that's weird.. Try changing the llm? sometimes that's the bottleneck, depending upon your prompting