Perhaps not quite what you're asking for, but along the same lines there's this "Incomplete JSON" parser, which takes a string of JSON as it's coming out of an LLM and parses it into as much data as it can get. Useful for building streaming UI's, for instance it is used on https://rexipie.com quite extensively.
Work could be done to optimize it, for instance add streaming support. But the cycles consumed either way is minimal for LLM-output-length=constrained JSON.
Fun fact: as best I can tell, GPT-4 is entirely unable to synthesize code to accomplish this task. Perhaps that will change as this implementation is made public, I do not know.
If the LLM did such a bad job that the syntax is wrong, do you really trust the data inside?
Forgiving parsers/lexers are common in language compilers for languages like rust or C# or typescript, you may want to investigate typescript in particular since it's applicable to JSON syntax. Maybe you could repurpose their parser.
I feel like trying to infer valid JSON from invalid JSON is a recipe for garbage. You’d probably be better off doing a second pass with the “JSON” through the LLM but, as the sibling commenter said, at this point even the good JSON may be garbage …
I've been looking into something similar for handling partial JSONs, where you only have the first n chars of a JSON. This is common with LLM with streamed outputs aimed at reducing latency. If one knows the JSON schema ahead, then one can start processing these first fields before the remaining data has fully loaded. If you have to wait for the whole thing to load there is little point in streaming.
Was looking for a library that could do this parsing.
https://gist.github.com/JacksonKearl/6778c02bf85495d1e39291c...
Some example test cases:
Work could be done to optimize it, for instance add streaming support. But the cycles consumed either way is minimal for LLM-output-length=constrained JSON.Fun fact: as best I can tell, GPT-4 is entirely unable to synthesize code to accomplish this task. Perhaps that will change as this implementation is made public, I do not know.