AI Data Extractor: web page or PDF to JSON by your schema

Extract structured data from any web page or PDF with AI. You name the fields you want; you get one clean JSON object per URL. A field the page does not state comes back as null, never as a guess.

Paste public page or PDF links and a small schema such as {"price": "number", "in_stock": "boolean"}. The actor fetches each document once, has a language model read it, and returns exactly those fields. No browser to run, no selectors to maintain, no prompt to write: the schema is the prompt.

What you get, per URL

• data: one entry for every field in your schema, in your order: the value the page states, or null when it does not.
• fields_found and fields_requested, null_fields (which fields the page does not state) and verbatim_fields (which values were found word for word in the text).
• kind (html or pdf), page_title, url (what you supplied) and final_url (after redirects).
• truncated and pct_read: a long document is read up to about 20,000 tokens (roughly 70,000 characters); truncated: true and the share read tell you when only the first part was seen.
• source and fetched_at.

Fields and types

Type	Returns
string	text
number	a number, such as a price or a year
boolean	true or false
date	a date as YYYY-MM-DD
string[]	a list of text values, such as authors or tags

One schema per run, 1 to 20 fields, the same for every URL.

Use cases

• Pull the same facts (name, price, contact, date) from a list of product, company or news pages without writing selectors.
• Extract title, authors, abstract and year from a list of paper or report PDFs.
• Turn filings, spec sheets or invoices published as PDF into rows for a spreadsheet or a database.
• Give an AI agent a dependable way to read one page into typed fields.

Input

Field	Type	Default	Description
urls	array of strings	prefill: the Solana article on Wikipedia	Public http or https links to web pages or PDFs (up to 50).
schema	object	prefill: five fields for that article	Field names mapped to string, number, boolean, date or string[], 1 to 20 fields.

With no input at all, the actor runs that example page and schema, so the default run always returns data. If you give URLs without a schema, a schema without URLs, or an invalid schema, the run stops with a clear message before any request is made, and nothing is charged.

How to run it

In the Apify Console: open the Actor, paste your links into the Input tab, edit the schema and click Start. The dataset has an Overview table (one row per URL, with the extracted data) and exports as JSON, CSV or Excel.

Through the API: one request starts a run and returns the items (replace YOUR_API_TOKEN with your Apify API token):

curl -X POST "https://api.apify.com/v2/acts/ochinimus~ai-data-extractor/run-sync-get-dataset-items" \
  -H "Authorization: Bearer YOUR_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://bitcoin.org/bitcoin.pdf"], "schema": {"title": "string", "author": "string", "abstract": "string"}}'

The answer is a JSON array with one item per URL that returned at least one of your fields.

Sample output

A real result for the default input (the Solana article on Wikipedia), captured live on 15 Oct 2026:

{
  "url": "https://en.wikipedia.org/wiki/Solana_(blockchain_platform)",
  "final_url": "https://en.wikipedia.org/wiki/Solana_(blockchain_platform)",
  "kind": "html",
  "page_title": "Solana (blockchain platform)",
  "data": {
    "name": "Solana",
    "original_author": "Anatoly Yakovenko, Raj Gokal",
    "initial_release_year": 2020,
    "consensus_mechanism": "proof-of-stake",
    "native_token": "SOL"
  },
  "fields_requested": 5,
  "fields_found": 5,
  "null_fields": [],
  "verbatim_fields": [
    "name",
    "original_author",
    "consensus_mechanism",
    "native_token"
  ],
  "truncated": false,
  "pct_read": 100,
  "source": "AgentFeed, page or PDF fetched once from the URL you supplied",
  "fetched_at": "2026-10-15T09:58:50.000Z"
}

Pricing

Pay per event: $0.02 per page or PDF extracted (event extracted-page), plus Apify's tiny per-run start fee.

You are charged only for a URL that returns at least one of your fields. A link that cannot be fetched, a site whose robots.txt forbids it, a private or blocked address, a file that is not a web page or a PDF, a document over 10 MB, a page that states none of your fields, and any data-source error return nothing and cost nothing. A schema that is invalid stops the run before any request.

Data source and its limits

• Source: AgentFeed, our own data service. It fetches the one URL you supply, once, on your request, and has a language model (currently Gemini 3.1 Flash-Lite) read it against your schema. The document is not stored.
• One URL, one result. It reads the page you give it and does not crawl links or follow pagination.
• Plain fetch, no browser. Only http and https links on ports 80 and 443 are fetched, redirects are followed, as HTML or PDF. Content a site builds only with JavaScript after the page loads may be missing, and pages behind a login or a bot wall cannot be read.
• Up to 10 MB per document, and about the first 20,000 tokens are read. truncated and pct_read say how much.
• robots.txt is honoured. If the site forbids automated fetching, the link returns nothing.
• Private, local and reserved addresses are refused.
• A model reads the page, so check the values that matter. verbatim_fields lists the values found word for word in the text; a value you need exact should be one of them, or checked against the page.
• The content belongs to its publisher, and you are responsible for how you use it.

Source code

github.com/seekdaseek/agentfeed-apify/tree/main/ai-data-extractor (MIT).