SQL Agent with Gemini 2.5 Flash

DEMO VIDEO LINK - https://drive.google.com/file/d/1zsSJWoBGnxqsaaVBHJxpLhSxRpmXSI38/view?usp=sharing

This project is a web-based AI SQL assistant that allows users to interact with a database using natural language. Users can upload CSV files, which are converted into a searchable SQLite database. The Gemini 2.5 Flash LLM is used to convert natural language questions into SQL queries, validate them, execute them, and return human-readable answers.

The project is a full-stack application with a FastAPI backend and a Streamlit frontend.

---

Live Application

• Backend (Render): https://sql-agent-mpka.onrender.com/
• Frontend (Streamlit): [https://aisqlagent.streamlit.app//)

The frontend reads the backend URL from st.secrets["BACKEND_URL"].

---

Project Structure

• app.py – FastAPI backend. Handles CSV uploads, SQL query generation, validation, execution, and converting results to natural language.
• frontend.py – Streamlit frontend. User interface for uploading CSVs and asking questions.
• requirements.txt – List of Python dependencies.
• .env – Environment variables including GOOGLE_API_KEY. Ignored by Git.
• uploaded.db – SQLite database storing uploaded CSV data.

---

Features

1. CSV Upload Upload CSV files and store them as SQLite tables for querying.

2. Natural Language to SQL Uses Gemini 2.5 Flash LLM to generate SQL queries that respect the database schema.

3. SQL Validation Queries are validated and corrected before execution to ensure correctness and security.

4. Human-Readable Answers Query results are converted into natural language summaries for easy understanding.

5. Modular Workflow The backend uses a state graph workflow:

   query_gen → query_check → execute_query → nl_output
   

   • query_gen: Generate SQL from natural language.
   • query_check: Validate and correct SQL.
   • execute_query: Run SQL on SQLite.
   • nl_output: Convert results into natural language.

---

Getting Started

Prerequisites

• Python 3.11+
• A valid Google Gemini API key

Local Setup

1. Clone the repository

git clone https://github.com/your-username/your-repo-name.git
cd your-repo-name

2. Create and activate a virtual environment

python -m venv venv
# Windows
venv\Scripts\activate
# macOS/Linux
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Set up environment variables Create a .env file:

GOOGLE_API_KEY="YOUR_API_KEY_HERE"

5. Run backend and frontend locally (optional) In separate terminals:

Backend:

uvicorn app:app --reload

Frontend:

streamlit run frontend.py

In deployed mode, the frontend reads the backend URL from st.secrets["BACKEND_URL"].

---

API Endpoints

Endpoint	Method	Description
/upload	POST	Upload CSV file and store as a table.
/ask	POST	Ask a natural language question. Returns human-readable answer.
/schema	GET	Return current database schema (tables + columns).
/	GET	Basic welcome message.

---

Deployment

• Backend: Deploy on Render (or similar cloud platforms).
• Frontend: Deploy on Streamlit Community Cloud, linking your GitHub repository.
• Secrets: Add BACKEND_URL in Streamlit Cloud secrets pointing to your Render backend.

---

Workflow Architecture

The backend uses a state graph workflow to dynamically generate, validate, and execute SQL:

START → query_gen → query_check → execute_query → nl_output → END

• Dynamic transitions:

  • If LLM returns an error → regenerate SQL.
  • If SQL is valid → execute query.
  • Convert query results into natural language for users.

---

Example Usage

1. Upload a CSV file as a table.

2. Ask questions like:

   "List the top 5 customers by total purchase amount."
   

3. Receive SQL-generated results in natural language:

   "The top 5 customers are Alice, Bob, Charlie, Dave, and Eve with respective totals of..."