the summarization becomes much better if we allow the model to first generate a candidate summarization and then improving on it.
doing the improvement step just once seems to significantly improve the summary.
we also now use an llm (mistral 7b) for the summarisations, as we can then use the same model for multiple tasks and serve it using gpus, thus significantly decreasing the latency.
it might be easier to score pages based on their rank of the sorted their centralities. for instance the centralities for page A and page B might be very similar numerically, but if a lot of pages are between A and B when looking at the sorted list, the highest ranking page might in reality be a better result than the lower ranking one.
the rankings are calculated using an external sorting algorithm to account for the fact that we might need to sort more nodes than we can feasibly keep in memory at once.
