Does this work well on kvps and tables? That is where I typically have the most trouble with tesseract and where the cloud provider ocr systems really shine.
tl;dr NER models work best when they process data similar in structure to their training data. This notebook shows a few examples of how one can use an NER model not trained on JSON to actually identify and redact both structured and unstructured values with in a JSON blog. The NER model used is free available but closed source, but one can expect similar results with open source NER models such as spacy.