🤖 Semantic Question Answering System

An intelligent QA system that understands meaning, not just keywords.

Semantic search over your knowledge base — ask in your own words and get accurate answers.

---

📖 Overview

This project is a semantic question-answering system that uses state-of-the-art sentence embeddings to match user questions to a curated knowledge base. Instead of brittle keyword matching, it uses the all-mpnet-base-v2 model (via Sentence-Transformers) to encode both questions and answers into dense vectors, then finds the best match using cosine similarity. That means you can rephrase questions, use synonyms, or ask in a different style and still get the right answer.

Aspect	Details
Model	sentence-transformers/all-mpnet-base-v2 (MPNet, 768-dim)
Backend	Flask (Python)
Data	CSV with question and answer columns
UI	Single-page web app with confidence scores and modern dark theme

---

✨ Features

• Semantic understanding — Matches by meaning, so "What's France's capital?" and "Capital of France?" both map to the same answer.
• Confidence scoring — Every response comes with a 0–100% match score so you know how reliable the answer is.
• Rich web UI — Dark theme, gradients, smooth animations, and a clean "Ask AI" experience.
• Easy to extend — Drop in a new CSV; the system indexes it at startup. No retraining required.
• GPU-ready — Uses PyTorch; if CUDA is available, encoding runs on GPU for faster responses.
• Fallback data — If the CSV is missing or invalid, a small built-in demo dataset keeps the app runnable.

---

🏗️ How It Works

┌─────────────────┐     ┌──────────────────┐     ┌─────────────────────┐
│  Your question  │ ──► │  MPNet encoder   │ ──► │  Cosine similarity   │
│  "Capital of    │     │  (sentence       │     │  vs. all stored      │
│   France?"      │     │   transformers)  │     │  question vectors    │
└─────────────────┘     └──────────────────┘     └──────────┬──────────┘
                                                             │
                                                             ▼
┌─────────────────┐     ┌──────────────────┐     ┌─────────────────────┐
│  Answer +       │ ◄── │  Return answer   │ ◄── │  Best-matching      │
│  confidence %    │     │  for best match  │     │  question index     │
└─────────────────┘     └──────────────────┘     └─────────────────────┘

1. Load — On startup, the app reads the CSV and encodes every question into a vector.
2. Query — Your question is encoded with the same model.
3. Match — Cosine similarity finds the stored question most similar to your query.
4. Respond — The corresponding answer and similarity score are returned to the UI.

---

📦 Prerequisites

• Python 3.11+ (3.11 or 3.12 recommended; the launcher will use what you have)
• pip (usually bundled with Python)
• ~500 MB disk for the first run (model weights are downloaded once and cached)

---

🚀 Installation & Run

1. Clone or download the project

git clone https://github.com/your-username/Question-Answering-System.git
cd Question-Answering-System

2. (Optional) Create a virtual environment

py -3.11 -m venv venv
.\venv\Scripts\activate   # Windows
# source venv/bin/activate  # macOS/Linux

3. Install dependencies

pip install -r requirements.txt

Package	Purpose
flask	Web server and API
pandas	Load and handle CSV data
torch	Tensor ops and optional GPU
sentence-transformers	MPNet model and encoding utilities

4. Run the app

Option A — One-click (Windows)
Double-click run_app.bat in the project root.

Option B — Command line

py -3.11 app.py
# or
python app.py

Then open in your browser: http://127.0.0.1:5000

⏳ First run: The app will download the all-mpnet-base-v2 model from Hugging Face (one-time). This can take 1–2 minutes depending on your connection.

---

📁 Project Structure

Question-Answering-System/
├── app.py                    # Flask app + SemanticQASystem logic
├── run_app.bat               # Windows one-click launcher
├── requirements.txt          # Python dependencies
├── 100_Unique_QA_Dataset.csv  # Default Q&A knowledge base (90 entries)
├── templates/
│   └── index.html            # Single-page UI (form + results + confidence)
├── static/                   # (optional) CSS/JS assets
└── README.md                 # This file

---

📄 Dataset Format

The system expects a CSV with exactly two columns:

Column	Description
question	The question or prompt (e.g. "What is the capital of France?")
answer	The answer to return (e.g. "Paris")

Example:

question,answer
What is the capital of France?,Paris
Who wrote 'To Kill a Mockingbird'?,Harper Lee
What is the largest planet in our solar system?,Jupiter

To use your own data, either replace 100_Unique_QA_Dataset.csv or pass a custom path when instantiating SemanticQASystem(csv_path="your_file.csv") in app.py.

---

🔧 Configuration

What	Where	Notes
CSV path	app.py → SemanticQASystem(csv_path=...)	Default: 100_Unique_QA_Dataset.csv
Embedding model	app.py → self.model_name	Default: all-mpnet-base-v2; other sentence-transformers models work too
Port	app.py → app.run(port=5000)	Change if 5000 is in use
Debug mode	app.py → app.run(debug=...)	Set True for auto-reload during development

---

🌐 API

The app exposes a simple POST endpoint for programmatic use.

POST /predict

• Content-Type: application/x-www-form-urlencoded
• Body: question=<your question>
• Response (JSON):

{
  "answer": "Paris",
  "confidence": 0.892
}

Example with curl:

curl -X POST http://127.0.0.1:5000/predict -d "question=What is the capital of France?"

---

❓ Troubleshooting

Issue	What to do
"No suitable Python runtime found"	Install Python 3.11+ or edit run_app.bat to use py or python instead of py -3.11.
ModuleNotFoundError: No module named 'flask'	Run pip install -r requirements.txt in the project folder.
First run is slow	Normal. The MPNet model is downloaded once; later starts are fast.
Low confidence on all answers	Your question may be far from any in the CSV; try rephrasing or adding more Q&A pairs.
Port 5000 already in use	Change port=5000 in app.run(...) in app.py to another port (e.g. 5001).

---

📜 License & Credits

• Sentence-Transformers — sentence-transformers (MPNet and utilities)
• Model — all-mpnet-base-v2 on Hugging Face
• Flask — pallets/flask

This project is provided as-is for learning and deployment. Adjust and extend as you like.

---

Built with Python · Flask · PyTorch · Sentence-Transformers

Ask in your own words — get the right answer.