LLM-Powered Database Assistant

An agentic LLM-powered assistant that converts natural language into SQL queries, executes them safely, and returns conversational answers from a PostgreSQL database.

---

Project Overview

This project implements a Text-to-SQL system with tool-augmented reasoning, enabling users to interact with a database using plain English.

Features

• Natural language → SQL query generation
• Safe SQL execution (read-only)
• Conversational response synthesis
• Dynamic schema awareness
• Modular workflow using LangGraph

---

Architecture Overview

---

Workflow

flowchart LR
    A[User Input] --> B[SQL Generator LLM]
    B --> C[DB Executor]
    C --> D[Result Synthesizer LLM]
    D --> E[Final Answer]

    C --> F[(PostgreSQL DB)]
    B --> F

Loading

---

Approach

Agentic Workflow (LangGraph-based)

This system follows a structured agentic pipeline:

• Text-to-SQL Agent Generates SQL queries from user input

• Tool Augmentation Uses:

  • Database schema inspection
  • SQL execution tools

---

Components

1. SQL Generator

• Converts user queries into SQL
• Uses schema context for accuracy

2. DB Executor

• Executes SQL queries
• Restricts to SELECT only

3. Schema Inspector

• Retrieves:
  • Tables
  • Columns
  • Relationships

4. Result Synthesizer

• Converts SQL results into conversational responses

5. Workflow Engine

• Managed via LangGraph
• Controls execution flow and state

---

Stack

• LLM: LLaMA 3.2 (via Ollama)
• Orchestration: LangGraph
• Framework: LangChain
• Database: PostgreSQL
• ORM: SQLAlchemy

---

Safety Measures

• Only allows SELECT queries (read-only)
• Blocks INSERT / UPDATE / DELETE
• Schema-aware prompting to reduce hallucinations

---

Installation

# Clone the repo
git clone <your-repo-url>
cd <your-repo>

# Install dependencies
pip install -r requirements.txt

---

Environment Variables

Create a .env file:

DB_USER=your_user
DB_PW=your_password
DB_HOST=localhost
DB_PORT=5432
DB_NAME=your_db

---

Running the App

python main.py

---

Example Usage

You: Show me all customer accounts with balances above 10,000
AI: Here are the accounts with balances above 10,000...

---

Project Classification

This project can be described as:

• Agentic LLM Workflow
• Tool-Augmented LLM System
• Text-to-SQL Application

---

Future Improvements

• Add memory (conversation history)
• Add retry & error handling
• Add query validation layer
• Integrate vector DB for hybrid retrieval
• Add evaluation benchmarks

---