This new search engine is definitely not an improvement. If I search on “Leap Year Day” what shows up on the preview page has 2 images that are appropriately on topic.
But when you click on “see more” (in other words the new and “improved” AI powered search) the results are all wrong. I didn’t find an actual Leap Year Day image until the 3rd page. I didn’t find another one even though I went through 10 more pages.
Even though that image is actually Trending right now.
The same is true if I search on Feb 29. Good result on the old search engine preview page, but pages and pages and pages of irrelevant and incorrect images with the new.
@BenLeong @KingDog is there anyway for us to submit feedback to the team working on this besides the “rate this search” option? It doesn’t allow me to submit screenshots to show what I mean.
6 Likes
yep its just not working the old search engine was much better
2 Likes
@MargJohnsonVA Thanks for raising this - you can send me more details (screenshots etc) in a DM, and I can make sure those get to the team that has been working on this.
The initial release of Vector Search for photos has generally been very well received by customers, but there will always be edge cases that we can fine-tune the system for. The Feb 29 / Leap Year Day search looks like one of those cases, so we can investigate further to understand what’s causing it.
A good example of how the new search works can be seen for photo searches related to the background term. This is one of the most popular topics for search by a huge margin (it’s roughly the same volume of all photo searches relating to “business” “office” and “texture” combined), but the old search produced a very haphazard selection of search results.
Vector search takes the text entered as a search term, and transforms this into a group of compound queries that can be used in our systems, to find images that match them. Each part of those queries can be tweaked, increasing or decreasing the weighting given to it - for example, we know that a lot of “background” image searches are actually looking for what we’d consider computer or phone wallpaper graphics, so we can boost the relevance of that part of the search, relative to the other parts.
The search results from doing that look a lot more consistently relevant, and that’s reflected in the search-to-download ratios for that content: we can see that customers searching for “background” images are now far more likely to license an image they find in their search results.
3 Likes
Thanks for the followup! You have Twenty20 to thank for the “background” issue. They encouraged the Insiders to add it to all photos. 
