Yeah... I'm just now realizing how relying on unit count in the listing is more problematic than I thought. Sellers say their unit count is 1, but the product title says it is 4 dozen. I need to figure out how to fix these inconsistencies.

ChatGPT was able to tell that there were 48 balls given only the URL...

Maybe add a "flag listing" button so users can alert you?

gemini flash!

Maybe you could build a heuristic around shipping weight? A single golf ball weighs about 45 to 50 g, so divide the shipping weight by, say, 50 g to account for boxing and so on and you get a rough estimate of the balls in the package.

You could make a list of all the metadata and pass it through a LLM to determine the quantity. You’ll need some sanity checking but if you prompt it with some examples it will do well. (Done something very similar myself.)

I’m not the person you replied to but I took a look at the data and this is an interesting one. You found a really cool data set and this will be fun.

Consider the top four most expensive golf balls on your current list:

TaylorMade 2021 TP5x (3+1 Box) 4DZ Golf Ball Pack, White — uses 4DZ in title, 48.0 in unit count in product specs.

Bridgestone Golf Tour B RXS Quadfecta - nothing in the title, unit count in product specs is 4.0. This one shows 4 dozen in a different spot than other balls.

TaylorMade Golf 2024 TP5 Golf Balls 3+1 Box Four Dozen — Four dozen in the title, unit count in product specs is 1.0 but it has 4.0 dozen in the same div as the Bridgestone balls.

Srixon Z Star Yellow Golf Balls - Buy 2 DZ Get 1 DZ Free — Title shows buy 2 DZ get 1 free. That’s represented as 2+1 or 3+1 in other data. In product specs it shows a unit count of 1.0.

— In that extremely limited sample, the product weight is a pretty good metric to show that the unit count is flawed though that only works in comparison to others. I wonder if you could do a multi pass approach, where you sort data first and then do a unit count versus weight check to find outliers and then start rocking through the titles? You’ll still end up digging through a lot of edge cases and that won’t be much fun but a multi pass would at least give you some insight into those weird edge cases.

The funny thing is, if your method becomes the dominant way of price discovery, a bad actor will simply try to circumvent the system to get their product ordered first, and you’ll be embattled in a Cold War.

See also: toilet paper sheet count comparisons.

what I've done for a similar script in the past:

    answer_initial = llm(prompt=prompt, site=site) # JSON with answer and any stuff needed to do heuristic checks.
    heuristic_results = heuristics(answer_final) # rule based.
    answer_final = llm(prompt-prompt, site=site, answer=answer_initial)
    mark_for_review = ... # basically just a bunch of hard-coded stuff I add flag possible failures for review.

You can use an extremely small/cheap model for something like this -- granite 4.0 micro works fine for me, 3.3 8b did as well, both run on my macbook. YMMV / try different models and see how it goes.

Alternatively, make your site so popular that manufacturers tune their description to work with your site :)

As much as I love simple deterministic things, this is a classic example of where NLP is better than hardcoding a list of keywords. Trying to guess every set of quantity keywords with various spelling, punctuation and how they interact ("1dzn box, two pack" is actually 24 balls) seems more brittle than an LLM.