im working on something custom built now that AI agents are useful and you can just rent a GPU for a few bucks to classify and transform data. Wouldn’t trust a company with that amount of personal data anyways and that’s what makes it hard to build a product in that space. Apple would be a great position with on-device processing but until then I’m running my own. I’m still in the middle of this but the idea is roughly:
- get all the data from takeouts GitHub - ivbeg/awesome-data-takeout: Awesome list of data export pages/tools for most common online services
- transform the data into something that’s queriable, has frontmatter metadata fields, a one line summary of what the doc is
- create an embedding for that data, maybe using https://www.cognee.ai that combines RAG with traditional SQL
- have many locally run AI agents use that data to create entities (people, events, automated journal of every day, summaries of weeks/months/years - my understanding is that cognee does a lot of this automatically) with links that reference source doc, MOCs
- fill in gaps manually
- create a compressed version of it all that fits into 200.000 tokens or less and use as context for personal questions
In a recent interview @kepano said he sees Obsidian mainly as a tool to link notes. Not sure if that’s the right strategic vision if AI can link notes for you (either via “copilot” type editing or embedding spaces that move related notes closer together)