While we wait for the patch to land on production, I'm noting here some quirks encountered in the local dev environment:

• when there are multiple occurrences of a search result highlight, the text fragment directive seems to match the first. Examples:
  • how old is donald duck? > generated URL = https://en.wikipedia.org/wiki/Donald_Duck%27s_50th_Birthday?wprov=sscw1#:~:text=Donald%20Duck%27s-,50th%20Birthday > page title highlighted, should be the lead section body. This might happen often, given that the lead section usually repeats the title
  • who is the Spanish prime minister? > https://en.wikipedia.org/wiki/Prime_Minister_of_Spain?wprov=sscw1#:~:text=Pedro%20S%C3%A1nchez > infobox text highlighted, should be the lead section Pedro Sánchez, of the Spanish Socialist Workers' Party (PSOE) ...

I think this is due to suffixes not being handled deliberately.
Context terms are there to cope with such ambiguous fragment matches.

At the time of writing this comment, Beta cluster's semantic search results are the same as lexical search.
Examples:

• https://en.wikipedia.beta.wmcloud.org/w/index.php?title=Special:SemanticSearch&search=can%20fish%20fly%3F&ns0=1&limit=10&sort=relevance&noredirect=
• https://en.wikipedia.beta.wmcloud.org/w/index.php?title=Special:SemanticSearch&search=can%20bees%20talk%3F&ns0=1&limit=10&sort=relevance&noredirect=

Works fine on enwiki: https://en.wikipedia.org/w/index.php?title=Special:SemanticSearch&search=how%20old%20is%20planet%20Mars%3F&ns0=1&limit=20&sort=relevance&noredirect=

I think we should wait for the latest patch to land on production before we can close this ticket.

Congratulations @NBattista-WMF for your first ticket!

I do not see a need for deployers to have the git push ability at all.

Totally agree, and apologies for the trouble.
For context: my goal was to squash 3 commits of localization updates, which are expensive to backport. I mistakenly pushed to HEAD:wmf/1.47.0-wmf.21, instead of HEAD:refs/for/wmf/1.47.0-wmf.21.

Fix confirmed on prod:

Instrumentation QA

Source of truth: spec.

@matthiasmullie the patch is ready for review. CI is crying because I've added a test that uses CirrusSearch, see config patch.

See also https://gerrit.wikimedia.org/r/c/mediawiki/extensions/ReaderExperiments/+/1339812/comments/5df0a09e_8f76d80a

✅ double-checked on beta.

Backport done, verified it works: https://es.wikipedia.org/wiki/Madrid?mpo=pre-image-carousel-retest-aa:control-5

In T437487#12327808, @matthiasmullie wrote:

Note: namespace-prefixed search are already rerouted within search (read more)

@matthiasmullie I'm not sure I get this: the thread says the requested namespace can be controlled from the query, a user searching for "wikipedia: is it really free?" might unfortunately be routed to a classic fulltext search on the "wikipedia" namespace. Patch demo seems to follow prod.

Highlights ready & implemented!

This ticket has become outdated.

Published notebook.
@MGerlach , I've tried to integrate the points in the ticket description. Would you mind signing this off as well, please?

@MGerlach , would you mind signing this off, please?

• new experiment’s machine-readable name is pre-image-carousel-retest-aa
• if the current page gets the carousel, then call TK’s send exposure - roughly here I think
• exposure should be sent to all groups, i.e., control, and control-2 to 5
• also send page_visit alongside. We could re-use anyPageVisit

Published notebook and asked Research for a review.

Metric backfilled, only applicable to enwiki.

Analysis published at https://gitlab.wikimedia.org/mfossati/scriptz/-/blob/main/T435301.ipynb.
See the last section for takeaways & recommendations.

QA instructions work as expected.