This should likely have a disclaimer that JSON you paste into the tool can be stored. I was intentionally providing bad input and was presented with a message, “Sorry, I couldn't parse this JSON. I've logged it and will try to add it to the list of use-cases. Please let me know if this error persists.”
Seems useful as I do occasionally get broken JSON from LLM responses and would like an easy solution to that. However, I don't see this as something I would use an API for. Especially seeing the supported fixes are relatively simple. Seems like it should be a python package. Maybe if you could somehow handle a comparison with a pre-specified json schema that would make it more valuable.
I have a simpler thing I do with a wrapper around jsonlint. Just pass a filename and it opens my editor (Sublime) right to the line and column of the error.
Very cool! One note: the random sample marked: 'Using a string "true" instead of a boolean true' leaves "true" as a string after parsing, instead of converting to an actual boolean.
Just imagine if we didn’t use heuristics to fix arbitrary inputs and instead we used some sort of learning algorithm that was trained on producing valid looking json - then all we’d need to do is add a prompt and start with some random noise… :P
