Sentence reordering seems to be messed up when app deployed

[thiborose]

I don't remember exactly how it was, but when the app was still working online, output sentences were super weird, i.e. mixed up one with the other.

I suspect that this is due to the fact that there is only 1 tsv file which is overwritten at each query. So if more than 1 user tries to predict, there will be a conflict.

[thiborose]

I tried to change the architecture of the model to avoid relying on files on the disk. However it was very complicated, and due to the lack of time I decided to create a new file on the disk for each query, hence no conflict.

I expect it to solve the problem. To be checked