Usage of vector databases

[WardVH]

You probably have considered using vector databases for storing the behaviourals? Would love to hear tour opinions about it.

[nvandessel]

Hi @WardVH , appreciate you checking out the project! This is a really great question!

So I'm not using vectors right now - retrieval is structural matching on context fields (language, file path, task) plus a spreading activation graph that learns associations over time through co-activation patterns. It's precise and debuggable, but it's rigid - if a behavior isn't tagged well, structural matching just misses it.

Vectors as a complement to the graph is something I find interesting though. A fresh floop store is kind of like a baby brain; it has to learn all its associations from scratch, through experience. The graph gets smarter through usage but starts cold. Vectors would give you fuzzy similarity from day one, which could help bootstrap things before the graph has learned enough connections.

I've been looking at https://github.com/hybridgroup/yzma (purego llama.cpp bindings) for a while now. This would give me a way to ship a local embedding model with the project - so no API dependency, and runs offline. Using that for vector-based retrieval alongside the existing graph definitely feels like a a natural progression!