AI can answer a lot of questions with varying levels of accuracy, it is even pretty good at creating SQL queries, but if it doesn’t know your database then it’s a very detached and incomplete experience. In this post I’ll discuss how I overcame this and set Claude.ai up so that it understands my database and can easily create SQL queries that you can immediately try for yourself.
This has another set of advantages as well which I will talk about later in this document.
As with any AI the way to success is all about context. You need to find a way of telling the AI as much as you can about your database or in my case, our ERP system.
You can of course construct a document that contains this information, which is a good idea and you can reuse the same document over again for future similar queries. This however becomes unwieldy if you want to ask for queries that you have not previously documented as context for the AI.
The database itself contains all the information about tables, views, indexes, foreign keys, constraints etc. that Claude will need so that it can write accurate queries for you.
Our ERP system has 1000+ tables and eleventy million other objects, there way no way I was going to even attempt to do anything manually. So I wrote a Python script and ran it against the database and the script generates a schema file containing everything the database knows about itself.
The image shows a small section of the schema file. Obviously for often used queries you can extract just the necessary information from this file and save it as a much smaller specific schema.
Now I’ve got this I just need to upload the schema file to Claude and ask it to generate a SQL query for me.
‘using the uploaded schema generate me a sql query to show me top customers by order value in 2026’
It will then give you some feedback before building the required query.
I mentioned that there are other advantages of showing Claude a complete picture of your database. Once Claude has that picture you can then ask it to explain how part of the database works, or ask it to create reference documents for you to use.
Below are a three screenshots of interactive HTML tools I asked Claude to generate from the uploaded schema:
Claude will not always be 100% accurate with SQL queries that it generates and it may make assumptions about things which are not obvious from the schema file and ideally needs access to the data itself.
I’ll go through this in the next post.