🤖 Trend Setters - E-Commerce Conversational Chatbot

Product submitted to Cognizant Kolkata through Digital Nurture TechnoVerse East Hackathon

Features • Installation • Tech Stack • Usage • Future Enhancements

📌 Overview

📝 Project Overview

• This project is an AI-powered conversational chatbot built for e-commerce platforms.
• It enables users to discover and explore fashion products through natural conversations instead of traditional keyword searches.
• By combining Natural Language Processing (NLP) with Vector Search (Qdrant), the chatbot delivers highly relevant and personalized product recommendations.
• The system is designed with a Flask backend, an interactive frontend UI, and integrates state-of-the-art LLaMA models (via Groq API and Hugging Face) for intelligent query understanding and recommendation.

---

🚀 Features

• Conversational AI Assistant: Powered by LLaMA 3 (70B) via Groq API and Quantized meta-llama/Llama-3.1-8B via HF for natural and human-like conversations.
• Intelligent Vector Search: Leverages Qdrant Vector Database to find the most relevant fashion products using semantic similarity.
• Multi-Attribute Query Handling: Supports dynamic filtering based on brand, color, category, gender, size, and more.
• Context-Aware Query Processing: Understands partial queries and prompts users for missing details to refine results.
• Fallback Recommendation System: Displays 10 random trending products when no attributes are provided, ensuring the user always sees results.
• Optimized Model Performance: Supports 4-bit Quantization (Linux/CUDA) and 8-bit LoRA (Windows) for faster, resource-efficient inference.
• Scalable Architecture: Easily extendable to integrate new product categories, recommendation strategies, or external APIs.

---

🛠️ Tech Stack

• Frontend: HTML, CSS, JavaScript
• Backend: Flask, Python
• Vector Database: Qdrant
• Groq LLM Model: Groq API using llama-3.3-70b-versatile (powered by META)
• Hugging Face LLM Model: Hugging Face Llama-3.1-8B (powered by META)
• Fine-Tuning: 4-bit Quantization for CUDA/Linux & LoRA (8-bit) for Windows
• Deployment: TBD

Note : Due to Limited Resource, We have used ChatGroq due to resource constraint. If proper GPU support is available we can use the Quantized FT version of the HF-LLM

---

📚 Dataset - Fashion E-Commerce Data Set

📎 Source: Custom-curated dataset integrated with Qdrant Vector Database

The Fashion E-Commerce Data Set is a structured product dataset containing detailed information about fashion items.
It is used to power semantic search and AI-driven product recommendations in this project.

📁 Dataset Highlights:

• Total Records: ~3,000 fashion products (expandable)
• Type: Tabular (Structured)
• Task: Product Recommendation & Attribute-based Retrieval

✅ Why This Dataset?

• Contains rich product attributes (brand, color, size, category, gender, ratings).
• Enables multi-attribute filtering through natural conversations.
• Optimized for semantic similarity search with vector embeddings.
• Suitable for building personalized recommendation systems in e-commerce.

---

🔧 Installation

1. Clone the repository

   git clone [email protected]:priyam-hub/Trend-Setters.git
   cd Trend-Setters

2. Create Conda Environment and Install Dependencies

   # CONDA ENV CREATED ALONG WITH DEPENDENCY INSTALLED
   conda env create -f environment.yml
   
   # ACTIVATE CONDA ENVIRONMENT
   conda activate <env_name>

3. Set up environment variables

   • Create a .env file in the root directory of the Project and add the following Credentials:

   GROQ_API_KEY              = "YOUR_GROQ_API_KEY"
   LLM_MODEL_NAME            = "YOUR_LLM_MODEL_NAME"
   HF_LLM_MODEL_NAME         = "YOUR_HF_LLM_MODEL_NAME"
   QDRANT_API_KEY            = "YOUR_QDRANT_API_KEY"
   QDRANT_CLUSTER_URL        = "YOUR_QDRANT_CLUSTER_URL"
   QDRANT_COLLECTION_NAME    = "YOUR_QDRANT_COLLECTION_NAME"
   HUGGINGFACE_LOGIN_TOKEN   = "YOUR_HUGGINGFACE_LOGIN_TOKEN"

4. Run the Flask server

   python web/app.py

5. Access the E-Commerce Platform Open http://127.0.0.1:5000 in your browser.

---

🔍 Usage

1. Enter a Query: Type a product-related query (e.g., "Show me blue Adidas shoes").
2. NLP Processing: The chatbot extracts attributes like brand, color, and category from your query.
3. Smart Filtering:
   • If attributes are provided → Relevant products are displayed.
   • If attributes are missing → It suggests similar items and prompts for more details.
4. Fallback Search: If no attributes are found, 10 random products are shown with a prompt to refine the search.

📸 Example Search Results

Query	Result
"I am Looking for women Black Top"
"I am Looking for Women Red Kurtis"
"Show me women blue jeans"
"Show me Kurta-Sets for my wife"
"Show me tops to gift in birthday for my sister"

---

⚡ Performance & Inference Time of the Model

We benchmarked meta-llama/Llama-3.1-8B across different hardware accelerators and configurations (4-bit & 8-bit).
The chart below shows the predicted inference time (in seconds) for different setups:

📊 Summary Table

Model	Hardware	Config	Predicted Time (s)
meta-llama/Llama-3.1-8B	CPU	4-bit	~60
		8-bit	~110
meta-llama/Llama-3.1-8B	T4 GPU	4-bit	10
		8-bit	17
meta-llama/Llama-3.1-8B	A100 GPU	4-bit	~3–4
		8-bit	~6–7
meta-llama/Llama-3.1-8B	L4 GPU	4-bit	~5
		8-bit	~9
meta-llama/Llama-3.1-8B	v5e-1 TPU	4-bit	~8
		8-bit	~13
meta-llama/Llama-3.1-8B	v2-8 TPU (Deprecated)	4-bit	~15
		8-bit	~25
meta-llama/Llama-3.1-8B	v6e-1 TPU	4-bit	~6
		8-bit	~11

---

📂 Project Structure

Trend-Setters/
├── .env                                     # Optional environment variables for API keys, database URLs, and other secrets
├── .gitignore                               # Git ignore rules to exclude unnecessary files from version control
├── environment.yml                          # Conda environment configuration with all dependencies
├── LICENSE                                  # MIT License for project usage and distribution
├── main.py                                  # Main script to run the application (entry point)
├── README.md                                # Project documentation and overview
├── requirements.txt                         # Python package dependencies for pip installation
├── config/
│   ├── __init__.py                          # Marks config as a Python package
│   └── config.py                            # Configuration settings for API keys, model paths, and constants
├── data/
│   └── V-1.01_Updated_Fashion_Dataset.csv   # Fashion dataset containing product attributes
├── images/
│   ├── image.png                            # General images used in README or UI
│   ├── search_result_1.png                  # Example screenshot of search result 1
│   ├── search_result_2.png                  # Example screenshot of search result 2
│   └── search_result_3.png                  # Example screenshot of search result 3
├── logger/
│   ├── __init__.py                          # Marks logger as a Python package
│   └── logger.py                            # Custom logging utility for tracking events and errors
├── models/
│   └── meta-llama/Llama-3.1-8B/             # Pretrained Hugging Face LLaMA 3.1 8B model
│       ├── model/                           # Model weights and architecture
│       └── tokenizer/                       # Tokenizer files for text processing
├── notebooks/
│   └── Trend-Setters.ipynb                  # Jupyter notebook for experiments, analysis, and prototyping
├── src/
│   ├── extractor/  
│   │   ├── __init__.py                      # Marks extractor as a Python package
│   │   └── extractor.py                     # Extracts product attributes from user queries
│   ├── llm/
│   │   ├── __init__.py                      # Marks LLM module as a package
│   │   └── llm_builder.py                   # Loads and configures LLM models (Groq & Hugging Face)
│   ├── parser/   
│   │   ├── __init__.py                      # Marks parser as a Python package
│   │   └── parser.py                        # Parses extracted attributes into search-friendly format
│   └── searcher/
│       ├── __init__.py                      # Marks searcher as a Python package
│       └── extractor.py                     # Performs vector search on Qdrant and retrieves products
├── test/
│   ├── __init__.py                          # Marks test folder as a Python package
│   └── test.py                              # Unit tests for different modules
└── web/
    ├── static/
    │   ├── css/
    │   │   ├── styles_home.css              # CSS for homepage
    │   │   └── styles.css                   # General styling for web pages
    │   ├── images/
    │   │   ├── icon/                        # Icons used in the UI
    │   │   ├── logo/                        # Logo images
    │   │   └── products/                    # Product images for frontend display
    │   └── js/
    │       ├── home.js                      # JS for homepage functionalities and search
    │       └── chatBot.js                   # JS for chatbot interactions and modal/chat window
    ├── templates/
    │   ├── index_home.html                  # Homepage template
    │   └── index.html                       # Main index page template
    ├── __init__.py                          # Marks web folder as a Python package
    └── app.py                               # Flask application file to run backend and route endpoints

---

🔮 Future Enhancements

🚀 Phase 1 – Core Improvements

• User Authentication: Enable login/logout with saved chat history.
• Better Query Understanding: Improve NLP pipeline for more accurate attribute extraction.
• UI/UX Enhancements: Add responsive chatbot widget with modern design.
• Randomized Recommendations: Smarter fallback mechanism with trending products instead of purely random ones.

🌍 Phase 2 – Advanced Features

• Multi-Language Support: Let users interact in multiple languages.
• Voice Search: Integrate speech-to-text for hands-free chatbot queries.
• Recommendation Engine: Personalized product suggestions using collaborative filtering.
• Image-Based Search: Upload product images to get visually similar results.

☁️ Phase 3 – Enterprise & Scalability

• Real-Time Inventory Sync: Connect with live e-commerce databases to track stock availability.
• Payment Integration: Allow direct checkout and transactions within chatbot.
• Analytics Dashboard: Provide admins with insights on queries, popular searches, and conversion rates.
• Cloud Deployment & Scaling: Deploy with Docker/Kubernetes on AWS/GCP/Azure for high availability.
• Security Enhancements: Implement JWT authentication, role-based access, and encrypted APIs.

---

📜 License

This project is licensed under the MIT License.

Made with ❤️ by Team Technocrats

↑ Back to Top