⚡️ Ideas
Sarah Arminta Sarah Arminta Mar 4, 2023

AI Embeddings for Similar Nodes

Explain the problem as you see it

For research and writing I should be able to quickly see related nodes, not just by exact wording or linking, but by related topics and including nodes that I perhaps had forgotten or didn't realize were related.

Why is this a problem for you?

Using AI Embeddings to surface related notes is now a priceless part of my workflow that I can currently only do in other apps. The amount of time savings is huge and the quality of my work has gone up considerably, including being able to make serendipitous connections that I never would have otherwise. I want to do this in Tana.

Suggest a solution

Embed and store embeddings for each node in a workspace and include a section at the bottom of each page, near the references section, to show the top 5-10 similar nodes.

⁨7⁩ ⁨Comments⁩

Oh this would be so amazing! Especially, if it was possible to prioritize nodes with the same (or related) supertags.

Imagine opening a node for a meeting, and seeing related meetings with similar agenda, or tasks, or discussion items.

Or opening a node tagged with #Flight, and see nodes related to the same trip.

Or opening a #book node, and seeing book nodes with similar content..

Just as when you are about to input a new idea here, the website suggests pre-existing ideas that may be about the same thing, when you tag a new node in Tana, it should make me realize I already have an existing similar tagged node elsewhere in my graph.

The ability to AI search / 'chat with my notes' (i.e. search my notes by semantic meaning using vector embeddings) is the single most helpful feature Tana could build IMO.

I respect how hard it is to prioritise features, and I'm sure the team talked at length with the community when deciding to prioritise the meeting agent, but I personally have very little need for it. This feature instead would be the killer one for me personally. I wonder if the legibility of the solution means few people explicitly ask for this, but I suspect it is a feature that drives many many acquisitions for mem and the others that support it.