I have recently fixed #385 (closed), by modifying the query which was calculating the score on document, filtering those out to not include trashed documents.

That fixed the issue, but I have actually spent considerably more time trying to write a test for it than code the actual feature; that smells like something is wrong here: if our tests are so complicated and error-prone to write, we either need a better testing infra, or we need something more on the backend side.

Either way this issue is about tracking time and bringing to completion this MR, while trying to understand why at the moment testing ngrams scores is so tricky.