The way openai implemented this is really clever, beyond how neat the plugin architecture is, as it lets them peek one layer inside your internal API surface and can infer what you intend to do with the LLM output. Collecting some good data here.