Content types
13 types, assigned at index time from page structure and markup rather than inferred from the URL — so the filter does what it claims.
Documentation
docsOfficial product, API and library documentation — reference material maintained by whoever built the thing, not third-party write-ups of it.
Code
codeSource files, repository content and code-bearing pages: the implementation itself rather than an article describing an implementation.
Academic
academicPapers, preprints, theses and conference writing, identified from document structure rather than from the domain that happens to host it.
News
newsReported news articles carrying a publication date and an identifiable outlet — reporting of an event, as distinct from commentary on one.
Article
articleLong-form editorial writing that is not news reporting: essays, analysis, features and engineering blogs. The general-purpose editorial bucket.
Reference
referenceEncyclopedic and reference material — definitions, tables, standards and specifications, written to be looked up rather than read through.
Q&A
qaQuestion-and-answer pages: one stated problem and one or more proposed solutions, with the answer structure visible in the page markup.
Forum
forumThreaded discussion — mailing lists, message boards and community threads, where the value is in the exchange rather than in any single post.
How-to
how-toProcedural content: step-by-step instructions for accomplishing a task, as opposed to reference material explaining how something works.
Product
productProduct and service pages, including pricing, specifications and feature listings, published by whoever sells the thing being described.
Opinion
opinionExplicitly argumentative writing — columns, editorials and position pieces, where the author is arguing a case rather than reporting an event.
Video
videoPages whose primary content is video, indexed through their text metadata and transcripts rather than through the video itself.
Document
documentStandalone documents published as files: PDFs, reports, filings and papers, where the file rather than the page is the thing you wanted.
Structure, not the URL
Content type is derived at index time from how a page is built — its markup, its structured data, the shape of its headings and its body — rather than from a rule about the path it lives at. That distinction is what makes the filter dependable. A documentation page hosted on a company blog is still documentation, and a marketing page sitting under /docs/ is not, and a URL-pattern rule gets both backwards.
It also means the assignment is a judgement that can be wrong, most often on pages that genuinely mix two forms: a tutorial that is half reference, a product page carrying a long technical appendix. Where a page could reasonably be two things, expect it to be filed under the one its structure most resembles.
Why this is the highest-leverage filter
For an agent, changing the content type changes the corpus rather than the ranking, and that is a much larger effect than any amount of query rewriting. Asking a general index about a library API returns a great deal of tutorial content paraphrasing documentation, frequently against an older version; restricting to documentation removes the entire category of that error in one parameter.
The corresponding risk is over-restriction. A type that is too narrow returns a thin set and the agent has no way to tell that from a subject nobody has written about. Probing the distribution first — counts and facets with no results at all — is the cheap way to find out which it is before you commit to the filter. Thetopic tags are the natural second half of the pair.
Narrow to the shape of the answer
10,000 free credits a month. The live vocabulary is served by /describe, so an agent can read it rather than hardcoding it.