This was after stemming, but it was using the default set of words. Either way, with so many documents (and all of them being written by a different person -- these are auction descriptions) the variety and number of truly unique and non-stopped words is going to be really high.

Can you (or anyone) confirm for me that my understanding is correct, in that you need the entire vector for each document to compare against your search query's vector when you cosine them? That is, you couldn't perform some function ahead of time on your vector to reduce the size of it (yet still have the same representation) and have it be compared properly to the search query vectors? Since the two are so interrelated, I don't know that there could be a way to do that since each search is unique and performing some shrinking computation on it ahead of time would destroy the ability to gather results against each different set of query words...

Like I say, I was just sort of hoping against hope here. I need a viable solution for searching on my site because I have hundreds of thousands of old (closed) auctions, thousands of existing (open) ones and then hundreds of thousands of posts in the forums. So as you can see, simple SQL queries are really not cutting it any longer - both in speed and accuracy (not to mention flexibility). I'm not even sure that I could properly use a reverse index method on this type of setup. Aaaargh.