Most production software is wrappers around existing libraries. The relevant question is whether this wrapper adds operational or usability value, not whether it reimplements OCR. If there are architectural or reliability concerns, it’d be more useful to call those out directly.
I fought with Tesseract for quite a while. Its good if high accuracy doesn't matter. Transcribing a book from clean, consistent non-skewed data its fine and an LLM might even be able to clean it up. But for legal or accounting data from hand scanned documents, the error rate made it untenable. Even clean, scanned documents of the same category have all sorts of density and skew anomalies that get misinterpreted. You'll pull your hair out trying to account for edge cases and never get the results you need even with numerous adjustments and model retraining on errors.
Flash 2.5 or 3 with thinking gave the best results.
I have a flow where i extract text from a pdf with pdf-parse and then feed that to an ai for data extraction. If that fails i convert it to a png and send the image for data extraction. This works very well and would presumably be far cheaper as i'm generally sending text to the model instead of relying on images. Isn't just sending the images for ocr significantly more expensive?
I have seen this flow in what people in some startups call "Agentic OCR", its essentially a control flow that is coded that tries pdf-parse first or a similar non expensive approach, and if it fails a threshold then use screenshot to text extraction.
By definition, OCR means optical character recognition. It depends on the contents of the PDF what kind of extraction methodology can work. Often some available PDFs are just scans of printed documents or handwritten notes. If machine readable text is available your approach is great.
What matters most is how well OCR and structured data extraction tools handle documents with high variation at production scale. In real workflows like accounting, every invoice, purchase order, or contract can look different. The extraction system must still work reliably across these variations with minimal ongoing tweaks.
Equally important is how easily you can build a human-in-the-loop review layer on top of the tool. This is needed not only to improve accuracy, but also for compliance—especially in regulated industries like insurance.
The "guts" are here: https://github.com/majcheradam/ocrbase/blob/7706ef79493c47e8...
Do people actually store their secrets in plain text on the file system in production environments? Just seems a bit wild to me.
https://github.com/tesseract-ocr
Flash 2.5 or 3 with thinking gave the best results.
Discussion is here: https://news.ycombinator.com/item?id=45652952
https://github.com/rednote-hilab/dots.ocr
Equally important is how easily you can build a human-in-the-loop review layer on top of the tool. This is needed not only to improve accuracy, but also for compliance—especially in regulated industries like insurance.
Other tools in this space:
LLMWhisperer/Unstract(AGPL)
Reducto
Extend Ai
LLamaparse
Docling