The search engine can’t even return an exact match. How did you release FAB in such a poor state?
I bet they hired the guy who did the Start menu search in Win 10/11 
5 Likes
Apparently the spec was: “show anything that doesn’t match what was typed”.
I think they nailed it 
6 Likes
It’s been like this since yesterday, if you type more than one word you’ll get mostly random stuff. I really hope they prioritize fixing this because text search is pretty much unusable now. The only way to find relevant results now is to narrow them down by clicking on the categories bellow the search field.
1 Like
Yeah I searched one of my products and it doesn’t appear at the top anymore, you have to scroll for a bit to find it despite being an exact name search. The first result isn’t even anything with the name I typed.
1 Like
in general, “full text search” engines actually do poorly on things that aren’t individual words.
For example, here there’s an expression “Will’s” which has two problems:
- it’s not all lowercase
- it contains an apostrophe
Separately, if they use a “semantic search” type solution (embedding based search) then the character match doesn’t even matter; the contents of the listing are embedded as one semantic (normalized) vector, and the contents of the search is embedded as another semantic (normalized) vector, and similarity is scored based on the dot product between these two vectors. The only difference to 3D dot products is that embedding vectors typically have 768-3072 components, not just three …
So, I can see a bunch of ways they’d implement “improved” search and actually break the basic cases, although for all of those ways, they have to be overly enthusiastic about ONLY supporting the new way, and not having any integration or acceptance tests against typical normal search cases. Which, you know, because they are game developers, is not unheard of. I love game devs, but automated testing isn’t commonly thought of as a core tenet of most game developers …
(After all, you cannot unit test “fun”)
1 Like
I didn’t understand anything you say hahaha
But look at this on Netflix, it’s the first results, it shouldn’t be that difficult
3 Likes
I have to say, it is terrible. I don’t think they ever tested it. It can’t even find the most basic things. Is it just lack of time, or another example of money targets steering the functionality away from what is actually required to keep the product usable?
Being able to block certain vendors would also be nice.
I’m particularly talking about these ‘sweatshops’ flooding the site with AI garbage. At the moment, audio.
Even if they just gave us a text doc about the possible parameters, that would be enough…
3 Likes
Nobody who is actually trying to optimize revenue (money) would ship a new search algorithm without testing it. If that were the goal, they would absolute A/B test even the smallest change in the algorithm for maximum revenue generation. So, we can be almost 100% certain that’s not the goal for this change.
So, there’s a few possibilities:
- They implemented semantic search as a replacement for text search, and didn’t comprehensively test it. “The feature list from above says semantic search, this is semantic search, please keep giving me good performance reviews!”
- They use some text search engine that’s bad at words that aren’t just plain lowercase un-stemmed words. Seems unlikely, but anything’s possible.
- They shipped some bug, either that ignores the search match score at all, or does something bad with it (like sorting a number as if it’s a string or whatever.)
It’s obviously they didn’t:
4. Run a comprehensive acceptance test suite as part of shipping, and act on the results.
FWIW, the word “will’s” alone doesn’t match anything at all, so it’s actually quite plausible that they’re in timeline 2. But, who knows?
2 Likes
Yes, agreed, money can’t be the goal, unless they’re really badly off target 
Whichever timeline they took, the target has been left in the bushes 