I just downloaded 4, and need to know about integration with Claude. I am working on a large novel, have a Claude Project so I can upload material to make it available for queries, and am using it for editing tasks, historical accuracy checks, and research. [Not writing/rewriting; that I do myself, thank you very much.] Need advice on whether integration with DTPro 4 would help with that, and if so for what, and how to go about setting it up. Thx
You should read the Getting Started > AI Explained section in the help or manual.
And there’s really not enough information to go on here, but DEVONthink hasn’t become an AI application where it will be ingesting a bunch of documents for you to process like you would in a dedicated application.
Hi Jim; Thanks for the quick response, as well as your patience and time for a too quick question. Didn’t phrase it well – I really only needed to know if it was possible. I’m too far behind on all this to catch up quickly, so mostly wanted to know if it was even feasible before I invested the time it would take.
You’re welcome!
While you can certainly examine and chat about documents, in DEVONthink it will be in smaller, very controlled batches - for privacy and performance.
Bespoke AI applications, like LM Studio, are employing vector databases to stored the embeddings of documents from which AI can draw. At this time, this is not a capability DEVONthink has. And you may hear tell of other apps building a vector database from DEVONthink’s data but this is neither officially supported nor sanctioned by us.
On the other hand, through scripting – and with a larger budget for AI – it is possible to process larger numbers of documents with external AI. However, it is not something built-in nor is it necessarily easily coded.
Hi Jim: Cost/benefit analysis says the API charges would bankrupt me given the size of my book 
and I’m managing at $20/month by uploading chapters into my Claude Project Knowledge for editing – plenty of room in there, but it isn’t well organized. Minor inconvenience given all I’m getting out of it – having previous chapters available means I can get structure and story/character continuity analyzed. But I’m in DT all the time, and will definitely give your new updated AI in 4 a try. All the research for my book is in there, and I expect it will be useful.
2 Likes
This is interesting! Does the Claude Project Knowledge create a “memory” for the chapters you’ve uplaoded? I was under the impression that the AI would have to process all the files every time.
I have a couple of shelved writing projects that I’d like to use AI in this way, but every time I try to upload chapters, I get too many token errors.
Newbie to this, but so far my experience is: I have Claude Pro, $20/month (annual cheaper) API Console is a separate subscription, so I’ve seen no extra charges.
No it doesn’t remember anything but current chat (not yet; it’s coming), but I can work around that. I’ve uploaded chapters into Project Knowledge (no extra charge, and seems to be no limit on data stored), plus any other relevant info – they say the more you put in there, the faster, more efficient a search will be. The Claude site has detailed information about the tiers, costs, how to make use of it, etc.
I had Claude help write a default set of instructions to be used for every chat. Said I was writing a novel, genre, setting, etc., wanted a thorough edit. The Claude response was based on what a professional, human editor would have been trained to look for, as in “Do a Line Edit, followed by a Structural Edit, then check for Historical Accuracy” with details about exactly what was wanted for each of those. I added some instructions such as "Do not rewrite my text; intersperse comments between paragraphs. Or “check if dialogue is appropriate for character, setting, etc.,” After some back and forth, I copied and pasted the result into the space for the default instructions, and it is now used for every chat without my having to retype it. It can be edited later based on your experience of what works/doesn’t.
When I start a new Chat, I say: “Find file “Ch X1” and perform default instructions.” After a response I can ask it to look at one or more other chapters to check for sub-plot continuity or character development, etc., or ask it to check for any redundancy with material in a previous chapter. If it makes an error or overlooks something, I redirect. Every chat is saved so you can go back and reread it, or copy and paste it to your computer.
Problems: I have had it hang up, which seems to be about server overload. If it is making an error, I can click on Pause and redirect, (one night it hallucinated and we had a chat about that). Sometimes it skips parts of the chapter and I have to tell it to go back to a particular paragraph (Claude told me to include a word count with the chapter name so it doesn’t overlook anything and that helped), or I’ll think of a relevant detail in a previous chapter and tell it to also include that in the analysis.
There’s a time limit - I’ll get a warning message that I have 1 minute; when it cuts me off I can come back in 5 hours). It doesn’t organize the file hierarchy in any way that makes sense, and doesn’t display the filenames, so I go by the filenames in my Finder when I make an inquiry.
I get the feeling it’s starting to remember my story, which is spooky, I must say. I understand there are AI apps that allow you to create an Avatar and tell it about yourself, and have it remember everything you say (major privacy risks there). Seems to be mostly used for boyfriend/girlfriend AIs, but probably good for helping the elderly with medications, reminders to make phone calls, etc.
Claude has announced Memory for it’s top tier and says it will be available for all subscribers in the near future. That would be ideal for a project like mine – I’ve uploaded an entire book and could reference all of it in a chat, instead of just one or two files, and every previous chat would be remembered – huge time saver and much better responses.
This has been a mind-blowing experience – I’m writing historical fiction and have spent hours digging into the internet for answers to questions, and I ask Claude or ChatGPT and blink, there it is! It provides the reference to the sources used (Wikipedia on steroids – and yes, it uses Wikipedia a lot).
So dust off that shelved writing project of yours and give it a whirl. Lots of fun.
3 Likes
Thanks for the detailed breakdown of your experience with AI and writing.
It’s very helpful, especially the part about setting up default instructions for editing; that’s exactly the kind of workflow tip I was hoping to learn. I like how it can assist with various aspects of plot flow and character continuity. I’ve had several minor characters where I’ve needed to pause and deep dive into “who” they were before continuing, because I forgot too much about them!
Ah, yes, our memories are definitely porous. I found it helpful to write up a brief bio of my characters – it clarified my thinking, and it’s the sort of think you could upload so Claude can refer to it.
Definitely ask Claude for help defining what you want in the default instructions – I started the query with something fairly general, and was amazed at how detailed and thorough the suggestions were – way beyond what I would have thought of myself – and then I edited it for my purposes. Also, you aren’t limited to one file per query – you can ask it to look at several, or at a chapter plus some supplementary material. It provides “enhancements” i.e., suggestions for how to improve writing, pacing, or whatever, and you can challenge that and get into a dialogue so you have various possibilities to choose from. Important: It’s Your Writing! Keep your voice; take or leave the suggestions.
I’ve also done a little with ChatGPT. So far it seems better on historical research (some of it is astonishingly detailed), and it will generate images to your description (like how you imagine a character to look) which is fun. It has a similar setup – you can have a Project, upload materials, etc., but with much smaller limits. Try both – can cancel at any time.
1 Like
PS: Just rechecked ChatGPT Plus. No uploads for Plus, only for Pro. But I have pasted in an entire chapter, asked for the usual analysis I described before, and got an excellent, clear response. That Chat is stored in my project and is available for any subsequent chat. Each chat shows in the sidebar, can be renamed, downloaded, or deleted. If you run out of time, you can return to a chat and continue. Or refer to it, and continue with additional questions in a new Chat. In many ways, it’s far more convenient than Claude.
It’s great that AI (LLM or whatevs) has gotten to the point where this is possible, as this is where I see the benefit. Being able to have it use my database or notes as the main knowledgebase for my questions is how I want to use it.
When you’re using it like this, does it automatically know to use all of the documents or do you specify which to use in your prompt?
For example, with writing… Having no dellusions of grandeur, I like to write for the fun of it and tend to go in waves. Which is how I have several unfinished works and many “barely started” ideas. Picking these up again usually means having to reread what I’ve written, along with all of the notes I have on it. If I know I only have a small window of opportunity to do some writing, this often is the reason why I don’t. It would be nice if I could ask a general question like, “Where did I leave off and where does the plot summary need to be picked up?”
Another use case is I’m currently in the process of buying a business and I have many documents regarding this and I’d like to ask random questions about when they pop into my head, as well as more serious questions pertaining to the loan, franchise rules, and so on. For some of these questions, I don’t necessarily know which documents may have the informatoin, or the AI might find something “hidden” in the massive amount of text that I was unaware of.
Yes, both apps can search in past chats or uploaded files, but the way Claude and GPT go about it is different. Being specific and concise in any query helps a lot, like: say where to look.
So far Claude makes more mistakes and wanders off into the weeds more often than GPT, but my default instructions may be to blame. They can help you write them. I also much prefer GPT for historical research; same test, both apps., Claude failed miserably, GPT results were awesome. They were trained on different books?
Good idea to have separate projects, i.e., writing, business. Chats you didn’t start in a Project can be moved. You can return to a Chat and continue. Some things seem easier to deal with in web version vs desktop app; but I prefer the latter.
If you don’t want to leave material on their site, you can download a pdf or doc of the Chat into your computer, then delete it. In terms of security, I get the feeling DT prefers Claude (Bluefrog, yes?) but I’ve dug around a bit and for what I’m doing both seem reasonably good.
Copyright? Most concerns are about AI taking your material (Claude says they don’t by default; option in GPT), and about people passing off AI writing as their own. I’m working with my own material and asking for editing and historical analysis, so I’m not worried about violating someone else’s copyright, but have some residual concern about someone violating mine.
They say my Project is encrypted and no one has access. Have fingers crossed that’s true. Maybe less is more would be a good approach, but when GPT or Claude have access to previous chapters during the editing process it’s invaluable for analyzing a book.