Different output generated for same prompt in chat mode and API mode using gpt-3.5-turbo