I could use AI in my DevonThink databases – but I have no idea FOR WHAT

DEVONthink Artificial Intelligence
1

I’m not against AI, to be clear. When I do research, I find it very helpful that these days I have the possibility to ask clever machine minds questions that Google would not have been able to answer five years ago (questions like “what’s the name of that part in an X that does Y?”, for example). I find a lot of AI-generated pictures funny or astonishing. And now I could use AI in my DEVONthink database … but for what?

What do you guys actually use it for? I honestly have no idea. Several times, I’ve let Apple Intelligence create a short summary of lengthy, meandering articles just to get an idea what the author might be talking about, but that’s very much it.

I would love to hear about real-world examples of use cases. Just to ignite my fantasy about what’s possible.

3 Likes
2

We often discuss AI internally – the good, the bad, and the ugly – and the thing we circle back to is: AI is a tool. Tools are for problems and tasks. While your initial interactions will be low-level trivial things, I recommend starting with that thought: “What problem do I have?” or “What task do I need to accomplish?”

Also, be willing to accept you may not have a use case for external AI at this point. Just because you have a tool, it doesn’t mean you have a need to use it. The nature of some peoples’ use of the DEVONthink line doesn’t lead toward AI providing much benefit. But a use case may appear in the future as well.

14 Likes
3

Organize your tags.

Make a summary of a document or part of a document.

Reformat things. Let ai re write, format the document how you want it.

I also use it for doing some adjustments. Like if there are some calculations in the document, I ask for adjustments which are more relevant to me.

What I miss: feed ai entire doc(s) and interact with it and ask questions such that it gives me answers only from the document. No hallucinations

4

Yes, that’s the point – I don’t have a use case for AI. But learning what others use it for might give me ideas for use cases that haven’t occurred to me yet.

2 Likes
5

Before wondering what do you do with the AI tool, how about sharing with the community what you do now with DEVONthink. If you’re gathering data in support of your plan to take over Ganymede, that’s going to lead to one set of AI uses. If you’re focused on curating your digestive biscuit recipe collection (world class), that’s another set of uses.

The point is … it depends. A hammer has dozens of uses, but not many people have need for all that.

6 Likes
6

I use AI privately and professionally for research, document comparisons, document summaries, and much more, but never without my MI, human intelligence, which always stands at the end of my usage and process.

In DEVONthink, I use it for the aforementioned as well, but especially for writing scripts that allow me to directly or through intelligent rules use many new capabilities tailored to my needs in DT. This has made DEVONthink even more valuable to me than it already was. Claude in combination with the MCP-Server is fantastic for this. A sincere hat tip to the entire DT team for this fantastic program!

(Translated with Claude-Haiku)

Ich nutze KI privat und beruflich bei der Recherche, Dokumentenvergleiche, Zusammenfassung von Dokumenten, u. v. m., aber nie ohne meine MI, die menschliche Intelligenz, die bei mir immer am Ende der Nutzung und des Prozesses steht.

In DEVONthink nutze ich es auch zu dem Vorgenannten, aber insbesondere zum Schreiben von Skripten, mit denen ich dann direkt oder in intelligenten Regeln viele neue auf meine Bedürfnisse zugeschnittene Fähigkeiten in DT jetzt nutzen kann. Dadurch ist für mich DEVONthink noch wertvoller geworden als es bisher schon war. Claude in Verbindung mit dem MCP-Server ist dafür großartig. Dem ganzen Team von DT für dieses phantastische Programm ein aufrichtiges Chapeau!

1 Like
7

Here is one concrete use case from my own workflow using AI.

I keep a number of DT databases covering very specific domains of research that span decades, across thousands of papers. Through DT I can query for keywords and strings, and I often find connections between what I am writing and papers I have long since forgotten, or even ones still fresh in mind. Where the newer AI tools have added value for me is in surfacing the near-adjacent, the connections between passages that do not share keywords but share ideas. When I am drafting an article and want to know which of my research papers on a topic quietly agree, disagree, or elaborate on the point I am making, I often hand a batch of papers to an LLM and get results I would not have found through DT alone.

I agree that AI is a tool, to be used on definable tasks, and not every DT user shares my need. Mine happens to be writing across large bodies of research where the connections between sources matter as much as the sources themselves. The cost of using external AI tools is real though. The more papers uploaded for analysis, the more tokens required, and the greater the expense.

One day soon, I look forward to when this workflow can be handled entirely within DT, and completely offline. That has been what I have long wished for since I started using DT more than a decade ago, and I suspect I am not alone. I am eager to find even greater connections across thousands of papers I haven’t yet begun to dream about.

8

Well, I am a novel writer, and I use DEVONthink mainly as a big bucket for everything I come across that might be useful material for a novel one day. I stuff everything in folders about topics (countries, sciences, crimes, etc.), and I use tags to mark documents that are of interest to novel projects I am already planning/developing. I use intelligent folders on the top level to see all tagged material for a novel project, and from there, I follow DEVONthink’s proposed connections (what we once called “AI”), keywords etc. to find more material.

(Before anyone asks: The one use case I am not interested in is having AI writing my novels for me. Writing is what I enjoy doing, and letting AI do it would be like hiring someone to go on my dream holiday instead of me.)

7 Likes
9

Would you mind giving me an idea what that means in currency?

10

BTW: “AI is a tool” … yes, but that is DEVONthink itself, too. It’s in fact a huge set of tools, a big toolbox, and I doubt there is anyone who uses all its functions. I’d even guess most of its users are still discovering everything that is available in this toolbox; at least that’s how it went for me over the course of more than ten years. Reading this forum has been a valuable tour guide on the way. So, a big thank you at everyone who contributed to it.

2 Likes
11

I sometimes use AI for research when there is an article that I am not 100% sure if it fits my needs. I then use this prompt to summarise the text (translated from German with deepl.com). Then I can see in one glance if I should dive deeper into the article or just leave it be:

Role

You are a research assistant and are tasked with creating a structured summary in accordance with academic standards based on the provided document text, summarizing the main arguments and the core message concisely. Paraphrase the text so that it can be used in an academic paper. Maintain a neutral, objective, and detached tone.

Source

  • Document text: %@

Basic Rules

  1. Text-based: Base ALL statements exclusively on the provided text. No personal interpretation, no evaluation, no external knowledge.
  2. If not addressed: Explicitly note missing categories as “Not addressed.”
  3. Page numbers: The page numbers printed in the text may differ from the PDF page numbers of the document. List both, if distinguishable: p. 12 (PDF p. 14).
  4. Quotations: Use direct quotations only if the exact wording is academically significant—in quotation marks with the source reference (p. X, lines Y–Z) .
  5. Language: Academic, with precise paraphrasing. No evaluative adjectives (“interesting,” “significant,” “good”). Gender is indicated not with * but with a colon to avoid conflicts with Markdown formatting.
  6. Length: 300–800 words, maximum 20–25% of the text’s length.

Output Format (Markdown)

Bibliographic Information

Author(s), Title, Year, Publisher/Journal, DOI (if applicable)—as far as can be determined from the text.

Source: %@

Abstract

A brief abstract of the text, either copied from the text (if available) or written by the author.

1. Research Interest & Research Question

Central question, objective, problem addressed. → Cite sources with page numbers and references.

2. Theoretical Foundations

Theories, concepts, traditions; definitions of key terms. → Cite with page/source.

3. Nature of the Subject Matter

Subject of investigation (phenomenon, discourse, population, artifact, era, etc.). → Cite with page/source.

4. Methods Used

Research/analytical procedures and their rationale. → Cite sources with page numbers and references.

5. Data Set & Sources

Data material, sample/corpus size. → Cite sources with page numbers and references.

6. Key Findings

Most important findings, theses, conclusions. → Cite sources with page numbers and references.

7. Limitations & Open Questions

Limitations identified by the text itself and areas requiring further research. → Cite sources with page numbers and references.

8. Relevance & Applicability

Author’s own statements regarding the contribution; key referenced works. → Cite sources with page numbers and references.

Quality Check (check yourself before submission)

  • Have all statements been cited with page numbers?
  • No external information added?
  • No evaluative adjectives used?
  • Word count within the target range (300–800 / ≤ 25%)?
4 Likes
12

I also recently used AI in DTP for my political work to generate texts for different purposes. I write a general and detailed draft and let it extract a Facebook post, an Instagram post and a Mail to local politicians about an issue we identified based on that general draft.

13

In DEVONthink or another app?

14

In DEVONthink with the PDF Selected

15

So it’s a custom summarization defined in Settings > AI > Summary?

16

Sorry, I forgot to mention that. Yes, it is

17

The newer AI features can assist in suggesting connections among the material at perhaps a deeper level or more nuanced than the concordance-based AI in DEVONthink. If “Search: Web” is also checked, it could also suggest connections in external sources.

Two caveats. One, this kind of use of AI can instantly be a rabbit-hole time-waster. Second, exercising the AI’s brain to suggest connections can easily be a substitute for exercising your own brain, which might be an unfortunate thing for a writer.

3 Likes
18

I use DT to manage my research and thinking (e.g. my notes on things). I’m anti-LLMs for various reasons there’s no need to re-hash here, but I also make a point to know about things so I have DT set up with a local LLM and have tested it.

I still don’t really have a use for it as I believe interacting with my files is part of the thinking process, but one thing I have found it useful for twice now is finding a specific stat in a report. Government reports can be hundreds of pages, and depending on the government and the nature of the report they can sometimes bury the numbers a reader might be looking for. I’ve had two cases recently where I knew the report would have the number I needed, but the key words would appear across the whole document and search was going to be monotonous. Both times, I asked the LLM to find the stat I needed and the page number and it worked.

(This example worked because I’m familiar with the files in my database and I knew which one would hold the data I needed, I just didn’t want to spend the time scrolling to find it. It would’ve been messier if I didn’t know where to look, or if I didn’t know if my database even had it.)

2 Likes
19

I more or less could have written the first paragraph that @straylor wrote. I’ve lately been using the MCP server via Claude quite a bit to interrogate databases.

As well, I have used various AI models to build applescripts. One is attached to Keyboard Maestro that converts selected DEVONthink records from HTML or Markdown into a formatted RTF or reformats existing RTF/RTFD records in RTF. Nothing special, I’ll admit, but it was more a trial to see if I could create some kind of automation.

Another script captures an active Safari web page into the Global Inbox, applies a specific tag, but uses specific capture handling depending the source (i.e., as we know, some news sites do not like being captured as, for instance, web archives, so this script checks the source and alters the capture method accordingly).

1 Like
20

This is something I considered, not least because I have material in several languages, so a concordance search wouldn’t be able to find connections. But then again, my four main databases pile up to 400 million words in total, and I guess it would be prohibitively expensive to have an AI read and analyze all of that. :money_with_wings: