+1 it wouldn’t be terribly useful if it were only caching the tokenizer output.