I've found that our kb queries are consistent with using the same model on the openai playground but with added wrappers like the personality agent, for instance