They are using a fine tuned roberta model and emotional analysis. The first step is to fetch/scrap the content and analysis. Perhaps the model is returning different values each time. I am not too familiar with LLMs but I think not 2 results for the same input are unique and thus these variations.