The rest of the discussion is happening on the phab ticket, but i'll answer this here.  We are using a language detection algorithm that has been trained against tweets. Tweets are not, on average, as short as  the searches we are detecting the language of but it does an ok job. Trey did a great job putting together an analysis[1] of this language detection algo.  We will also be using his work there to evaluate other language detection methods and perhaps change what we are using in the future.

So the short of it is, we chose russian instead of serbian because the machine learning algorithm said so.

[1] https://www.mediawiki.org/wiki/User:TJones_(WMF)/Notes/Language_Detection_Evaluation