<< All versions

Skill v1.0.0

currentLLM-judged scan90/100
opensourceagi/qwksearch-research-agent/ask-extract-cite
──Details
PublishedOctober 21, 2026 at 10:27 AM
Content Hashsha256:9820ca2b055e07ca...
Git SHAa544f53d79cb
──Files
Files (1 file, 3.7 KB)
SKILL.md3.7 KBactive
SKILL.md · 54 lines · 3.7 KB

version: "1.0.0" name: ask-extract-cite description: Guide to extract-cite (packages/extract-cite) — regex/metadata citation extraction (extractCite) plus extractCiteLLM, an OpenRouter-by-default LLM pass that completes the full APA citation, scores confidence per part, flags parts for review, extracts author qualifications and formats APA/MLA/Chicago/Harvard/IEEE/BibTeX. Use when a citation is wrong or incomplete, when adding a citation style, or when changing the prompt or the checks on the model's reply.

Working With extract-cite

packages/extract-cite, published as extract-cite. It owns the citation code that used to sit in extract-webpage/src/html-to-cite; extract-webpage imports it and re-exports extractCite, convertURLToDomain, isURLValid.

Picking the right call

You wantCall
Fast, offline, partial citation from HTMLextractCite(documentOrHTML, { url })
Full citation + confidence + review flags + author biosawait extractCiteLLM({ url, apiKey, model })
Same, for a page you already holdextractCiteLLM({ html, url }) or { text }
Another providerbaseUrl — any OpenAI-compatible chat-completions API
Person-vs-org detection with the 92k names DBawait loadHumanNamesDB() (lazy, from CDN) or import extract-cite/full (bundled)
One style from existing dataformatCitation(citation, "mla")
Is the extracted article complete, or a paywall stub? What to cut?extractCiteLLM({ html, url, content: article.html }) → contentCheck

How the LLM pass works (src/llm)

  1. extractCite gives the partial; missingFields names what it cannot supply.
  2. preparePage builds the model input: head + tail of the text, citation <meta>, JSON-LD.
  3. callLLM — one chat/completions call, temperature: 0, JSON object reply. Defaults: OpenRouter, anthropic/claude-haiku-4.5, key from apiKey or OPENROUTER_API_KEY.
  4. parseCitationReply normalizes and checks the reply (see below) and builds needsReview.
  5. formatCitations writes the requested styles.
  6. Content check (content-check.ts, on unless checkContent: false): prepareContent sends the first contentWords (3000) words of content (default: the page text) and pageOutline, the page's blocks as selectors; CONTENT_CHECK_PROMPT is appended to the system prompt; parseContentCheck returns contentCheck (verdict, signals, note, tips, contentSelector, selectors in the per-domain-selector shape). Not full → a content item in needsReview.

Checks that must stay

Author names must appear in the page text (else capped at 0.2 and flagged); qualification evidence must be quoted from the page (else the qualifications are dropped and flagged); dates must be real YYYY[-MM[-DD]]; strings lose markup; a missing author/title/date is flagged, not invented. Regex/LLM agreement adds 0.1. Content-check signals and examples must appear in the text, and selectors must be valid and match an element on the page, or they are dropped.

Gotchas

SymptomCause
Three-word person name cited as an organizationSlim entry: no names DB loaded. await loadHumanNamesDB() (CDN) or import extract-cite/full.
extract-webpage tests can't resolve extract-citeIt is consumed as built dist. bun run build in packages/extract-cite (CI does it).
401 No API keyNeither apiKey nor OPENROUTER_API_KEY is set.
NYT and other sites 403They block server fetches. Fetch the HTML elsewhere and pass html.
Site typecheck fails after changing types.tsUpdate extract-pdf/site/types/extract-cite.d.ts.

Demo: Citation tab of packages/extract-pdf/site (/demo#cite, POST /api/cite).

All versions