What reliability, cost, performance problems do you run into when using LLM APIs? For example, what issues have you had when using Open AI’s Files API?