Pull the fieldsout of any document.
A vision model reads the document and returns the values you ask for — by specific fields, a template, or auto-detected from the document type. It reads checkboxes and tables, and abstains on anything it can’t see rather than guessing. Extraction, not a verdict: to check a document is genuine, use Verify.
What should it pull?
The document is classified, then the right fields for its type are pulled automatically. Unknown types fall back to every labelled field.
About document field extraction
Document field extraction pulls structured values out of any document — a payslip, tax invoice, bank statement, receipt, or an arbitrary form. You say what to pull four ways: name the fields yourself (a plain list, or each with a type — text, amount, date, or boolean — and a short description); pick a template (payslip, tax invoice, bank statement, or receipt); ask for nothing and let it auto-detect, where it classifies the document and then applies that type’s fields; or point it at an unrecognised type, where it works schema-free and returns every labelled field it can read.
A vision model reads up to a few pages and handles the awkward bits — checkboxes (returned as booleans), merged-cell tables, and nested tables. It is extraction, not verification: every value is simply what the document shows, never a judgement that the figure is correct or that the document is genuine. To check whether a document is authentic, use the document verifier instead.
- 01
Give it a document
Paste a public link to a PDF or image, or upload one. It reads up to the first few pages.
- 02
Say what to pull (or don’t)
Name specific fields, choose a template, or leave it on auto-detect. No schema is required — auto works on its own.
- 03
It reads the page
A vision model lifts the values, reading checkboxes as booleans and pulling figures out of merged-cell and nested tables.
- 04
Unrecognised types fall back
On a document type it doesn’t have a template for, it goes schema-free and returns every labelled field it can find.
- 05
You get the fields back
Each value arrives with a confidence and the page it came from; anything asked-for but absent is listed rather than guessed.
What you get back
fields
A map of field name to its value, the confidence in that read, and the page it was found on.
document_type
What the document was taken to be — either detected automatically or the template you chose.
mode
How the fields were decided: your named fields, a template, auto-detect, or schema-free.
not_found
Fields you asked for that aren’t clearly on the document — abstained on rather than invented.
Common uses
- Data entry from invoices and payslips into a spreadsheet or system
- Capturing the rows of a statement, schedule, or other tabular document
- Populating a record or form from a supplied document automatically
- Reading checkbox-driven forms where ticked boxes carry the meaning
Extraction reflects what the document states, not whether those values are correct or the document is genuine. Treat the fields as the document’s own claims — verify authenticity separately when it matters.
Questions
- Do I need to define a schema?
- No. Leave it on auto-detect and it classifies the document, then pulls the right fields for that type — or, on a type it doesn’t recognise, returns every labelled field it can read. Naming fields or picking a template is optional, for when you want exactly a known set.
- Does a value being extracted mean it’s true or genuine?
- No. Extraction reports what the document shows — nothing more. A figure that was wrong, altered, or on a forged document is extracted just the same. To judge whether a document is authentic, use the document verifier.
- Can it read checkboxes and tables?
- Yes. Checkboxes come back as boolean fields, and the vision model reads merged-cell and nested tables — not only flat key-value labels.
- What happens if a field I asked for isn’t there?
- It’s returned in not_found rather than guessed. The tool abstains on anything it can’t clearly see on the page instead of inventing a value.
- Do you store the documents I submit?
- No. Extraction is stateless — your document is read to pull the fields and is not retained.