The math used to update the relevance score seems strange. In some cases, it doesn't add up. It sometimes appears that instead of incrementing the relevance score with the individual action scores, we are in fact replacing the current score with the individual score.
This needs to be researched more to confirm specific use cases that can be reproduced consistently, but I wanted to start this ticket so we do a bit more rigorous testing on this issue.
Version: unspecified
Severity: normal