You can try to switch to "hybrid" in the KB Agent config, as GPT-3.5 turbo will be tried first and is 30x cheaper than GPT-4