I ran this under Devel::NYTProf and found that this loop is the hot spot:

        for(0 .. $tf)
        {
            my $last=vec ($wektor, $index++, 32);
                        $pos{$docID}{$last}=$order;
        }
[download]

I changed it to this, following a suggestion from dvryaboy and also getting rid of the block completely:

    $pos{$docID}{vec($wektor, $index++, 32)}=$order for 0 .. $tf;
[download]

That got somewhat faster.

I tried taking out the constant reference to $pos{$docID} like this:

    $pos{$docID} ||= {};
    my $did_ref = $pos{$docID};
    $did_ref->{vec($wektor, $index++, 32)}=$order for 0 .. $tf;
[download]

...but that didn't make much difference to the hot loop, and it was more expensive outside the loop than the savings it got inside.

This is without really understanding what's going on, though. I wouldn't be surprised if what you're doing would benefit from just using a better algorithm.