Help using AI via Devonthink with API key (newbie)

AI
1

Ok, Exploring moving from interactive ChatGPT sessions to using Devonthink as my primary AI interface.

I understand that I will need to get an API key from the LLM provider of choice (ChatGPT) and from what I have determined, that is billed a la carte for usage, but is also completely independent of my existing paid $20/month ChatGPT plan.

So, if I want to only use API key access, am I correct:

1 - I will have to backup saved chat sessions and ‘projects’ in my current ChatGPT app workspace to external files (which I can later incorporate into my DT files/database)

2 - Cancel my ChatGPT monthly account

3 - For interative, casual chatting, I would have to do that inside Devonthink using the DT Chat Inspector and DT Chat popover (if I want to keep archive/history of my sessions as I do now)?

4 - If using the web directly, or the ChatGPT macOS app, I would be using the free versions that don’t keep memory or history of my sessions?

5 - Lastly, how would I use ChatGPT from an iOS device? Does DT have a “Chat Inspector” as part of DevonThinkToGo or as a standalone app for iOS?

(I do plan to consider other LLM’s in addition, or in place of ChatGPT, but think replicating my current setup to use ChatGPT and API tokens inside DT is my first step to changing my AI workflow to benefit from DT’s support for AI work before going on a side-quest of comparing various LLM’s or even running a model locally.)

Open to any and all suggestions, including not using DT, but I already own/use DT heavily.

I guess I should explicitly say - my primary goal is to eliminate the friction in merging LLM output into some of my existing writing-heavy projects in DT. Keeping track of multiple branches of exploration, and avoiding all the formatting problems I’ve been having by manually cut & paste I am doing currently.

Have no interest, at this time, in using DT for image generation/editing, or any video/audio manipulation via AI. (Have those covered in other tools)

1 Like
2

Last I checked you can’t return to a previous chat in DEVONthink. Apparently this is scheduled to be added in the future. This alone makes it extremely suboptimal for use as your main AI tool.

You can save your chat to text file and then add a text file to another chat in the future if you really want to refer to it, but who can be bothered doing that all the time.

Plus it lacks stuff like snippets and roles you get from Mac based AI tools, and (last I checked) it can’t do stuff like create Word Docs or html files that the online AI tools can do.

I don’t think DT is really suitable as a main chat tool, and it’s not trying to be tbf. Its AI functions are in support of its document management. Not quite as bad as playing Doom on your fridge, but same idea

If you want a better connection between AI output and DT, I’d suggest using Hookmark. If/when DT makes it possible to save and return to a chat about a specific file, you that may tip the balance.

3

I would say that it playing Doom on your fridge is quite far from optimal :joy: I think the AI features in DT4 are good and useful, I use Claude mostly, but I agree that it is not optimal for chatting (but much closer to optimal than Doom on a fridge by a large degree).

4

You should buy API credits at platform.openai.com (there are also other options). Subscriptions to ChatGPT doesn’t work with DEVONthink.

Not in DEVONthink To Go 3. There is a beta for DEVONthink To Go 4. I’m haven’t used it so I don’t know what type of AI support it has. Have you asked over at https://discourse.devontechnologies.com/ ?

5

Thanks everyone. I might explore DT further buying some credits, but definitely feels like for general AI use, I still need to keep a native chat interface (free or paid).

Since I still need to manage that, it seems simpler to continue using that as my regular workflow and just cut/past results back into DT or other tools.

If DT or something else can truly be a superset “AI command center” for both my general and document/workflow centric needs, I’ll take another look.

Just to clarify - this discussion is independent of planning to take advice here and evaluate Claude to replace my using of OpenAI/ChatGPT, which I will try to set-aside some time to give it a serious side-by-side comparison.