Provide guidelines to maximise items cardinality

[minottic]

I think the documentation should note that RIs are strongly recommended to compute the weights by using as many items as possible to compute the weights. Maybe I've missed that part, but at first, I was confused at why the .ipynb was using private datasets even if they will not be later shown by the search-API. I think the reason is indeed what was mentioned at the beginning of this issue, namely computing a more accurate scoring.