Hi Rashmi, Relevancy needs some kind of training data which can lead to a chicken and egg problem. If you dont have that training set, then you need to come up with it or train manually (provide some seed). Our existing search had 2 years worth clickstream data, i.e. we know if someone searches for "ipod" they clicked on a UPC which was an iPod 4th gen or an iPod 5th gen 32GB etc.
So, we have used that data to build an internal lookup table of millions of queries which look something like this: ipod 32gb -> music^1000, apple^1000, 32gb^991, 8gb^800.... We wrote an algorithm which computes the "keyword relevancy score" which is used as the boost value. Now, when a query like "ipod 32gb" comes in, we lookup this table, get the boost values and query solr with these boost values and its score. We are happy with the results. Our usecase was product search (title+description) of about 60M documents, not sure how will this approach work with a different usecase. Thanks, -Utkarsh On Tue, Jan 28, 2014 at 9:22 AM, tamanjit.bin...@yahoo.co.in < tamanjit.bin...@yahoo.co.in> wrote: > You may also want to look here > <http://wiki.apache.org/solr/SolrRelevancyFAQ> > > > > -- > View this message in context: > http://lucene.472066.n3.nabble.com/implement-relevency-tp4113964p4113983.html > Sent from the Solr - User mailing list archive at Nabble.com. > -- Thanks, -Utkarsh
