Build a clean training corpus

Raw HTML is mostly navigation, cookie banners and footers. Feeding that to a model teaches it about cookie banners. Extraction returns the article and nothing else, as text or Markdown, with a word count so you can filter thin pages before they reach your pipeline.

• Each page reports a word count taken before any truncation, so thin pages can be filtered out cheaply.
• Up to 8 URLs per call, and a URL that fails to fetch is never billed.
• Markdown preserves headings, lists and links, which is the structure most chunking strategies depend on.