Hi,
I'd like to prepare some personalized search. Let's say there is a user
vector which describes a long term profile of the user. There are some
values which ranks parameters used in queries and there is a function which
uses parameters to compute dynamic ranking (paramets are just index fields
- eg. Integer, Float). I tought that I can use boost query to prepare such
a ranking ({!boost b=$bb v=$qq}, where $bb is boost function and $qq is
query). After test I realized that using boosts is much "heavier" than just
make standard queries. The problem is that I have about 50milion of
documents and about 3000rps. So if query will be much more CPU consuming it
will be expensive to make such a search system.
Moreover I noticed that making scoring cutoff (some filter that makes
frange on scoring - {!frange l=<some_value>}query($qq) ) makes queries much
slower.
I also tought that I could prepare some indexes with precomputed values for
ranking for many groups of users. I couln't make index per each user
becuase there are milions of users - so I think it is bad idea.
Did somebody have a chance to prepare personalized ranking for quite heavy
loaded system? Do you have any advices?
--
Regards,
Pawel
