This is fascinating, but how are you retrieving this data? Are you running the fan-out prompts separately to see the citations or are you somehow harvesting from OpenAI backend response file?