Extract a balance sheet PDF to JSON
Everything below is the response the live API returned for a balance sheet, generated when this page was built. Not a description of what it would return — the output itself, including what it scored low on.
Who this is for
Credit analysis, underwriting and financial-data providers reading statutory accounts that only exist as PDFs.
The job, in the words people search for: “get balance sheet figures out of a filed PDF”.
What makes this shape awkward
Two comparative year columns, subtotal rows interleaved with detail rows, and indentation carrying the hierarchy. The indentation is whitespace, not structure — so "Total current assets" is a row like any other, and nothing in the file says it is a sum of the rows above it.
The measured result
| Pages read | 1 |
| Time | 6 ms (6 ms/page) |
| Labelled fields found unprompted | 4 |
| Tables | 1 (11 rows) |
| Input | balance-sheet.pdf (28 KB — synthetic, generated from a spec in the repo; real documents of this type cannot be published) |
Fields, with a schema
Ask for the fields you want by the label printed on the page. Each answer carries the text exactly as printed (raw), the value coerced to the type you asked for, a confidence, and a bounding box you can draw on the page to check it.
curl -X POST https://api.pdfcraft.dev/v1/extract \
-H "Authorization: Bearer $PDFCRAFT_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"file": "<base64 of your balance sheet>",
"schema": {
"entity": "string",
"registered_number": "string",
"period_end": "date"
}
}'What came back
| Field | raw | value | confidence |
|---|---|---|---|
entity | Halvorsen Industrial Ltd | "Halvorsen Industrial Ltd" | 60% |
registered_number | 08817422 | "08817422" | 60% |
period_end | 2026-03-31 | "2026-03-31" | 70% |
Every label it found without being asked
Send no schema and you get all of these, keyed by the label as printed. Useful for discovering what a new supplier’s layout actually contains before you write a schema against it.
Entity, Registered Number, Period End, Currency
Tables
Table 1 · 11 rows × 3 columns · page 1
Table confidence 75%. No header was promoted (confidence 10%). Nothing in the geometry distinguishes row 0 from the rows under it, and inventing a header would silently eat a row of your data.
| Intangible assets | 1,204 | 1,388 |
| Property, plant and equipment | 8,441 | 7,902 |
| Investments | 620 | 620 |
First 3 of 11 rows.
What this does not do
- No OCR. This reads the PDF’s text layer. A scanned or photographed balance sheet has no text layer and returns
422 extraction_failedwith the page numbers that were blank — deliberately, and quickly, so you can route it somewhere that does OCR rather than waiting on a guess. - No model. Run the same file twice and you get the same bytes back. That is the trade: it cannot infer a field that is not printed, and it cannot hallucinate one either.
- Ambiguity is reported, not resolved. A date like 03/04/2026 stays a string, and a bare
$returns a null currency, because guessing wrong on either is not something you can recover from downstream.
Try it on your own balance sheet
The playground takes a file and shows the same JSON, with every bounding box drawn over the page. No key and no signup for the first few; a free key gives you 100 pages a month.