What I'm experimenting with is building a graph out of "entities", which I'm calling people, places and things found in each card. The point I guess is to allow the LLMs to make connections between cards, but this won't be replacing user made connections.

This lets me pick up all cards that reference something like, I don't know, "factories" as a concept, which isn't something the user will necessarily want to directly connect.

Downside is that deduplicated taxonomy: I have, from a scan through my list, "factories", "Factory" and "Factories". My attempt to head this off is by vector searching based on the proposed entity keywords and having an LLM decide if they are the same or not. I think I need to work on my prompt for this though.