I can't tell you how many times I've had the feeling that finally I was doing anything at all that was original and then to find out (usually within five minutes, but this time at least a lot of years later) that it was already old hat when I was a toddler.
Oh well. Keep trying I guess.
One problem with stuff like this is simply to find what is out there given a problem description.
At least.
If I would have found it at all. This is one of those areas where I think I can really feel that having English as a first language - or at least a degree in computer science - would have helped tremendously, just knowing the right terminology can be such a huge advantage.
I can parse your description of it but I could never come up with it in those terms. I'd be looking for "compressed index" or "candidate records algorithm" or something like that.
Obviously, you can't really do it so I first pre-sorted the hotels with some common criteria in large chunks, then sampled random bits from the rest of the criteria to fill the memory based portion of the index. When a query would come in a mask was made for the various criteria in the same way as during the indexing, and this mask was then used to get the indices of hotels that might be right, candidates so to speak. Then the candidates would be loaded from disk and checked against all the criteria and an answer would be returned to the caller. It worked pretty good, the disk block caching mechanism really saved the day.
This was on a QnX system before they had a 32 bit/large model version...
That's when programming is fun!
Of course 64K is still not really 'constrained' ;)