Semantic FAQ Chatbot 🚀

An intelligent FAQ-based chatbot built with SentenceTransformers. It combines semantic similarity and keyword matching to provide accurate answers to user queries. The chatbot supports API integration, fine-tuning, logging, and automatic FAQ updates.

---

✨ Features

• Hybrid Search: Combines semantic similarity (SentenceTransformers) with keyword matching.
• Fallback Mechanism: Returns a default response if similarity is too low.
• Embedding Storage: Saves embeddings with Pickle for faster startup.
• FastAPI Integration: Provides REST API endpoints for chatbot interaction.
• Automatic Logging: Stores all queries and responses in CSV logs.
• FAQ Correction: Dynamically update chatbot answers.
• Fine-tuning: Retrain the model on custom FAQ data.
• Evaluation with BERTScore: Measure model accuracy before and after fine-tuning.
• Dataset: Uses the Ecommerce FAQ Dataset.

---

📦 Requirements

Install dependencies:

pip install pandas sentence-transformers torch datasets scikit-learn fastapi uvicorn bert-score

---

⚙️ Usage

1. Run FastAPI server

uvicorn your_file_name:app --reload

The server will start at: http://127.0.0.1:8000

2. API Endpoints

• Ask the chatbot:

POST /chat
{
  "query": "What are your payment methods?"
}

• Correct chatbot response:

POST /correct
{
  "question": "What are your payment methods?",
  "correct_answer": "Payment is available only via credit card."
}

---

🧠 Training & Evaluation

• Evaluate baseline accuracy: Uses BERTScore to measure model performance on test data.

• Fine-tuning the model:

fine_tune(df, epochs=2, batch_size=8)

• Re-evaluate after fine-tuning: BERTScore is run again to compare performance before/after training.

---

📊 Project Structure

• ecommerce_faq.csv → Main FAQ dataset
• faq_embeddings.pkl → Pre-computed embeddings
• chat_logs.csv → Query/response logs
• fine_tuned_model/ → Saved fine-tuned model

---

🔮 Next Steps

• Add a user-friendly interface (web/Telegram).
• Multilingual support (English + Persian).
• Database integration (SQLite, MongoDB).
• Speed optimization using FAISS or Milvus.

---