This is not the issue I’ve brought up though, and yet it is still directly related to the OCR process, just not the “recognition” part of the process. ABBYY is correctly identifying virtually all individual characters, so in that sense, optical character recognition is “working” at the level of each individual character. It’s even recognizing the column order correctly in many cases, and where it doesn’t, other solutions fail as well.
The issue I’m reporting is that the way the layout is being sequentially written to the invisible layer is lacking compatibility outside of DEVONthink. Something about those instructions is not readable by other readers, and results in a jumbled text sequence. This is the issue that is unique to ABBYY’s implementation within DEVONthink.
If writing what has been recognized by ABBYY to the invisible layer (line/layout sequence instructions, in particular) is not part of the OCR process (I have assumed ABBYY handles it from start to finish), then this would only reinforce my point that there is an issue with DEVONthink. More likely though, the implementation being used has a bug that needs to be worked out with ABBYY. Either way, the fact that it used to work without this issue suggests that something broke, and I’ve identified when that most likely was.
Good to know. That’s probably where it’s happening then. Hopefully they can fix it then since it used to work better when I compare earlier OCR compatibility with later ones. Is DEVONtechnologies reaching out to them, or is this a matter of passively hoping they might notice and fix it in the unknown future?
Excellent, thanks. I wasn’t trying to prod. It just wasn’t clear to me that this issue is even being recognized. I love what DEVONthink does and has to offer, and simply hope for the best with it.
Unfortunately not as this (and other reported issues) completely depends on Abbyy. But OCR has become less important in DEVONthink 4 as the Pro/Server editions make every imported PDF document automatically searchable (using macOS’ Vision framework) and support also macOS’ live text.
Last time I checked (about two months ago), Vision wasn’t perfect, either. It does occasionally mix up the sequence of words on the same logical line. So “searchable”, yes.
That’s a shame. I hope they’ve been notified of the issue, as it does not occur when using their dedicated software.
MacOS vision isn’t very competitive for historical documents with less-than-perfect scans.
Would you guys consider allowing us to bring an API key from MistralOCR as a working alternative?
Yeah, exactly. Which is a pity that they’re sharing an inferior solution with you guys! I hope they discount whatever contractual agreements they have with you for providing you an increasingly nerfed product since the deterioration began.