Will we have a DT 4 in the near future?

I like dt3 a lot, but I still use obsidian for writing and using plugins like excalidraw to sort out my ideas… I often combine dt with obsidian. If only DT had better editing skills. At the same time,The LLM gpt has improved a lot in the last two years. Will we see a dt4 that incorporates the large language model? In the near future?

1 Like

See DEVONtechnologies | Our Business Principles


Welcome @harry.yanghao

If only DT had better editing skills.

“better” is a matter of opinion. Many people, including myself, write in DEVONthink all the time. Please provide more concrete information on what you feel is lacking.

At the same time,The LLM gpt has improved a lot in the last two years. Will we see a dt4 that incorporates the large language model?

This has been discussed at length on these forums. DEVONthink will not incorporate a LLM as AI on a local machine is slow, mediocre in results, or both. We are looking at potentially allowing use with online models, but we consider performance and most importantly privacy to be the highest priority.


If I may, there are many little Apps that provide inline LLM.
You connect the App to the API of your model of choice (I do with GPT4), then you can trigger the inline feature via a keyboard shortcut.

You can check Elephas (maybe the most complete? Also, it’s on SetApp) or some of its alternative.
If you feel geek enough, you can also check LibreChat, but I’m not sure about an inline feature.

Hope it helps.

As for an upgrade of DT, I really hope for a better support for themes, as like as Obsidian.
Ok we can tinker with CSS, but it’s not the same.

1 Like

DEVONthink is not HTML/CSS-based and therefore there are currently no such plans, I’m sorry.


Do you know of any GPT-4 app with an API that can summarize or query large PDF files?

I know of no way to do that through an API without writing a custom app with LangChain or similar. Am I missing something?

You may have a look at PDF Pals, that is under SetApp too.
You upload the PDF, the App indexes it (I think it indexes the PDF local ,but I’m not sure), then you can query the PDF via a chat on the side of the App.
You link the App to the OpenAI API.

Elephas does the same, though.
Maybe there are alternatives I’m not aware of.

Hope it helps.


Both of those fail almost entirely with any significant length PDF

The only apps I have seen do a good job store your files in the cloud rather than locally

Just wondering but how many words do these files contain? GPT-4 Turbo has a 128k context window, that’s sufficient for smaller books already.

Typical for the files I am using is 150 Mb, 1500 pages, 350,000 words

There are some ChatGPT Custom GPTs using GPT-4 which can currently do a stunningly helpful job summarizing or querying such a file as long as I am willing to upload the file to the web. The files are larger than the context window, but they use chunking or other algorithms to deal with that; it has limitations of course and all output needs to be reviewed, but it is extraordinarily useful even at the level of professional document review.

The context window number is a bit misleading because it includes both input and output. Typically the goal in analyzing a large document is a detailed summary or retrieving a large number of responses in context. So the context window needs to be much larger than just the document that you are analyzing. Also a number of the dev sites say that the AI models slow down when over 50% of the context window is used and I have noted that as well.

Thus even with a 128K context window it is necessary to use algorithms to iterate multiple AI queries or index documents or chunk documents or otherwise manipulate the data. The apps which handle AI locally do not seem optimized for that function. I can get a stunning result from a query of a large PDF as above using a cloud-based custom GPT in under a minute, whereas PDF Pals locally on my Mac Pro says “Indexing - wait 10 minutes” and in fact takes way more than 10 minutes to prep the file.

If you want to compare the performance, try PopAI

or try AIDrive


Both have stunning performance. But hosting files in the cloud can be a privacy/security issue for some applications.

Hence I think the best solution at present is to use Langchain to build a custom app where the documents either are stored locally or are stored in a trusted cloud location I control such as Dropbox.

what about AI based-translation? . I read foreign scientific documents sometimes, and it will be very helpful if DT can translate the documents using LLM.
I paid a translation tool “https://immersivetranslate.com” for foreign websites, but not good to use for pdf documents.

What about DeepL? I think you can feed it the text of PDFs


Interesting that Obsidian more and more comes up with people working with both Obsidian and DTP. A while ago, just as ChatGPT was becoming a thing, I wrote a comparison:


By now, like many people, I’ve GPT running using plugins, like SmartConnections, all the time.

I think what we’re really observing is Kristensen’s “Disruptive Innovation.” We’ve an incumbent, DTP, that’s fast, established, does exactly what many existing customers want it to do. Hence it is reasonable both for those customers as well as for the providers to play it conservative and keep delivering what they’re delivering. And boy are they resistant to even the slightest change (just “strike out view in the trash” comes to mind).

I guess we’ll see what we always see there: The moment that the newcomer is “just good enough”, the market will flip over. Has happened just so many times, and every time it’s the same pattern. Obsidian is clearly the long term winner, or should I rather say, not Obsidian in particular, but the open approach. Just see the enormous ecosystem that’s come up around it.

Hence my verdict stays:

  • DTP: Long term memory / storage. Fast, reliable retrieval. I go there only in the same way that I would go into my basement to fetch something out of a shoe box.
  • Obsidian: Short term memory. Very dynamic. My work center every day.

The way out of that for DTP would clearly be to become much more open. Like, radically more. But that’s not happening, and has not happened since I wrote in the above article:

I see Obsidian as a strong competitor to DTP. It is not there yet for some of the reasons I mentioned. But because of its openness and extensibility, there is no reason whatsoever why it shouldn’t be able to. So in that sense, it is like Sharing vs. Telling:

  • DTP: Telling. Developers literally tell you how you should use it. You don’t like it, you often have no way to influence it. They do listen though, like when years ago I found a bug in their metadata storage in extended attributes and fixed it myself. But more generally, it is very much 1990s.
  • Obsidian: Sharing. Literally. You have the power to change every single aspect of it.

Therefore, while I’m using them both next to each other, for different purposes, I clearly can see that DTP has found its disruptor there. I’ll keep using, as Kristensen would say, DTP only as long as Obsidian is not yet good enough for me to do what I do with DTP; the moment it is, I’ll jump ship for these remaining use cases, and never look back (like with Scrivener).

1 Like

They complement one another and do not compete.

Moreover - while for many people both are better than only one, if you had to pick one, DT3 is the only one that can do the job of both.

If you don’t have Obsidian, you can still take notes albeit not with all the bells and whistles that Obsidian offers.

But If you do not have DT3 then Obsidian does not function as a document database except in the most simplistic of situations.

Anyone who gives up DT3 for Obsidian never had a true use case for DT3 to begin with.


While Obsidian has been a media sensation, it’s not clear if Obsidian actually is, or will be, a “winner”. Their CEO estimated a million users, based on metrics including app bundle downloads, in a video dated April 2023. Whether that is an impressively large user base is up to debate.

No one I know in real life is able to use Obsidian for longer than 3 months, although none of them has coding experience.

My opinion is that Obsidian’s influence lies in its “mostly-free” business model. It allows users to experiment with things without a monetary commitment. Many of these users share successful experiments with the community, which in turn grows into an “ecosystem” in which everything is free of charge. (I refrain from labelling Obsidian freemium.)

This model is naturally optimized for social media. A user’s spectacular experiment could become a YouTube sensation with a million of views, again, without any monetary investment. When you become aware of that video and the magic stone software, the curation engine pushes more and more Obsidian-related content onto your screen, creating a powerful hype.

In other words, Obsidian power users are voluntarily advertising for Obsidian, and social media sites involuntarily push these ads towards you. Of course, your seeing many ads for a brand does not automatically make that brand a “winner”.

DT is a paid software with a $99 starting price tag. Few people would buy DT just for experimenting, without an actual use case in mind. Perhaps that means DT’s business model is not well aligned with social media’s UGC ecosystem. In any way, I would say your point that …

… is an illusion.

And it may not be taken for granted Obsidian is decidedly more “open” than DT. Both are not open source software. Obsidian has an admirable marketplace for plugins and themes, whereas DT has excellent scripting support. To some extent, plugins and scripts are different terms for the same thing. While it is true that Obsidian, with all the plugins, can do certain things DT cannot do, this alone does not mean one is winning over the other.

Update: I actually read your blog post. My takeaway is that you prefer Obsidian because it allows you to write your own code to integrate various parts of your workflow. Good for you. However, the majority of users of either DT or Obsidian are not developers. I don’t think Obsidian can be a game-changer for them because of Dataview or inline JavaScript, or because one can “simply launch the developer tools” for visual customizations.


Don’t you know this is bad for business! You’re supposed to exploit user privacy as much as possible to get an increase in short-term profits! :face_in_clouds:

1 Like

I’m expecting Apple to introduce their own local LLM model and API built into MacOS and iOS within the next two years. From the rumors I’ve read, I don’t think it’ll make it in time for this year’s WWDC (but you never know). I’m guessing at that point it’ll be good enough for Devon Technologies, but who knows. (IMO, small models will get there but I’m unsure when due to the amount of resources required to properly train small models.)

I’ve used DevonThink off and on for over a decade. I find that I like the idea of note taking and storing items more than my ability to actually do those tasks as part of my workflow. Being able to query data via an LLM does make me lean more towards using those. I just moved to Bear from Obsidian due to my brain liking the old Notational Velocity method of note taking and searching vs the links/backlinks/etc method. Obsidian does have a lot of nice power features, but it also has more of an overhead than I want to deal with.

I haven’t looked into what DT provides scripting/plugin wise. I just made a todo for me to look into that and seeing what it’d take to pull the data into a vector database. I’m guessing if we want this anytime soon someone will need to implement something to regularly go through devonthink and make it easy to use with a LLM. If I end up doing this, it’ll most likely be something with RayCast.

1 Like

Haha! Well, our “bad business practices” have kept us solvent and sailing steady for 22 years now. :thinking: :slight_smile:


Welcome @soleblaze
Thanks for sharing your thoughts.
Apple would have to do something either very limited in scope or very magical :wink: to make a local LLM (1) broadly useful, (2) private, and (3) performant.

You can run Ollama on a Mac but it’s not up to the tasks our userbase are expecting. For example, you can’t feed it a PDF to summarize. And it still guides you to (stay online for better results) :thinking:.