Add high-fidelity document import to your form builder.

One JSON document per import: the form itself, a review note for every judgement call, a verification verdict, and what the source turned out to be. Here is a piece of one.

{
  "form": {
    "version": "1",
    "title": "Patient intake",
    "sections": [
      { "id": "s2", "heading": "Insurance", "description": null, "items": [
        { "id": "q7", "kind": "single_choice",
          "label": "Do you have coverage through an employer?",
          "description": null, "required": true, "show_when": null,
          "choices": ["Yes", "No"], "other": null },
        { "id": "q8", "kind": "short_answer", "label": "Policy number",
          "description": null, "required": null, "format": null,
          "show_when": { "question": "q7", "operator": "equals", "value": "Yes" } },
        { "id": "q9", "kind": "short_answer", "label": "Attach a photo of your card",
          "description": null, "required": null, "format": null, "show_when": null }
      ] }
    ]
  },
  "review_notes": [
    { "code": "kind_downgraded", "item_id": "q9",
      "message": "The original was a file upload. This editor does not support uploads, so it came through as a short answer." }
  ],
  "verification": { "status": "needs_review" },
  "source": { "format": "pdf", "page_count": 3, "char_count": 4812 }
}

Field names and structure are the same for every import, whatever the document was.

Converting a document means reading it, so the honest version of this page starts with where it goes. Everything below describes the API this page is calling right now.

Where your document goes

Structuring the form is done by a large language model. We use Google Gemini, OpenAI and Anthropic, tried in that order, falling back only when one is unreachable or returns something unusable. The request carries the text of your document, our instructions and the output schema — and nothing else. No account id, no key, no file name, no user data goes with it.

What we store

Not your document. The file, the text extracted from it and the JSON we return are held in memory for the length of the request; nothing is written to a database, to object storage or to disk, and the container runs with no volumes attached. What we do keep is one metadata record per import: counts of bytes, characters, pages and questions, the review-note codes without their text, which model ran, token counts, duration and cost.

No way to fetch it back

There is no endpoint that returns a past import. The result is sent once, in the response to your own request. This page holds it in the browser tab's memory and nowhere else — not local storage, not session storage — so closing the tab is the end of it.

Keys

An API key is 32 random characters from a cryptographic generator. We store only its SHA-256 hash, never the key, and show it once when it is created. The database is closed to browser access entirely; only the server can read it. Keys carry a quota and can be revoked.

Importing from a URL

A URL import is restricted to http and https on publicly routable addresses. Loopback, private ranges, link-local and cloud metadata addresses are refused, as are internal hostnames and URLs carrying credentials. Redirects are followed at most three times and every hop is checked again. The fetch sends no credentials of any kind, times out at 20 seconds and stops at 20 MB.

Limits and access

Browser access to the API is an explicit origin allowlist, with no wildcard. Requests are rate limited per key and per source address. Uploads stop at 25 MB. Errors come back as plain messages, never stack traces.

Reviewing us on behalf of a customer? Send the questionnaire to the address below and you will get a straight answer, including to anything this page does not cover.