Tags behavior? (DEVONthink and Chat)

Hi everyone,

I have a few questions about the tagging on import behavior:

What happens if both “Assign existing tags” and “Add Chat …” are selected? Does DEVONthink use its own classification logic to add tags, and then Chat makes an AI API call? What if the tags suggested by DEVONthink and Chat (“AI API”) are the same? How many tags does Chat assign…?

thanks!

I’ve also found, that the rules to not create new tags are not being respected.

on import, lots of irrelevant (and new) tags are being created

What do you import, how, and what are your settings?

I’ve tried on PDFs, images (e.g., PNG).

here are my settings:

upon import, I get a lot of inaccurate tags (which previously did not exist in my db)

But you’ve set your prefs to import finder tags and convert hashtags to tags as well as assigning existing tags. All these settings might well cause new tags to be created, I guess.

“Convert Hashtags to tags” — does it mean, that if a document includes any hashtags, those will be automatically added?

as for finder tags — not using these.

Ich lese mal das Handbuch vor. Das ist die verbindliche Quelle für solche Fragen.

Hashtags: Popularized in social media, hashtags are a type of tag prefixed by a # . If enabled in the Tags settings, DEVONthink can detect hashtags in the contents of text-based files and convert them to tags for the file automatically. Removing the hashtags from the content will remove the tags from the file. Hashtags do not support spaces in them and they should be listed on a separate line. They will not be detected within existing paragraphs. Of special note, hashtags supports creating nested tags as noted here. The first tag just needs to be prefixed with the # to be detected as a hashtag..

okay, this is not the case, since I was importing PDFs and images (screenshots) as well…

still got a bunch of not so helpful tags…

Yes, both can be enabled and they act independently. Redundant tags aren’t made if Chat suggests an existing one.

Also, the tags suggested by Chat will vary depending on the AI engine and model you’re using, e.g.,…

  • Gemini 2.5 Flash: Strawberry, Ice cream, Cartoon, Bee, Cute
  • GPT 4.1 mini: cartoon, ice cream, cute
  • Mistral Medium 3: cartoon, animation, 3D, character, food
  • Claude Sonnet 3.7: Animation, Cartoon bee, Strawberry, Ice cream, Digital art

Chat suggestions honor existing tags in import, if the option is selected…

so I have turned off the chat suggestions tags:

I have then imported “stoic-mantra” image into DEVONthink:

DEVONthink has assigned the following tags to this image:

tags: furniture, shopping, tableware: these are the tags I was using recently today for other items!

I’m starting to think, there is something off in DT, because earlier today, I was assigning tags to my imported bookmarks using chat (“AI”). and while the tags assigned by the chat were mostly ok, there were always off by one entry.

That is, entry 2 had tags which were more appropriate to entry 1; entry 3 had tags for entry 2 and so on and on…

could this be a bug?

update: I have now imported random screenshot into DT, and it also got assigned the same tags: furniture, shopping, tableware…

I’m not sure why this is happening

If you’re really importing a PNG, it does not have any textual information. What do you suggest DT uses as tags in this case if you ask it to assign existing tags?

The question should not be really addressed to me, since I don’t know how DEVONthink works under the hood.

Instead: why DT assigns some garbage tags if there no textual information?

btw, I just imported DEVONthink’s own manual (PDF) and here are the tags I got:

I mean, most of the tags are almost correct. But the choice of “fashion” is interesting

Make sure your backups are current and do a File > Rebuild Database on the problematic database.

Also, what are your Files > Import settings?

Here are the settings:

Q: what makes you think this database is a problematic one…?

:thinking:

I’m starting to think, there is something off in DT, because earlier today, I was assigning tags to my imported bookmarks using chat (“AI”). and while the tags assigned by the chat were mostly ok, there were always off by one entry.

That is, entry 2 had tags which were more appropriate to entry 1; entry 3 had tags for entry 2 and so on and on…

If it were me and the tags were relevant to me, I’d assign them manually. Or with an automation that I control. Leaving it to some unspecified algorithm doesn’t make much sense to me. Especially not if the data provides no information at all. As is the case with images.

3 Likes

Thanks, that’s the approach I have settled after all…

The database works fine, I probably won’t be rebuilding it any time soon. But will be assigning tags manually

Please zip the file that you imported and upload it or send it to cgrunenberg - at - devon-technologies.com. It’s unlikely that it’s a bug so far but I would like to check this.

Finally, did you verify that the file has no Finder tags and that none of your smart rules could cause this?

I turned it off as DT did assign existing tags, but I got far more than I wanted. For example; it would assign the tag “news” if the document contained the word ‘new’ anywhere.

I prefer to assign tags manually or by using a script that assigns tags based a document containing a few key words like ‘meeting’ to tag it with “meeting”.

2 Likes