There is a specific and very common problem this solves, and if you have hit it you already know. You want a model to answer questions about a website. The website is HTML, full of navigation, cookie banners, footers, tracking scripts and layout divs, and if you hand that to a model you waste most of your context on rubbish and get a worse answer. What you want is the actual content, as clean text, with the structure preserved.
That conversion is annoying to do well, everybody rebuilds it badly, and this is a service that does it properly. The quickstart is the best thing here. Within a few minutes of the first page you have made a request and got back readable markdown, and that speed is not a trick of the documentation, it is genuinely how the product works. Anyone who has spent a weekend with BeautifulSoup and readability heuristics trying to get the same result will appreciate what has been saved.
The four core operations get explained clearly and the distinction between them is worth internalising because getting it wrong is how you waste money. Scrape takes one page. Crawl follows links and takes many. Map gives you the URLs on a site without fetching the content, which is the one people skip and then regret, because it lets you look at what you are about to crawl before you commit to crawling it.
Search combines a query with retrieval. Those are four different jobs with four different costs and the docs are clear about which is which. Output quality is the real product and it holds up. Navigation and boilerplate get stripped, headings survive, tables mostly survive, and links come through in a form you can still follow.
There is also structured extraction where you supply a schema and get typed fields back, which works well enough on consistent pages and less well on the messy ones, roughly as you would expect. The self-hosting documentation deserves credit. Plenty of companies open source a repository and then make the setup instructions just bad enough that you give up and pay them. This is a genuine guide, and for anyone with volume or data residency concerns that is a real option rather than a gesture.
Now for what is missing, and it is not a small thing. Scraping is a legal and ethical area, not just a technical one. Sites have terms of service. Some of those terms prohibit automated collection.
robots.txt exists and expresses a preference that you may or may not be obliged to respect depending on where you and the site are. Personal data pulled off a page is still personal data and does not stop being covered by data protection law because a model is going to read it. Copyright applies to text you copy at scale. None of this is complicated to raise and the documentation essentially does not raise it.
A developer following the quickstart is pointed at a URL field and left to work out the rest, and the honest position is that a meaningful number of the crawls people run with tools like this are not clearly permitted. That absence is a real mark against it as a teaching resource. Cost predictability is the practical complaint. Everything is measured in credits, and the number of credits a crawl will consume depends on how many pages it finds, which you do not know until it has found them.
Set a crawl going on a documentation site with a lot of versioned pages and you can eat a monthly allowance before lunch. Limits exist and you should use them, and the documentation could be considerably louder about setting them before your first crawl rather than after. The hard cases are still hard. Sites behind logins, aggressive bot protection, infinite scroll, content that only appears after user interaction.
There is support for some of this and the honest summary is that the failure modes are quieter than they should be. You get something back, it looks plausible, and it is missing half the page. Nothing warns you. A pipeline that silently ingests partial content is worse than one that fails, and the docs do not push you towards validating what you got.
The teaching, as with most vendor documentation, is product-shaped. You will finish knowing this API. You will not finish knowing how to build a crawling pipeline that handles retries sensibly, deduplicates content across near-identical URLs, detects when a site has changed its layout, or decides what is worth re-fetching and how often. Those are the things that separate a demo from something you can leave running, and they are not covered.
My three point seven is for solving a real and irritating problem properly, documenting it clearly, getting you working fast, and being honest enough about self-hosting to show you the exit. Marked down for near silence on the legal questions the tool makes it very easy to walk into, for cost behaviour that surprises people, and for quiet failures on the difficult sites. Good documentation for a good tool. Bring your own judgement about what you point it at.