Live demo

The site lives in the monorepo at packages/extract-pdf/site. It is one Cloudflare Worker that serves:

worker/index.ts is the Worker's entry: it answers /api/* from worker/api.ts and hands every other request to the docs.

The same site also deploys to Vercel as a plain Next.js app (vercel.json, build-vercel.sh): the docs and /demo are the same pages, and app/api/[...path]/route.ts answers every /api/* route with the same worker/api.ts. Set the variables below as Vercel environment variables; there is no KV there, so settings saved in /admin last only as long as one function instance.

The PDF APICopy Anchor Link

# Upload a file
curl -F file=@paper.pdf https://YOUR-SITE/api/convert

# Raw PDF body, with page-number markers
curl -X POST -H 'Content-Type: application/pdf' --data-binary @paper.pdf \
  'https://YOUR-SITE/api/convert?addPageNumbers=true'

# Remote URL
curl 'https://YOUR-SITE/api/convert?url=https://arxiv.org/pdf/1706.03762'

The response:

{
  "source": "https://arxiv.org/pdf/1706.03762",
  "status": "ready",
  "title": "…",
  "author": "…",
  "html": "<p id=\"page-1\">…",
  "bytes": 2215244,
  "ms": 1377,
  "ocr": {
    "needed": true,
    "pageCount": 15,
    "pagesNeedingOcr": [3, 7],
    "reasons": { "3": ["table-caption", "numeric-grid"], "7": ["sparse-text", "figure-caption"] },
    "enhance": null
  }
}

ocr is the text-layer scan: the pages that look scanned or carry tables and figures, and why. When pages were flagged and OCR is configured, status is "ready_with_pending_ocr" and enhance is { "endpoint": "/api/enhance", "maxPages": 10 }.

Errors are { "error": "…" } with 400 (bad input), 413 (too large), 422 (unparseable PDF) or 502 (the URL could not be fetched).

The OCR follow-upCopy Anchor Link

Workers have no canvas, so they cannot rasterize pages, and a vision model is far too slow to run inside a request anyway. The demo splits the work:

1. POST /api/convert returns the text-layer HTML at once, with the flagged pages.
2. The page renders only those pages to PNG in the browser with PDF.js. For a URL source it reads the PDF again through /api/source.
3. It sends each image to POST /api/enhance as { "page": 3, "imageBase64": "…" }. The Worker forwards it to the Docling processor with its secrets and returns { page, html, ms }.
4. The page swaps each OCR'd page into the result as <section class="ocr-page" id="page-N">, one at a time. A page whose OCR fails keeps its text-layer version.

Page images pass through the Worker and the Space in memory only: neither stores them, and every API response is Cache-Control: no-store.

When OCR is configured, the PDF tab calls POST /api/warmup as soon as it opens. The Worker asks the Space to start loading its model and answers within a few seconds with { ocr, modelLoaded, loading }, so a Space that was asleep is booting while the visitor picks a file.

To turn it on, deploy the Docling Space and give the Worker its URL and token:

bunx wrangler secret put DOCLING_PROCESSOR_URL   # https://YOUR_HF_USERNAME-extract-pdf-docling.hf.space
bunx wrangler secret put DOCLING_API_TOKEN       # the Space's DOCLING_API_TOKEN secret
bunx wrangler secret put HF_SPACE_TOKEN          # private Space only: a Hugging Face read token

Without DOCLING_PROCESSOR_URL the demo still works and only reports which pages would benefit from OCR. Optional tuning variables: DOCLING_MAX_PAGES (10), DOCLING_MAX_TOKENS (1500) and DOCLING_MAX_IMAGE_MB (8).

The webpage APICopy Anchor Link

curl 'https://YOUR-SITE/api/extract?url=https://en.wikipedia.org/wiki/Web_scraping'

# HTML you already have; the URL makes links absolute and helps the citation
curl -X POST -H 'Content-Type: application/json' \
  -d '{"html": "<html>…</html>", "url": "https://example.com/post"}' https://YOUR-SITE/api/extract

The response is extractContent's article plus where it came from: { input, ms, title, author, author_cite, date, source, cite, url, word_count, html }. Flags images, links and formatting default to true. Pasted HTML is capped by MAX_HTML_MB (2).

The citation APICopy Anchor Link

curl -X POST -H 'Content-Type: application/json' \
  -d '{"url": "https://arxiv.org/abs/1706.03762", "apiKey": "sk-or-…", "model": "anthropic/claude-haiku-4.5"}' \
  https://YOUR-SITE/api/cite

apiKey is an OpenRouter key, used for that one request and never stored. Leave it out when the Worker has its own (bunx wrangler secret put OPENROUTER_API_KEY; CITE_MODEL sets the default model). html can replace url for pages that block server fetches, and styles narrows the output. The Worker first runs extract-webpage on the page and passes the article as content, so contentCheck says whether that extraction is the full article or a paywall stub, and which blocks to cut. The response is extractCiteLLM's result plus input. Errors are { "error": "…" } with 401 (no key, or the provider rejected it), 429 (the provider's rate limit), 400, 413 or 502.

Run it locallyCopy Anchor Link

bun install                      # from the monorepo root
cd packages/extract-pdf/site
bun run dev                      # the docs, /demo and its /api/*, on one port

The site imports extract-pdf, extract-webpage and extract-cite from their source through Vite aliases, so there is no package build first. Under bun run dev the API runs in Node, and its variables come from the shell, for example DOCLING_PROCESSOR_URL=http://127.0.0.1:7860 DOCLING_API_TOKEN=local-test-token bun run dev. bun run preview runs the built Worker in workerd.

Deploy your ownCopy Anchor Link

cd packages/extract-pdf/site
bun run build
bunx wrangler deploy

PDF parsing is CPU-bound. On the Workers Free plan (10 ms of CPU per request) most real PDFs fail with error 1102; on the Paid plan uncomment "limits": { "cpu_ms": 30000 } in wrangler.jsonc. See Running on Workers.

Admin panelCopy Anchor Link

/admin signs in with a password and sets the site's global keys (the OpenRouter key and default model for the Citation tab, and the Docling OCR URL and tokens) without a redeploy. A value saved there wins over the same name in the Worker's dashboard; clearing it falls back to the dashboard value.

bunx wrangler secret put ADMIN_PASSWORD   # until set, /admin is disabled

Keys are stored in the SETTINGS KV namespace (created on first deploy) and are shown masked, never sent back to the browser. Sessions last 8 hours in an HttpOnly, SameSite=Strict cookie signed with the password, so changing the password signs everyone out. Eight failed logins from one IP lock it out for 15 minutes. Use a long random password.