I am really curious how do people actually evaluate scrapers?
It is not easy to evaluate scrapers unless you have had to deal with lots of poorly written websites in your life. Just using it on a few highly structured well-maintained sites can be impressive but if you are using it to acquire data from many websites or large websites things get hairy fast.
Most scrapers today are some combination of extracting xpaths and reducing them to the loosest common form, parsing semantic (or easy to identify, like links) or highly structured content that has discoverable patterns, and LLMs.
The actual best way to scrape a site is to determine if they are populating the data you want with API calls and replicate those. People are usually more reluctant to completely change back-end code but will make subtle breaking changes (to your scraper) to front-end code all the time. For example small structural or naming changes. This has become more problematic since people have been moving to more SSR and semi-SSR injection. There can also be a problem with discovering all the pages on a site if it doesn't have poorly designed or implemented paging or search.
Some of the worst sites to scrape are large WP sites that have obviously been through a few developers. If you really want to test a scraper find some of those and they will put it to the test.
Cloudflare is another issue. Not necessarily an issue with this plugin, but because so many sites use it, you typically have to spin up multiple automated headless browsers using residential proxies for any type of large-scale scraping.
Some things that LLM does shine at related to scraping is interpreting freeform addresses, custom tables (meaning no TR,TD, but just divs and CSS to make it look like a table, and lists that are also just styled divs. Often there are no tags, attributes, keywords, or generalized xpath that will help you depending on how the developer put it together.
Surprisingly there is a pretty old library from Microsoft of all places called Prose if you can still find it (they keep updating but using the same name for different things and trying to inject AI) that is really good at pattern matching and prediction and is small, fast, and free and generally great at building a generalized scraper. Only drawback is I believe the only one I could find was .NET at the time.
Comments
It is not easy to evaluate scrapers unless you have had to deal with lots of poorly written websites in your life. Just using it on a few highly structured well-maintained sites can be impressive but if you are using it to acquire data from many websites or large websites things get hairy fast.
Most scrapers today are some combination of extracting xpaths and reducing them to the loosest common form, parsing semantic (or easy to identify, like links) or highly structured content that has discoverable patterns, and LLMs.
The actual best way to scrape a site is to determine if they are populating the data you want with API calls and replicate those. People are usually more reluctant to completely change back-end code but will make subtle breaking changes (to your scraper) to front-end code all the time. For example small structural or naming changes. This has become more problematic since people have been moving to more SSR and semi-SSR injection. There can also be a problem with discovering all the pages on a site if it doesn't have poorly designed or implemented paging or search.
Some of the worst sites to scrape are large WP sites that have obviously been through a few developers. If you really want to test a scraper find some of those and they will put it to the test.
Cloudflare is another issue. Not necessarily an issue with this plugin, but because so many sites use it, you typically have to spin up multiple automated headless browsers using residential proxies for any type of large-scale scraping.
Some things that LLM does shine at related to scraping is interpreting freeform addresses, custom tables (meaning no TR,TD, but just divs and CSS to make it look like a table, and lists that are also just styled divs. Often there are no tags, attributes, keywords, or generalized xpath that will help you depending on how the developer put it together.
Surprisingly there is a pretty old library from Microsoft of all places called Prose if you can still find it (they keep updating but using the same name for different things and trying to inject AI) that is really good at pattern matching and prediction and is small, fast, and free and generally great at building a generalized scraper. Only drawback is I believe the only one I could find was .NET at the time.