How to design and scale a chat service with LLM: choose between SSE and WebSocket for client-server communication, organize streaming responses from the model, and handle requests?