Also, the only way I can imagine that this reliably works is by transcribing the audio & ocr-ing the video and subsequently extracting the recipe instructions and structuring the output with a LLM