Similar to @mpopov recent work comparing our zero result rate to other search engines, we should look over the discernatron data to see if there are particular query features that we should be taking into account to return better results.
- Are the high ranked results in our top 20, 100, or 1000? Are we just giving them too low of a score?
- If the high ranked results are not in the top 100, or 1000, why not? Synonyms? Spell correction? Relaxing query constraints?
- Probably others, these are just a couple off the top of my head.