Yeah, but that could require multiple queries, which isn't very efficient. Training model just to fix JSON would be better.