Help understanding how to use the AI feature in DEVONthink

Dear community,

I’m looking for a video tutorial or any other resources that can help me better understand how to use the AI feature in DEVONthink.

I’ve read the built-in Help section about AI and have connected my ChatGPT account, but I’m still struggling to use the feature effectively. A few specific issues I’ve encountered:
• Sometimes, when I select a document and open a chat tab in the right pane, I receive a message saying nothing is selected or that I need to upload something first — even though a document is selected. Other times, it works as expected.
• I’m unclear on the difference between the “main” chat window and the one in the inspector pane. When should I use each one?
• Lastly, although I’ve selected the database as the context source in the AI settings, I can’t seem to perform contextual searches within a database, group, or tag. For example, I’d like to ask the AI to help me find a specific recipe in my PDF documents (e.g., in the selected group / tag, db), but I haven’t been able to get that to work.

I’d really appreciate any tips, clarifications, or links to helpful tutorials. Again, I’ve read the help page several times — but I still find the practical use a bit confusing.

Thanks in advance!

1 Like

I don’t know if I can help but I always include “the document” in the prompt. Sometimes it is a little unclear to me what “the document” refers to. For example, if I follow some links in the document and the linked document is shown in the main window. What is now “the document”? The one I started with or the shown document? I think it is the document I started with. At least that is the info that is shown in the right pane.

1 Like

interesting. including “document” in the prompt in the chat window in the inspector / right pane made it better:

this time, I was able to get a meaningful response.

Q: why is it necessary to include document? is it not implied by the currently selected document…?

thanks anyways, that’s already a progress for me! still struggling with the other questions, though

1 Like

Did you read the Getting Started > AI Explained and AI in Practice section in the Help?
Or Tasks > Leveraging AI?
Or Documents > AI and Your Documents?

1 Like

I’ve read “Inspectors → AI”.

the one you have mentioned I didn’t see. thanks, I’ll take a closer look!

I’ve went through all the links you have sent. Thanks. I’ve now discovered a few new features, e.g., summarize via chat or renaming document names.

However, I still have the issue that the Chat window in the Inspector / right pane does not deliver the results consistently. Sometimes it works with the document selected, but more often (90% of the time) it tells me to provide the contents…

here is an example:

I have selected a text file. and ask to translate it in the chat window. but it does not work.

You are only using the GPT 4.1 Nano model. As noted in the help, there isn’t one model that’s suitable for all cases, so you should be trying different models to see what works for the specific situation. I generally use a mid-tier model, like Claude 3.7 Sonnet, for these kinds of chats.

I now tried Claude 3.7 with the same command (translate into German) with a selected document and it worked.

but! it replaced the contents of the document, and now I’ve lost the original text :smile: is there a way to restore that…?

also, I’m surprised that the same prompt sometimes works and sometimes does not work with ChatGPT…

AI is not a static repository of information, like an encyclopedia. Not only will responses vary between AI engines, e.g., ChatGPT and Claude, but the same prompt to the same model will often elicit different replies.

it replaced the contents of the document, and now I’ve lost the original text

What was your prompt?

the prompt was “translate this document into German”.

it replaced the contents of the file with the German translation (which is fine, but I’d prefer to keep the original text…). Unfortunately, the versioning is not enabled for this db (this was also new to me, I thought it’s on by default).

Okay… now logically consider what you asked it to do, because it did exactly that.

Unfortunately, the versioning is not enabled for this db (this was also new to me, I thought it’s on by default).

No, versioning is certainly not enabled by default. Just as sync and AI are optional, so is versioning. We believe in opt-in for more than opt-out.

well, my expectation was that it would place the translation (“output”) into the chat window. and not change the document

1 Like

While I get where you’re coming from, that’s not what you asked it to do. It’s just something you expected but didn’t make clear. Remember these mathematical acrobatics are interesting and impressive, but it’s still a computer you’re dealing with. You need to think about the specifics of what you’re asking or commanding, not assuming it will know.

In fact, here is a variation of your query. And note it may look similar but it is not the same thing you asked…

And the AI even suggested saving it as a separate document. I could have included that in the prompt as well.

And here is a more specific prompt…

4 Likes

yes, I see.

can you, perhaps, suggest a better prompt…? without replacing the original text automatically?

great, thanks!

I’ve learned a few things today, that was helpful!

I’m impressed by the AI features, but need to play more with them, but already see the potential.

thanks a lot for the great work!

You’re very welcome!
In the help, I tried to provide more practical examples (and information on a variety of topics), but it can’t be exhaustive. Part of using AI will be learning how to talk to it.

2 Likes

There is a setting in DT4 where you can allow or deny the LLMs to modify your files. You have it enabled?

Repeating the same prompt on the same document brings different outputs. As you say, we are not querying a database. The output is, if I understand correctly, somewhat stochastic.

1 Like

yes, it’s enabled. tbh, I thought it would ask me to confirm or something, before actually changing the file :slight_smile: