Thanks - we used to use a bunch of k-shot prompts (particularly with our previous idea before we pivoted when we got into YC), but with the davinci model we were sending ~3.5k tokens per invocation, which in the long term was costing far more than finetuning!