How was fault tolerance organized in the recommendation service with OpenAI? How were prompts built, how were failures handled, what were the bandwidth limits?