vector DB that stayed out of the way with vLLM

[thetafield]

What worked for me was pretty simple. Before rebuilding anything, I'd test one clean baseline with Llama and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.