Text Recognition via Chat repeatedly transcribes first page for multi-page PDFs

DEVONthink Feedback
,
1

As someone with an immense library of handwritten PDFs, DEVONthink 4’s LLM-based PDF transcription has the promise of being more useful and much more robust than the the rube-goldberg automations I have been relying on.
Oddly, though, it presently seems to repeatedly transcribe the first page of the multi-page PDFs. Here, there are three pages, and the transcribed comment has 3 slightly different renditions of the first page –

CleanShot 2025-06-09 at 16.40.37@2x
CleanShot 2025-06-09 at 16.40.37@2x1920×1241 251 KB

This happens with both Claude 4 and GPT 4.1, using the latest DEVONthink 4.0beta3.
Let me know if I can provide other troubleshooting details.

Also, it would be lovely to customize the prompt used for transcription, e.g. to insert markdown formatting or with instructions on how to handle equations and images.

With congratulations on the most thoughtful integration of LLMs to personal knowledge management I’ve seen,
-Wug

1 Like
2

Thank you for the bug report! Does this always happen or affect only certain PDF documents? Which app did you use to create them?

3

I was able to reproduce the issue, the next release will fix this.

1 Like
4

Welcome @professorwug
Thanks for the report and very kind words of encouragement. They’re very appreciated.

5

You could use a Chat - Query action of smart rules & batch processing and the option With Document (Vision).

1 Like