Cite with an LLM

import { extractCiteLLM } from 'extract-cite';

const cite = await extractCiteLLM({
  url: 'https://www.npr.org/2023/12/28/1221827923/2023-hottest-year-record-climate-change',
  apiKey: process.env.OPENROUTER_API_KEY,
  model: 'anthropic/claude-haiku-4.5', // optional
});

cite.formatted.apa?.text;   // the full APA 7 entry
cite.needsReview;           // [{ field, confidence, reason }]

The regex pass (extractCite) is instant and free but partial: it cannot tell a byline from a job title, never finds a DOI, and misses dates a page only writes in prose. extractCiteLLM runs it first, then sends the model the page text, its citation <meta> tags and JSON-LD, the partial citation as candidates to verify, and the list of parts still missing.

OptionsCopy Anchor Link

What comes backCopy Anchor Link

• citation: every part a bibliography entry needs: authors (given, family, suffix, or an organization), title, containerTitle, publisher, publishedDate (as precise as the page is), doi, volume, issue, pages, sourceType.
• confidence: a 0–1 score per part, and overallConfidence.
• needsReview: the parts to check by hand, each with a reason: low confidence, missing, an invalid date, an author the page never prints, or the model asking for review.
• origin: whether the regex pass and the model agreed (both) or the model supplied it (llm). Agreement raises confidence.
• partial and missing: what the regex pass found and what it left for the model.
• formatted: one { text, html } per style. The HTML is escaped apart from <i>.
• contentCheck: whether the extracted content is the full article. See below.

Content checkCopy Anchor Link

The same call reads the first contentWords (3000) words of content, which is the article you extracted or, by default, the page's own text, along with an outline of the page's blocks as CSS selectors, their word counts and opening words. It returns:

• verdict: full, paywalled (a subscribe or sign-in wall cuts the body), truncated (a teaser, abstract or "Read more" with no wall), blocked (bot check, consent wall, error page), not-article (a home page or listing), or unknown when the model gave none. isFullContent is true only for full.
• signals: the phrases that show it, such as "Subscribe to continue reading", and a note for the reader.
• tips: clutter the content still holds (header, nav, sidebar, footer, ads, related, newsletter, share, comments, cookie-banner), each with the selector that holds it, an example and a one-line tip on cutting it.
• contentSelector: the element that holds the article body.
• selectors: { content, remove }, the shape of an entry in extract-webpage's extract-selectors-per-domain.json, so a site's tips can be pasted in as a rule for the Readability and Mercury extractors.

Signals and examples must be in the text and selectors must match an element on the page, or they are dropped. Anything but full is also added to needsReview as field: "content". Pass checkContent: false to skip it.

Author qualificationsCopy Anchor Link

For each author the page describes, citation.authors[n].qualifications holds jobTitle, affiliation, credentials, expertise, highlights and evidence, the bio sentence copied from the page. It is read in the same call as the citation, from the author bio, byline tagline or JSON-LD, and only from those: a bio quote that is not actually on the page discards the qualifications and flags them.

Checks on the modelCopy Anchor Link

The page is untrusted text, so the prompt fences it and tells the model to ignore instructions in it. The reply is never trusted as it comes: strings are stripped of markup and length-capped, dates must be real, an author whose name the page never prints is capped at 0.2 confidence and flagged, and qualifications need a verbatim quote.

Re-style without another callCopy Anchor Link

formatCitation(cite.citation, 'mla') and formatCitations(citation, styles) are pure and offline.

Hosted demoCopy Anchor Link

The Citation tab of the live demo runs the content check on the article extract-webpage pulls out of the page, and has presets for news (New York Times, NPR, Scientific American), academic (arXiv, Nature), organization and reference pages (WHO, Wikipedia) and an essay blog. See its API.