Realistic Corpus size

[David Webb]

I have about 2.5million documents that I can analyze with LSA.  I currently have a test sspace that I generated from 10K of those documents.

Is there a magic number of documents to analyze, where the return diminishes?  By return, I mean the time to generate, load, or query the sspace file.

Assume that all 2.5 million are similar documents (resumes) and a whatever sample size I choose should provide a good representative sample of the entire set.

Thanks.