Many of the responses from the API are very large and eat valuable context window. We should think through strategies on how to give the model options for smaller results. That might be*_small tools that return trimmed results or possibly the option for the model to pass in a nested set of keys to select.