Lecture: Building an End-to-End LLM Project for a T-Shirt Store

Overview

• The project aims to build a tool similar to ChatGPT that converts natural language questions into SQL queries to interact with a MySQL database.
• The store data includes t-shirt inventory and discounts.
• The project will use Google PaLM, LangChain, and other technologies.
• The goal is to allow the store manager to query the database in plain English.

Project Requirements

• Store Data: Includes four brands (V Hussein, Levis, Nike, Adidas) stored in a MySQL database.
  • Tables:
    • t-shirts: Maintains inventory count, price per unit.
    • discounts: Maintains discount percentage for each t-shirt.
• Key People:
  • Tony Sharma: Store manager, uses software to query the database.
  • Lok: Data analyst who writes SQL queries, often busy.
  • Peter P: Data scientist tasked with building the query tool.

Technical Architecture

• LLM (Google PaLM): Converts natural language questions into SQL queries.
• LangChain Framework: Utilizes Google PaLM within LangChain's SQLDatabaseChain class.
• Few-Shot Learning: Improves query accuracy by preparing a dataset of sample questions and corresponding SQL queries.
• Vector Database: Stores embeddings of the few-shot learning data.
  • Options include Pinecone, Milvus, ChromaDB (using ChromaDB for this project).
• Streamlit: Used to build the UI.
• API Keys: Set up through Google Maker Suite.

Implementation Steps

MySQL Database Setup

• Use MySQL Workbench for database operations.
• Create necessary tables (t-shirts and discounts) and populate sample data.

Jupyter Notebook Experimentation

• Setup: Import necessary libraries and set up Google PaLM API key.
• Data Connection: Connect Jupyter Notebook to the MySQL database.
• Creating LLM Object: Create and test Google PaLM LLM object.
• Sample Queries: Run sample queries to ensure the setup works.
• Handling Errors: Use few-shot learning to improve the model's performance on more complex queries.

Few-Shot Learning

• Purpose: Improve the LLM's ability to handle complex queries by providing sample question-query pairs.
• Embeddings: Use Hugging Face to generate embeddings for the sample data.
• Vector Database: Store embeddings in ChromaDB.
• Example Selector: Use semantic search to find the most similar examples from the vector database.
• Custom SQL Prompt: Use LangChain's pre-built SQL prompt to guide the LLM.
• Testing: Run various queries to test improvements.

Streamlit Interface

• Setup: Create a main Python file (main.py) to house the Streamlit code.
• UI Elements: Text input for questions, display answers.
• Integration: Transfer code from Jupyter Notebook to Python files, modularize the code.
• Run Streamlit: Test the user interface and functionality.

Conclusion

• The tool allows the store manager to query the database using natural language, improving efficiency.
• Encouraged to practice coding and explore the code provided in the video description.

Additional Resources

• LangChain Basics Video
• Vector Database Explanation Video
• MySQL Tutorial Video
• Embeddings Introduction Video