Just a quick non-thought-through idea but would it be possible to build an index in a way that allows clients to download only parts of it based on what they search? I.e. the search client normalizes the query in some way and then requests only the relevant index pages. The index would probably be a lot larger on the server but if disk space is not an issue...?
(Though at some point you have to ask yourself what the benefits of such an approach are compared to a search server.)
(Though at some point you have to ask yourself what the benefits of such an approach are compared to a search server.)