/output
• Implement feedback mechanism

📄 License

[Add your license here]

🙏 Acknowledgments

• Model: E5-Multilingual-Large by Microsoft
• Framework: Sentence Transformers, FAISS, Gradio
• Data: [Add your data source]

📞 Support

For issues or questions:

• Open an issue in the Space discussions
• Contact: [Your contact information]

---

Medaase! | Thank you! 🇬🇭