Intelligent E-Commerce FAQ Chatbot with Emotion Analysis

A complete end-to-end chatbot system that combines Retrieval-Augmented Generation (RAG), emotion analysis, and feedback collection for intelligent customer service automation.

---

Project Overview

This project implements a sophisticated chatbot system that:

• Answers FAQ questions using semantic search and local LLM generation
• Analyzes user emotions in real-time using fine-tuned DistilBERT
• Collects user feedback and ratings for continuous improvement
• Provides analytics dashboards for monitoring and optimization

---

Key Features

RAG (Retrieval-Augmented Generation)

• Semantic Search: FAISS vector database for FAQ retrieval
• Local LLM: TinyLlama-1.1B-Chat for response generation
• Smart Context: Combines relevant FAQs for accurate answers
• Fast Performance: 3-6 second response times

Emotion Analysis

• Real-time Detection: 28 different emotions using DistilBERT
• GoEmotions Dataset: Trained on 40,000 samples for accuracy
• Emotion Categories: Positive, Negative, Neutral, Complex emotions
• Business Intelligence: Correlate emotions with satisfaction ratings

Interactive Chat Interface

• Streamlit Web App: Clean, modern chat interface
• Smart Small Talk: Instant responses for greetings and casual conversation
• Session Management: Unique session tracking with timestamps
• MySQL Storage: Complete conversation logging

Analytics & Feedback

• Feedback Collection: Star ratings and text comments
• Emotion Dashboard: Visualize emotion trends and patterns
• Session Analysis: Detailed per-conversation emotion journeys
• Continuous Improvement: Data-driven optimization suggestions

---

System Architecture

User Input → Small Talk Handler → RAG Pipeline → Response Generation
     ↓              ↓                   ↓              ↓
Emotion Analysis → MySQL Storage → Analytics Dashboard → Insights

Core Components:

1. FAQ Embeddings: Sentence Transformers + FAISS
2. Response Generation: TinyLlama-1.1B local inference
3. Emotion Analysis: DistilBERT + GoEmotions dataset
4. Data Storage: MySQL with session and message tracking
5. User Interface: Streamlit with real-time chat
6. Analytics: Plotly visualizations and trend analysis

---

Installation & Setup

Prerequisites

• Python 3.8+
• 8GB+ RAM (for local LLM)
• MySQL server
• 4GB+ disk space

1. Clone and Setup Environment

git clone https://github.com/Aparnashree11/ChatBot_with_emotion_analysis.git
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

2. Install Dependencies

pip install -r requirements.txt

3. Setup MySQL Database

-- Create database
CREATE DATABASE chatbot_feedback CHARACTER SET utf8mb4 COLLATE utf8mb4_unicode_ci;

-- Create user (optional)
CREATE USER 'chatbot_user'@'localhost' IDENTIFIED BY 'your_password';
GRANT ALL PRIVILEGES ON chatbot_feedback.* TO 'chatbot_user'@'localhost';
FLUSH PRIVILEGES;

4. Environment Configuration

Create .env file:

MYSQL_HOST=localhost
MYSQL_USER=chatbot_user
MYSQL_PASSWORD=your_password
MYSQL_DB=chatbot_feedback

---

Quick Start Guide

Step 1: Generate FAQ Embeddings

python embeddings_gen.py

• Processes your FAQ data (CSV format)
• Creates FAISS vector store for semantic search
• Output: faq_vector_store.faiss + metadata

Step 2: Train Emotion Model

python emotion_classification_training.py

• Fine-tunes DistilBERT on GoEmotions dataset
• Training time: ~2-3 hours for 10,000 samples
• Output: distilbert-goemotions-10k/ model directory

Step 3: Launch Chat Interface

streamlit run new_app.py

• Opens chat interface at http://localhost:8501
• Users can chat and provide feedback
• Features: Real-time chat, session management, feedback collection

Step 4: View Analytics Dashboard

streamlit run main.py

• Analytics dashboard at http://localhost:8501
• Features: Emotion trends, session analysis, rating correlations

---

Usage Examples

Basic Chat Interface

from model_testing import create_balanced_chatbot

# Initialize chatbot
chatbot = create_balanced_chatbot("faq_vector_store")

# Ask questions
response = chatbot.ask("How do I return a product?")
print(f"Answer: {response.answer}")
print(f"Confidence: {response.confidence_score}")

Emotion Analysis

from emotion_classification_training import EmotionPredictor

# Load emotion model
predictor = EmotionPredictor("./distilbert-goemotions-10k")

# Analyze emotions
emotions = predictor.predict_top_k("I'm so frustrated with this!", k=3)
print(emotions)  # [('anger', 0.87), ('annoyance', 0.72), ('disappointment', 0.54)]

Streamlit Apps

# Chat interface
streamlit run new_app.py

# Analytics dashboard  
streamlit run main.py

---

Performance Benchmarks

Response Times

• Small Talk: 0.1 seconds (instant)
• FAQ Questions: 5-10 seconds average
• Cold Start: 10-15 seconds (model loading)

Accuracy Metrics

• FAQ Retrieval: Confidence-scored semantic matching
• Emotion Detection: F1-macro ~0.45-0.50 (28 emotions)
• User Satisfaction: Tracked via ratings and feedback

Resource Usage

• RAM: 2-3GB with model quantization
• Storage: 3-4GB total (models + embeddings)
• GPU: Optional but recommended for training

Configuration Options

RAG Chatbot Settings

chatbot = LocalRAGChatbot(
    embeddings_path="faq_vector_store",
    similarity_threshold=0.5,    # Higher = more selective
    max_retrieved_faqs=2,        # Number of FAQs to use
    device="cpu"                 # "cpu" or "cuda"
)

Emotion Model Settings

config = TrainingConfig(
    max_train_samples=10000,     # Training data size
    num_epochs=2,                # Training epochs
    batch_size=32,               # Batch size
    learning_rate=5e-5           # Learning rate
)

---

Troubleshooting

Common Issues

Model Loading Errors

# Check if model files exist
ls -la distilbert-goemotions-10k/
# Should contain: config.json, pytorch_model.bin, tokenizer files

Memory Issues

# Reduce batch size or use CPU
config.batch_size = 8
config.device = "cpu"

MySQL Connection

# Test MySQL connection
mysql -u root -p chatbot_feedback
# Verify tables exist: SHOW TABLES;

Slow Performance

• Use GPU if available
• Reduce FAQ database size
• Lower similarity threshold
• Use smaller models

---

Analytics & Insights

Emotion Analytics Dashboard

• Overall Trends: Emotion distribution across all chats
• Session Analysis: Click any session for detailed emotion journey
• Rating Correlation: How emotions affect user satisfaction
• Time Patterns: Emotion trends by day/hour

Business Metrics

• Customer Satisfaction: Average ratings and positive emotion rates
• Content Gaps: Identify FAQ areas needing improvement
• Performance Monitoring: Response times and confidence scores
• User Behavior: Conversation patterns and emotional journeys

---

Continuous Improvement

Feedback Loop

1. Users chat → Provide ratings and feedback
2. System analyzes → Emotion patterns and satisfaction
3. Identifies gaps → FAQ content and response quality
4. Generates insights → Optimization recommendations
5. Updates system → Better responses and user experience

Optimization Areas

• FAQ Content: Add missing questions, improve answers
• Model Parameters: Adjust similarity thresholds and context size
• Response Quality: Fine-tune prompts and generation settings
• User Experience: Enhance based on emotional feedback patterns

---

Contributing

Development Workflow

1. Fork the repository
2. Create feature branch: git checkout -b feature/new-feature
3. Make changes and test locally
4. Commit changes: git commit -m "Add new feature"
5. Push to branch: git push origin feature/new-feature
6. Create Pull Request

Code Standards

• PEP 8: Python style guide
• Type Hints: Use typing annotations
• Documentation: Docstrings for all functions
• Testing: Add tests for new features

---

Support

For questions, issues, or contributions:

• Create an Issue: Use GitHub Issues for bugs and feature requests
• Documentation: Check inline code documentation
• Community: Join discussions in project Issues

---