How I made an AI-powered query language with ChatGPT

So I had recently picked up ChatGPT prompt engineering, and I wanted to make something sort of practical. So I thought of making something like GraphQL, but with natural language. So I whipped up the OpenAI playground, and started writing a prompt in the format that I usually use. It took a long while to get it to do what I wanted, but I eventually got it to work more than half the time.

How the prompt works

The prompt works by taking in 2 things. The natural language query and your API schema. (Your API schema should also be natural language) I took advantage of the system message by appending the schema to the system message, right under the the section labeled “Schema”

I instructed the AI to respond with this JSON format:

{
    "headers": { node-fetch-style headers here },
    "body (optional)": { body here },
    "url": "[api endpoint]",
    "method": "POST",
    "pipe (OPTIONAL)": 
    { 
        "url": "url to pipe to",
        "method": "request method",
        "headers": { node-fetch-style headers here },
        "pipeTo" : "pipe items from body to body, here is an example of syntax: "message=>username;status=>password""
    }
}

This, with the exception of the pipe key allows me to just plug in the values into a fetch request. I also added the ability to pipe the response of one request to another request, but it’s quite buggy.

Finally, to coax the AI into responding with just JSON, I added a small example conversation I had with the AI and sent that with the system message.

The UI

I made a simple UI in figma but I didn’t like it so I asked my friend @bddy to improve it. He did a really awesome job with the UI and you should check him out! I also hooked up the UI to the backend for him so he wouldn’t have to do that lol.

Demo