I never had the opportunity to play with GPT-3, but I’m very curious how this could fit the Note-taking picture.
I’m pretty sure that many people would feel great when being helped to generate specific types of repetitive content, like some office work, forms, things with known structure.
Perhaps using it in a personal vault would generate lots of noise and ‘distant’ text. The note-taking activity is rather intimate, needs emotional involvement; here, more is not better (at least for me).
But it would be cool to give it a try and discover when one could use this.
I would like something that helps me diminish complexity, which proposes to delete some notes, merge two notes that are too similar, summarize, etc.
I’m curious with what are you doing with your bookmarks?
With obsidian notes I’m experimenting with showing text of other notes that is similar to my current text. Also doing page-rank to my notes. It’s really useful.
I’m building a recommendation engine for my bookmarks. Right now it’s based on TF-IDF. I tried fancier embeddings like bert and doc2vec, but I’ve had the best results with TF-IDF.
With this similarity metric, I’m trying things like building a graph based on the top 5 most similar and then running some type of clustering on top of it.
It’s actually working surprisingly well and was born out of me not wanting to carefully organize every blog post, article, and paper I find interesting. It’s gonna be the first filter I use to determine if I want to distill an information source into my obsidian knowledge base.
It’s cool to hear your also using ML to understand your notes I think there’s a huge opportunity for this!
Text Generator Plugin’s goal is to have at the same time the capabilities of Obsidian with all my notes and the capabilities of Open AI GPT-3, which hopefully makes my writing very fluid, fast, and powerful.
I don’t know a lot about how natural language processing works and how models are trained but I wanted to share anyway because I thought this could be a game changer for obsidian especially as a tool marketed as a second brain. It would be great if a natural language model can be trained using one’s knowledge base from the md files stored in their vault and then an AI can respond to any questions the user asks based on that knowledge.
I think this would be the ultimate and most effective method for information retrieval and it could also allow users to identify creative links between their ideas that might otherwise not have come to notice.
I consolidated a few FRs/plugin ideas here. We are aware that there is hype/demand for this and, believe it or not, this is not technically challenging provided that you are willing to send your data to ClosedAI. In fact, I think we are getting closer to ten plugins that interface with OpenAI.
If you want to do things locally (Obsidian’s ethos) and cross platform on a regular computer or smartphone, things get much tricky. We’ll see where this goes.
For anybody wanting to kick start their own script for conversational interface over given notes with OpenAI API, here is a notebook I’ve just banged out to see if their cookbook is worth anything. It turns out it is.
The script takes just one pdf file, applies bespoke transformations to extract chapters, calculates the one having smallest distance with the user’s query and sends it as a context to gpt3.5 API call.
Just a poc, but may be useful for anybody wanting to build their own stuff. GitHub - rpast/tsunset
