[swalsh]

#2 is what would transform this from a nice niche tool, to something very valuable. In the ecommerce space, tracking competitor pricing is a great example of this type of thing. I can also see use casese for recipe's, finance, healthcare, you name it. Those b2b use cases are worth real money.

Just curious, in your experimentation, have you found it necessary to train a new model for each "category"? Or have you found a way to generalize it?

[tsergiu]

Training a new model for each category is already possible today, but doesn't achieve the goal (mass-specialization).

The problem is that when you pre-train a model, you can only solve for the lowest common denominator of what every customer might want.

In ecommerce, for example, you might pre-train to get price, product name, reviews, and a few other things that are general to all ecommerce. But you won't pre-train it to get the mAh rating of batteries, because that's not common to the vast majority of customers (even within ecommerce). It turns out that most customers need at least a few of these long-tail properties that are different than what almost every other customer wants, even if most of the properties they need are common.

And so the challenge is to dynamically train a model that generalizes to all "battery sites" based on the (very limited) input from a customer making a few clicks on a single "battery site".

[colinchartier]

I worked on this for a long time -

1. it's possible to make it "easy to switch" by having common building blocks and only changing the "selector" across sites - lots of companies in the space do this

2. it's impossible to do "just DOM" or "just vision/text" if you want to be able to generalize "get the price of the items"

- DOM doesn't represent spacial positioning very well (see: fixed/absolute positioning, IDs and dom changing without the visuals changing, ...) so you'd need the equivalent of an entire browser rendering engine in your "model" anyways!

- vision/text is messed up by random marketing popups (see: medium, amazon, walmart, ...), it's significantly more computationally expensive to do, and can't currently get >95% accuracy (which makes it useless, scraping needs very close to 100% accuracy in most use cases)