DT's OCR and AI/LLM?

[Blake]

Hi Devonthinkers – A bunch of papers on the Internet say that it should now be possible to harness Large Language Models to vastly improve OCR. (Even when I simply I ask Google’s Bard to “correct” a typically faulty text yielded by DT’s ABBYY reader, it easily gets rid of ALL of its errors, without even needing to see the original image that was OCR’d. I guess it substitutes statistically likely readings for the statistically impossible ones in the OCR’d text.)
Does anyone know if there’s a way – or soon to be a way – to implement/automate AI correction of documents OCR’d within DT? It seems that ABBYY has some such pairing of AI and OCR, called “ABBYY Vantage” but, as a non-specialist (an art historian, in fact) I haven’t been able to tell if it’s what I need.
Any thoughts much appreciated. I OCR a lot of historical documents, and then use DT to do Boolean searches of their contents, but the sheer number of errors in the OCR almost guarantee that my searches miss a lot of relevant material.

[chrillek]

@Blake

Does anyone know if there’s a way – or soon to be a way – to implement/automate AI correction of documents OCR’d within DT

I don’t, but did you try putting your text through LanguageTool (which is free for up to 3000 characters)? I find that quite helpful, though it doesn’t (fortunately) autocorrect. And I have no idea how that works with old languages.

@Blake

It seems that ABBYY has some such pairing of AI and OCR, called “ABBYY Vantage”

That looks rather business-oriented, and you have to train it. I guess your texts are really rather about the probability of one word following another one. The Vantage stuff has to consider different regions of text on a page.

[Blake]

Thanks. My texts are all in modern English, in fact, so language isn’t a problem. Just looking for basic text-recognition, which is surprisingly hard for letter-by-letter (non-contextual, non-semantic) OCR to get right. I think GPT-4 might correct uploaded texts even better than LanguageTool – but feeding each of my thousands of PDFs from DT into GPT would be too laborious to be practical. Hence my question about somehow automating such a process into DT. (I’m so tech-ignorant that I’m afraid I have no idea what that would involve, or if it’s just an absurd ask of a database program.)

[chrillek]

@Blake

feeding each of my thousands of PDFs from DT into GPT would be too laborious to be practica

OTOH, ChatGPT/OpenAI would certainly be grateful for another bunch of free training material.

@Blake

Hence my question about somehow automating such a process into DT.

I seem to remember that @rkaplan published a script here – did you search the forum?

[rkaplan]

It was here

OpenAI Integration with DT3 through JXA Script

Below and attached are Version 1.0 of a JXA script integrating OpenAI with Devonthink. As we have discussed extensively on the forum, AI has tremendous potential to help with document analysis that many of us are involved with. ChatGPT/OpenAI and this script are in early stages. Do not rely on the software without verifying the facts yourself. Above all do not upload any private/privileged documents to OpenAI. That warning aside, I am quite intrigued by what OpenAI is presently capable of …