Who should read tasks from the queue and send them to LLM, and how to publish chunks and final results of the response back to the queue?