At the beginning of August, Platform team ran the image suggestions model via T285816: Add an image: generate static file of suggestions, and Search team loaded those suggestions to the search index via T285817: Add an image: load static file to search index. Our hope was that the articles and suggestions from August would still be sufficiently fresh for a release in October/November.
Now that we are nearing release, we see that a sizable number of articles with image suggestions in the search index (especially from Bengali Wikipedia) are now illustrated. This may have happened because of concerted efforts by the communities to illustrate articles, unrelated to our work. In short, the modeled suggestions have gotten stale faster than we expected.
We always knew that the modeled suggestions would grow stale, and so the Growth team's interface handles these moments by telling the user that the suggestion is actually not available after they click on it. But we don't want this to happen frequently to users, as it will make for a poor experience.
Therefore, we would like to refresh the apparatus in advance of our deployment, if possible. Our deployment is November 29. After talking with @sdkim, it sounds like the main things that would need to happen are:
- Re-run the model.
- Make the refreshed suggestions available for the image suggestions API
- Load the refreshed suggestions to the search index
On this task, we want to discuss whether such a refresh would be simple enough to do before November 29.