Talk to Your Database: AI-Powered Natural Language Queries in 10 mins

Feb 01, 2025

Why It Works

🔹 Human-to-database chat interface

🔹 Schema-aware AI avoids guesswork

🔹 One-time setup, endless queries

Real-World Use Cases

Marketing team pulling campaign metrics via Slack
Executives accessing real-time KPIs in plain English
Developers debugging without direct DB access
Support team quickly looking up additional relevant information

The last one just happened at our team at the moment of me writing it.

Prerequisites

Database credentials:

Host address
Port number
Username & password
Database name

2. LLM access (Claude, OpenAI GPT, local Ollama or similar)

3. Basic terminal

Step-by-Step Implementation

1. Extract Your Database Schema

Sample LLM prompt:

Generate a command for SingleStore to export schema without data
using these credentials:
`host=db.example.com port=3306 user=admin dbname=analytics`.

Provide instructions to install the required tools
and execute the command in the Terminal app.

(Replace “SingleStore” with your database type and remember to insert your actual credentials.)

In the resulting message, the LLM will provide you with clear instructions on how to prepare and execute the command, along with the command itself.

Proceed to execute the command and enter your password when prompted. This will create a `.sql` file in your working directory.

2. Create Python Query Script

Upload or point LLM to the file (if working locally) and ask:

Using the database schema and credentials you already have,
build a Python script to transform my natural language requests
into relevant database queries and show me the query results.

Implement rate limiting on queries and add a validation layer for SQL commands.

Test the connection and test the flow with a sample request.

The AI will do its work (go through corrections if needed), and you will get the `.py` script along with other needed files (like `config.yaml` for storing your credentials).

3. Use Your AI Assistant

With the necessary files in hand, create an AI agent using OpenAI’s custom GPT feature or local tools like Goose with operating system control capabilities. Alternatively, you can simply save this prompt and send it to the LLM (along with the files) each time you need to find information within your database stack:

Here are the files needed to interact with my database
and query it based on my natural language requests to you,
and provide me with answers.

Analyze the files and get ready to receive a request from me.

Never under any circumstances execute a write operation on a database
without prior confirmation. Always confirm write operations first.
Read commands are okay to execute without confirmation.

(The last paragraph is kind of important, you might guess why.)

Your assistant is ready at this point \(^ヮ^)/

4. Improve your workflow

As you use your agent, ask it to remember which queries it used and which tables it looked into to find information for you in a particular use case. This way, you will gradually lower the number of queries and the time to get a result.

Safety measures

Better to start with non-critical databases
Store your database credentials securely
Provide detailed instructions to the LLM when necessary

For example, if you have different types of user data in your database, such as internal users and external customers, tell the AI which users you want it to search for.

I am a VP of Engineering, @Telegram analytics lead. Ex @Yandex, @Mapbox.

Everything: https://arbatov.dev / X: https://x.com/vladzima

Talk to Your Database: AI-Powered Natural Language Queries in 10 mins was originally published in Vlad Arbatov — The Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

Vlad Arbatov