3 ms·
I scrape website content regularly (usually as one-offs) and have a hand-crafted extractor template where I just fill in a few arguments (mainly CSS selectors a
by thrdbndndn 1y ago
I scrape website content regularly (usually as one-offs) and have a hand-crafted extractor template where I just fill in a few arguments (mainly CSS selectors and some options) to get it working quickly. These days, I do sometimes ask AI to do this for me by giving it the HTML.
The issue is that for any serious use of this concept, some manual adjustment is almost always needed. This service says, "Refine your scraper at any time by chatting with the AI agent," but from what I can tell, you can't actually see the code it generates.
Relying solely on the results and asking the AI to tweak them can work, but often the output is too tailored to a specific page and fails to generalize (essentially "overfitting.") And surprisingly, this back-and-forth can be more tedious and time-consuming than just editing a few lines of code yourself.
Also if you can't directly edit the code behind the scenes, there are situations where you'll never be able to get the exact result you want, no matter how much you try to explain it to the AI in natural language.
- throwup238 1y agoI’ve had no shortage of trouble using LLMs for scrapers because for some reason they almost always ignore my instructions to use something other than the class name for selectors. They love to use the hashed class (like emotion/styled/whatever css-in-js library de jour) names that change way too often.