On Tuesday 20 June 2006 18:42, Dan Climan wrote: > >Paul Elschot <[EMAIL PROTECTED]> > >>On Tuesday 20 June 2006 12:02, Marcus Falck wrote: > >> After a lot of debugging and some API doc reading I have come to the > > conclusion that the static encodeNorm method of the Similarity class > > will encode my boost value into a single byte decimal number. > >> > >> And I will loose a lot of resolution and will get severe rounding > >> errors. > >> > >> Since I need the exact float value as boost representation this isn't > >> good enough in my case. > > >An exact float value is a bit of an oxymoron. > >How exact do you need it to be? > > >The range of values that can be encoded by the existing encodeNorm() > >and decodeNorm() is quite big (about 10e-10 to 10e+10 iirc), > >and since there are only 255 possible values in there (excluding 0), > >the rounding errors can be severe indeed. > >However, with a smaller range, the rounding errors would also be smaller. > > >Are 256 different values enough for your case? > > >Regards, > >Paul Elschot > > I, too, have found that 256 values were not enough. I tried changing the > encodeNorm function to use a narrower range of values, but the 256 limit > makes it degrade quickly if I get any results outside the expected range. > This was true when we tried various algorithms for boosting results based on > external factors. > > FunctionQuery(not currently in core lucene) from the SOLR project may be be > an alternative. Can it replace all uses of the norm? > > Now that omitNorms is part of the core, the impact of allowing a 2 byte (or > even 4 byte norm) is not nearly as severe on memory. Any suggestions for how > to create a multi-byte norm as an option, but enable the same code to > reading existing 1 byte norms?
No more than to add your needs here: http://wiki.apache.org/jakarta-lucene/FlexibleIndexing, and to subscribe to java-dev. Regards, Paul Elschot --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
