Feature request: full score matrix

[ManuelFay]

If I understand correctly, 8192 scores are fully computed anyways. Does this mean on corpora with less than 8192 docs, we could get the full num_queries x num_docs score matrix ?

This could be very useful for compatibility with some evaluation libraries such as MTEB (and where we often have less than 5000 docs anyways) !

If it is possible, could we have a method for this ?

[raphaelsty]

By setting top_k to the actual number of documents you can get the whole scoring but I do get the point of having a matrix, I'll add this :)

A .matrix method which would return all the scores within a torch tensor would be fine ?

[ManuelFay]

awesome ! one key thing here might be to preserve the order (or reorder if needed) ?

[raphaelsty]

The matrix method would return a tensor

Q*D where Q is the number of queries and D the number of documents. The row and column of the matrix would be ordered with respect to the input queries and documents orders. Q_i D_j would be the score of query Q_i for document D_j