[1] http://www.miislita.com/information-retrieval-tutorial/svd-l... [2] http://nlp.fi.muni.cz/projekty/gensim/index.html
The other systems I've used the approach for have all been along either bibliometric/bibliographic lines, or have been relating to content-based image retrieval. It's a pretty robust approach, but can take a little bit of tuning to get just right- coming up with a good evaluation strategy is important to getting the most out of it, I've found.
As far as references that were useful:
Ilya Grigorik has a very accessible getting-your-feet-wet tutorial on his site: http://www.igvita.com/2007/01/15/svd-recommendation-system-i...
It might be a little dated w.r.t. specific libraries or APIs, but the basic technique is there. For a more comprehensive look at the SVD-IR approach, take a look at:
Berry et al. Using Linear Algebra for Intelligent Information Retrieval. SIAM Review (1995) vol. 37 (4) pp. 573-595
The SVD approach falls in the same family as Latent Semantic Analysis, which is a whole black art unto itself- I'd actually suggest going back to the early papers by Landauer, Dumais, etc. if you're really interested- those guys did a great job writing up what was at the time really novel stuff.
My contact info should be in my profile, drop me an email if you have any more questions (or to let me know what you end up doing!).
having said that, i'm always happy to hear about success and failure stories from users. so if there's something you find broken/missing in gensim, just let me know.
Some searches suggested: http://www.picloud.com/
HOWEVER- the big catch with naïve (i.e., simple to implement and understand) approaches to LSA are that every time you add a new datum, you have to recompute the SVD of your data- which is time consuming, and raises some interesting logistical challenges when done in the context of a web app (i.e., how you persist your matrices, etc.). There are more sophisticated LSA approaches that get around some of these problems, but they're more complex to implement.
Nothing insurmountable, but kind of a pain nonetheless, and w.r.t. the system described in the parent post, I'm not sure how well they'd integrate with GAE- can you use NumPy/GSL/etc. on GAE? Or would you be stuck using some sort of Java numerical computing library?