I wrote something like this using FUSE and Postgres a couple years ago. The major problem with this approach is scalability: on a large, multi-user FS, you can get TONS of tags to sift through in your root folder.

My solution, which I never got around to implementing, was to have "hierarchical" tags. e.g. "recipes:cookies:chocolatechip" is a single tag for all chocolate chip cookie recipes, all of which have the implied tags "recipes" and "recipes:cookies" (but not "cookies" or "chocolatechip"). The advantage is that your root directory will not be littered with the "cookies" and "chocolatechip" tags when these collections are fairly irrelevant.

This idea can be extended to solve multi-user conflicts (all tags are prefixed with the username) and to make sensible tags for dates -- it makes little sense to tag a file with "2011", "May", and "22", since no-one cares to find all files from the 22nd of every month; but the hierarchical tag "2011:May:22" is perfect for this situation since these files will also be present under "2011" and "2011:May".