---
title: Autonomous SQL Agent
emoji: 💬
colorFrom: yellow
colorTo: purple
sdk: gradio
sdk_version: 5.42.0
app_file: app.py
pinned: false
license: mit
short_description: 'An autonomous SQL agent, based on Qwen 2.5 (fine-tuned). '
hf_oauth: true
hf_oauth_scopes:
- inference-api
models:
- manuelaschrittwieser/Qwen2.5-1.5B-SQL-Assistant-Prod
- Qwen/Qwen2.5-1.5B-Instruct
tags:
- agent
- sql
- text-to-sql
- qwen
- qlora
---

# Autonomous SQL Assistant Agent

## 📋 System Overview

The **Autonomous SQL Assistant** is a demonstrative AI agent designed to bridge the gap between natural language inquiries and database execution. Unlike standard "Text-to-SQL" generators that strictly output code, this agent operates within a closed-loop environment: it **generates** syntax, **executes** it against a live database, and **retrieves** the actual data for the user.

The system is powered by **Qwen 2.5 (1.5B)**, fine-tuned via **QLoRA** on the `b-mc2/sql-create-context` dataset to ensure high fidelity in SQL syntax generation.

**[🔗 View Source Code & Documentation](https://github.com/MANU-de/Autonomous-SQL-Agent)**

---

## 🏗️ Technical Architecture

The application runs on a lightweight CPU environment and consists of three core components:

### 1. The Inference Engine
*   **Model:** [manuelaschrittwieser/Qwen2.5-1.5B-SQL-Assistant-Prod](https://huggingface.co/manuelaschrittwieser/Qwen2.5-SQL-Assistant-Prod)
*   **Optimization:** The model runs in full FP32 precision (CPU optimized).
*   **Role:** Translates user intent (e.g., *"Who earns the most?"*) into executable SQLite syntax, utilizing the provided schema context.

### 2. The Execution Sandbox
*   **Database:** A transient **SQLite** instance.
*   **Schema:** `employees` (id, name, department, salary, hire_date).
*   **Lifecycle:** The database is re-instantiated upon every application restart/build to ensure a clean state for testing.

### 3. The Agent Logic
The `SQLAgent` class orchestrates the workflow:
1.  **Ingest:** Receives natural language prompt.
2.  **Contextualize:** Injects the `CREATE TABLE` schema into the system prompt.
3.  **Generate:** produces the SQL query.
4.  **Act:** Connects to the SQLite cursor, executes the query, and fetches results.
5.  **Sanitize:** Catches execution errors (e.g., syntax errors) and reports them for debugging.

---

## 💻 Usage Instructions

### Interface Guide
The interface is a chat-based UI. You act as the user querying the HR database.

*   **Input:** Type natural language questions regarding the `employees` table.
*   **Output:** The agent provides a two-part response:
    1.  **"Brain" (Internal Monologue):** The generated SQL query.
    2.  **"Result" (Data):** The raw tuples returned from the database.

### Example Queries
Try copying these prompts to test the agent's capabilities:

| Complexity | Query |
| :--- | :--- |
| **Simple** | *Show me the names of all employees in Sales.* |
| **Conditional** | *Who earns more than 60000?* |
| **Aggregation** | *Count how many employees work in the Engineering department.* |
| **Logic** | *List employees hired after 2020.* |

---

## ⚙️ Local Reproduction

To run this Space locally on your machine (requires Python 3.10+):

1.  **Clone the Repository:**
    ```bash
    git clone https://huggingface.co/spaces/manuelaschrittwieser/sql-assistant-prod
    cd sql-assistant-prod
    ```

2.  **Install Dependencies:**
    ```bash
    pip install -r requirements.txt
    ```

3.  **Launch Application:**
    ```bash
    python app.py
    ```

---

## ⚠️ Limitations & Scope

*   **Inference Latency:** As this demo runs on **CPU Basic** hardware, generating the SQL query may take 2-10 seconds depending on server load.
*   **Sandbox Restrictions:** Database modifications (INSERT/DROP) are possible but will persist only until the application restarts.
*   **Hallucinations:** While fine-tuned, the model may occasionally generate invalid SQL for highly complex queries not covered in the training distribution.

---

## 📜 License

This project is open-source and available under the **MIT License**.