Hello!
It would be nice to create an Inkdrop plugin for ChatGPT so that we can search our notes by using natural language or augment ChatGPT capabilities with our data stored in Inkdrop.
More info:
Cheers.
Hello!
It would be nice to create an Inkdrop plugin for ChatGPT so that we can search our notes by using natural language or augment ChatGPT capabilities with our data stored in Inkdrop.
More info:
Cheers.
Hi David,
Yeah, that looks awesome.
I haven’t read the docs yet though, does it support accessing localhost? Because notes are end-to-end encrypted, it has to work locally, like the Inkdrop web clipper browser extension and live-export.
I could add this with GPT4all although it will run offline but it wont be as much fun as chatgpt
I guess it can simply use OpenAI APIs from plugins instead of using their retrieval plugin.
It has to send note content to the API though, it’d be easy for everyone.
It will look similar to how Notion does.
I just moved all my notes for a project over to Intellij as an experiment, because I have a CoPilot license and the extension is integrated. It instantly indexed the existing notes and the suggestions are immediately useful and reasonable. I would say this kind of LLM-assist feature has rather suddenly become table stakes. I love the Inkdrop UI and paradigm, but I’m not going back until this feature is available.
Adding that 30 minutes later my next insight is that this is table stakes for a much more profound reason than editing experience, the assistance of the automcomplete etc. It’s that this is actually a different paradigm for a knowledge tool. Specifically, this approach turns all the current and future notes you create into a continuously improving system, where the new content is indexed and the assistance continues to improve in ways specific to your notes. This is a feedback loop where the more you use the tool to the better it gets so the more you use it etc.
Hi @Mark_Weiss ,
That sounds interesting.
I am actually only licensed for GitHub Copilot, not ChatGPT, and yes, even just indexing over Markdown notes it’s been offering helpful suggestions. But I do expect that for general use cases a more generally trained LLM would do much better.
I can’t really show you exactly my setup, since this is for work, but this is representative.
The setup itself isn’t remarkable. In fact, the editing experience is a bit worse than Inkdrop, but the split view with rendered text side by side is nice for some use cases and another feature you should consider. Notice the file tree on the left, just like Inkdrop. Now imagine all those files were indexed and you were getting back useful, domain-specific autocomplete! And again, think about how this makes the product more sticky and more valuable to the user. The more notes they add, the better the index gets. This effect is concentrated, the more domain-specific and domain-focused the notes are too, so for the use case of using Inkdrop at work it is definitely compelling.
Your privacy point is a good one. I think you should experiment with non-trivial sized local sets of notes, and look at open-source language model libraries targeting smaller data sets. If you can get useful results, you would only be running indexing locally. There is the additional issue of backups to the cloud, but that is an existing issue now. For example, I don’t back up to the cloud on my work instance of Inkdrop because you are not an approved vendor. So I do a local backup and push it to an approved storage vendor. So I would say you would only need to also honor this setting if you added this indexing feature and never back up the index off the device if the user didn’t approve that.
Thanks for the explanation.
I didn’t understand how much helpful it is just by reading your comment, but that sounds intriguing.
I haven’t been using CoPilot but I’ll try it again and see it from the note-taking perspective.
BTW, Inkdrop supports side-by-side mode:
+1 for Github Copilot. I often used it to write my technical comments/notes because, unlike other AI writing assistants, Github Copilot only generates a short text instead of a full paragraph (which is good when I just want some hints in writing docs)
Or, instead of Github Copilot, any LLM auto competition (with a shorter / configurable completion length) would be great. I am looking at the writer-ai plugin, but sadly, it is not what I want. It generates text instead of suggestions