Any AI tool specifically good at generating JSONata queries?

[moksh-1]

Hi,
We’re currently using OpenAI to generate JSONata queries based on input/output JSON structures. It works for simple cases, but struggles with complex queries or sometimes generates invalid syntax which leads to multiple retries and higher costs.

Do you know of any AI tool, model, or method that works better or more reliably for generating complex and correct JSONata queries?
Any suggestions or best practices would be appreciated!

Thanks!

[markmelville]

I would suggest a tool specifically geared for developers, not just general chat. so Claude Code, Cline, etc.. there are so many any I don't even know all the latest. Because of my FTJ, I have had mainly used Amazon Q and Google Gemini. I just fired up Gemini in the browser and gave it a real example of a JSONata query that had me stumped last week, and it did OK.

I made suggestions changes to the expression it produced, based on what I know, and it was straight up wrong on these two points:

• it thought the else was required in a ternary expression
• it thought that undefined & "string" would be an illegal operation
  When I said it was wrong, it was like "let me go re-read the docs more carefully.. oh yeah I see where I got confused."

I then gave it a scenario where a filter produced only one item in the sequence and so it was not in an array, and I asked "why did it do that?", and it was able to actually explain sequence unwrapping, and that the expression need to be wrapped in [] to keep it in an array. I told it it should have put that in the expression.

Because of the mistakes it made and failing to identify the nuance of sequence unwrapping (which in my experience is a common stumbling block for adopters), I asked it if Gemini will learn anything from my corrections. It did say that engineers may review some interactions to see how to improve the model. So, eventually, Gemini or any other tool that we use to help us with JSONata will get better at it.

This leads me to some practices I would suggest when working with AI.

• Don't just ask something and either accept or reject its answers. Follow up with your insights and suggestions. Ask it why it gave the answer that it did. Gemini actually has a section you can expand to let you see what the AI is "thinking", which is very insightful.
• Ask it how you could have prompted better to get the result you wanted.
• In your initial prompt, set the stage: tell it to verify things when it's unsure, instead of making things up. Tell it not to be a sycophant but to question and challenge what you say with evidence.