For longer tasks, we should save the replies from the Ollama server so it's possible to recover from failure in longer annotation pipelines.