---

AI SQL Query Generator | Chat with Database

This repository provides an AI-powered tool for generating SQL queries from natural language questions, with a focus on schema-aware, context-restricted query generation. It leverages large language models (LLMs) and provides both a web interface (via Streamlit) and supporting scripts for experimentation and database interaction.

---

Table of Contents

• Features
• Project Structure
• Setup & Installation
• Usage
  • Web App
  • Bonus Script
  • Command-line Experiments
  • Jupyter Notebook
• Model & Tokenizer
• Data Files
• Notes & Limitations

---

Features

• Natural Language to SQL: Converts user questions into SQL queries using an LLM (e.g., Gemma).
• Schema Awareness: Requires the user to input the database schema; the model is instructed to reject out-of-schema queries.
• Streamlit Web UI: User-friendly interface for schema and question input, and for displaying generated SQL.
• Database Execution: (Bonus) Optionally executes generated SQL on a local SQLite database and displays results.
• Experimentation Scripts: Includes scripts and a notebook for running and testing LLMs locally.

---

Project Structure

SQL-Query-Generator/
│
├── app.py                  # Main Streamlit web app for SQL generation
├── bonus.py                # Streamlit app with database execution features
├── try2.py                 # Command-line and subprocess experiments with LLMs
├── try.ipynb               # Jupyter notebook for model/tokenizer experiments
│
├── abc.db                  # Example SQLite database
├── train.csv               # Training data (CSV)
├── test.csv                # Test data (CSV)
├── validation.csv          # Validation data (CSV)
├── wikisql_train.json      # WikiSQL training set (JSON)
├── wikisql_test.json       # WikiSQL test set (JSON)
├── wikisql_validation.json # WikiSQL validation set (JSON)
│
├── model_directory2/
│   ├── model/              # Model weights and configs (e.g., Gemma)
│   └── tokenizer/          # Tokenizer files and configs
│
└── README.md               # This file

---

Setup & Installation

Prerequisites

• Python 3.8+
• Streamlit
• Requests
• sqlite3
• (For LLM) Ollama or compatible local LLM server
• (For notebook) transformers, bitsandbytes, peft, trl, accelerate, datasets, torch

Installation

1. Clone the repository:

   git clone https://github.com/Kush2806/SQL-Query-Generator.git
   cd SQL-Query-Generator

2. Install dependencies:

   pip install streamlit requests sqlite3
   # For notebook experiments:
   pip install bitsandbytes==0.42.0 peft==0.8.2 trl==0.7.10 accelerate==0.27.1 datasets==2.17.0 transformers==4.38.0 torch

3. (Optional) Set up Ollama or your LLM server:

   • Ensure the LLM (e.g., Gemma) is running and accessible at http://localhost:11434/api/generate.

---

Usage

Web App

Run the main Streamlit app:

streamlit run app.py

• Enter your database schema (e.g., table definitions).
• Enter your natural language question.
• The app will generate an SQL query using the LLM and display it.

Bonus Script

Run the bonus app for SQL generation and execution:

streamlit run bonus.py

• In addition to generating SQL, this script will attempt to execute the query on abc.db and display the results.

Command-line Experiments

try2.py contains various experiments for running LLMs via subprocess and PowerShell. You can run it directly:

python try2.py

• It demonstrates how to interact with the LLM from the command line and process outputs.

Jupyter Notebook

try.ipynb provides a playground for loading, quantizing, and running LLMs (e.g., Gemma) using HuggingFace Transformers and related libraries.

---

Model & Tokenizer

• The model_directory2/model/ and model_directory2/tokenizer/ folders contain the model weights and tokenizer files, respectively.
• These are used for local inference in the notebook and can be adapted for other scripts.

---

Data Files

• abc.db: Example SQLite database for query execution.
• train.csv, test.csv, validation.csv: Datasets for training, testing, and validation.
• wikisql_train.json, wikisql_test.json, wikisql_validation.json: WikiSQL datasets for more advanced experiments.

---

Notes & Limitations

• Schema Required: The LLM is instructed to only answer questions within the provided schema. Out-of-schema queries are rejected.
• No Explanation: The model is prompted to return only SQL queries, not explanations.
• Local LLM: The system expects a local LLM server (e.g., via Ollama) running the Gemma model.
• Database Safety: The bonus script executes generated SQL on the database. Use with caution and only on test databases.
• Model Files: Model and tokenizer files are large and not included in the repo. Place your own compatible files in model_directory2/.

---

Acknowledgements

• Streamlit
• Ollama
• HuggingFace Transformers
• WikiSQL Dataset

---

For questions or contributions, please open an issue or pull request.

---