Hi, I'm building a custom scorer for LTR module. Right now, I want to optimise performance of TF and DF computation in scorer class. The profiler shown that TermContext.build is a hot path, it looks like to take some amount of time to seek in a disk to find a statistic about terms.
Here is a sample gist of my LTR Feature implementation: https://gist.github.com/tanapoln/733f7e783ca3f9dd28702ddc48936f13 I also attached a profiling result screenshot in the gist too. I'm not sure that caching computation results of TF and DF to reduce disk seek is a good idea. Is there anyway to increase performance of computation? Best regards, Tanapol
