Topic on Talk:Cross-wiki Search Result Improvements/Flow

The wiki blocks are ordered by recall (most to least number of articles returned from each project). . Large wikis are likely to be ordered first frequently. Concerning wikivoyage there's a small variation, a filter wikivoyage results on title.

Commons uses boosted-templates (https://phabricator.wikimedia.org/T163223).

I just tried out searching for some popular travel locations, and Wikivoyage came out last or next to last (i.e. below the fold) in each of these examples:

• "Cape Cod" on enwiki
• "Mallorca" on dewiki
• "Hawaii" on enwiki
• "Rome" on enwiki
• "Alaska" on enwiki(voy:Alaska ranks below q:Alaska Thunderfuck 5000 and b:Solitaire card games/Alaska)
• "Ko Samui" on enwiki (voy:Ko Samui ranks below wikt:come and q:Urusei Yatsura)

(Examples are not cherry-picked, these were just the first few that that came to my mind.)

It seems to be pretty well documented:

https://wikimedia-research.github.io/Discovery-Search-Test-CrosswikiSidebar/

https://commons.wikimedia.org/wiki/File:Second_Test_Of_Cross-wiki_Search_-_Helping_More_Users_Discover_Content_On_Wikipedia%E2%80%99s_Sister_Projects.pdf

https://phabricator.wikimedia.org/T149806

(Props to the project managers and analysts for reporting it so well).

It will probably more interesting to read the outcomes and analysis (spoiler: it is not random).

Projects should – in theory – be ordered according to recall (most to least number of articles returned from each project). And this is mostly true if you open each sister project's search results page separately.

Looking at InterwikiSearchResultSetWidget in MediaWiki Core, it does not appear there is explicit front-end code for ordering the projects when the SERP is rendered and the order is determined by Cirrus (IIRC) on the back-end when it returns the interwiki results – which should be according to recall.

Looking at InterwikiSearcher.php in Cirrus source, we still have code from the first cross-wiki search A/B test where results can be returned in a random order if the configuration requests it, but they should be returned according to recall in production. Although maybe we're accidentally still using the static order that the switch statement defaults to. I'll reach out to @EBernhardson (WMF) and @DCausse (WMF) for clarification.

Absolutely, the wiki blocks are ordered by recall. Large wikis are likely to be ordered first frequently. Concerning wikivoyage there's a small variation. During the RFC it was requested to strongly filter wikivoyage results on title. Today we ensure that 80% of the search terms (stop-words excluded) appears in a title for wikivoyage results. In other words it decreases recall for wikivoyage and probably one of the reason you feel that wikivoyage is ranked so badly. Without the title filter wikivoyage would be ranked #3 (just below wiktionary) for the query Alaska.

Might be good to expand the documentation :Extension:CirrusSearch/Scoring#Cross-wiki as this will be asked again ...