DeepSeek-OCR, Tesseract miss PDF JSON

The hot take is that there isn’t a lone SOTA OCR model that cleanly solves scanned-PDF-to-JSON under tight cost limits. The practical win is a hybrid stack: OCR for text, layout/table extraction for structure, then a schema-aware JSON normalization step.

• –Tesseract is still a baseline, but it struggles on tables unless you add custom layout segmentation and post-processing.
• –OCRmyPDF is useful as a preprocessing step because it adds a searchable text layer, but it does not solve structured extraction by itself.
• –For richer structure, local tools like Docling or PaddleOCR-style pipelines are a better fit because they target reading order, tables, and document layout.
• –If the end goal is JSON, the last step should be deterministic schema mapping or a local LLM pass over extracted blocks, not raw OCR output.
• –DeepSeek-OCR is interesting, but the complaint here is the common one: OCR quality alone is not the same as document parsing quality.