This approach will directly retrieve the knowledge summaries from an LLM, and use them to construct the input and output for LLM fine-tuning. However, the result would not be as good as the original ...