If a LLM starts streaming answers and fail before return, it should be possible to automatically retry.