Today we are introducing Page Extract API: describe the fields you want in a JSON Schema, and get them back from any PDF as JSON that fits it. Long filings, scans, and tables of thousands of rows included.
Page Extract is powered by PageIndex Agent. Where the agent answers a question, Page Extract fills a schema: it finds where each field lives in the document, reads those pages, and returns the values, in the shape you asked for.
Results
We ran Page Extract on ExtractBench: 370 documents, from one page to 192, each with a schema and the values a careful reader would extract. The measure is value F1 over every field, next to what each method costs a page:
| Method | Value F1 | Cost a page |
|---|---|---|
| Page Extract pro | 96.80 | $0.05 |
| Page Extract standard | 96.37 | $0.03 |
| LlamaExtract Agentic Plus | 96.17 | $0.075 |
| Reducto Deep Extract | 96.08 | $0.0635 |
| Pulse (Effort) | 96.07 | $0.105 |
| LlamaExtract Agentic | 95.65 | $0.0312 |
| Codex (GPT-6.1 Sol) | 95.31 | $0.0713 |
Page Extract pro is the most accurate method on ExtractBench. Page Extract standard still scores above every other method, at $0.03 a page: 60% less than LlamaExtract Agentic Plus, the most accurate of the others.
The other rows are from ExtractBench's leaderboard; ours are our runs of November 8, 2026, priced at what we charge. Every document's result and score, and a script to run it again, are on GitHub.
How to use it
Install the PageIndex SDK, describe what you want, and extract:
pip install -U pageindex
from pageindex import PageIndexClient
client = PageIndexClient(index="cloud") # reads PAGEINDEX_API_KEY
schema = {
"type": "object",
"properties": {
"company": {"type": "string", "description": "The company's name, as printed on the cover"},
"segments": {
"type": "array",
"description": "Each business segment in the segment results table",
"items": {
"type": "object",
"properties": {
"name": {"type": "string"},
"revenue_usd_millions": {"type": ["number", "null"], "description": "Revenue, in millions of dollars"},
},
},
},
},
}
job = client.extract("report.pdf", schema) # sends the file, waits for the result
print(job["extract_result"])
The result fits the schema. A value the document does not give comes back as null, and every row of a list comes back, in the order printed, however many there are. In Python the schema can also be a Pydantic model, with the result parsed back into it.
What it can do
- Your schema is the instruction. Each field's description tells Page Extract what to look for; see Writing a schema. Anything about the whole document, such as "amounts are in thousands", goes in the instructions.
- Long tables, in full. A 13F filing's holdings table comes back as all of its 3,000-odd rows, not the first page of them.
- The page each row came from. With
cite_sources, every list item carries the page it was read from, to check any value against its page. - Two tiers.
standard($0.03 a page) for most documents;pro($0.05 a page) for the hardest: values worked out across many pages, diagrams, dense charts and long drawings. - Some pages only. Extract from a range of pages, and pay only for those.
- In the background. Each extraction is a job: start it, come back for the result, or list, cancel and delete jobs. See Options and jobs.
- One account with Page Search. The same key, credit and documents work with Page Search API.
Get started
Create a key in the Developer Dashboard and follow the Quick start, or paste this into Claude Code, Cursor, or any coding agent that can read a URL:
Read https://pageindex.ai/docs/extract/quick-start.md and follow it to set up Page Extract API in this project.
Read more on the Page Extract API page.