Extract a PDF Table to JSON for API or Application Use

A clean two-page report becomes one flat array

The most common happy path: a PDF with a clear header row. Note the keys come from the first line, every value is a string, and page 2's rows are appended to the same array — there is no nesting per page.

PDF (page 1, then page 2 with the same columns):
SKU        Name             Price   InStock
A-100      Widget           12.50   42
A-101      Gadget           9.99    0
--- page 2 ---
SKU        Name             Price   InStock
A-102      Sprocket         4.25    inf

Downloaded JSON (one flat array):
[
  { "SKU": "A-100", "Name": "Widget",   "Price": "12.50", "InStock": "42"  },
  { "SKU": "A-101", "Name": "Gadget",   "Price": "9.99",  "InStock": "0"   },
  { "SKU": "SKU",   "Name": "Name",     "Price": "Price", "InStock": "InStock" },  <- page-2 header row
  { "SKU": "A-102", "Name": "Sprocket", "Price": "4.25",  "InStock": "inf" }
]

Drop the repeated header rows before POSTing

Because every page's first line becomes data after the very first page, repeated print-headers leak in as bogus rows (see the third row above). Filter them where a key equals its own header value.

const rows = JSON.parse(json);
const clean = rows.filter(r => r.SKU !== "SKU");   // drop repeated headers

await fetch("https://api.example.com/products", {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify(clean),
});

Cast strings to a typed payload with Zod

The tool never guesses types. Put a schema at the boundary so a missing cell or a stray row fails validation instead of silently posting "Price": "Price".

import { z } from "zod";

const Product = z.object({
  SKU: z.string().regex(/^[A-Z]-\d+$/),
  Name: z.string().min(1),
  Price: z.coerce.number().positive(),
  InStock: z.coerce.number().int().nonnegative(),
});

const payload = z.array(Product).parse(clean);
// payload[0].Price is now a real number

Reshape keys to match your API contract

PDF headers rarely match your endpoint's field names. Map once, explicitly, so the rest of the pipeline is stable even if the PDF's wording changes.

const body = clean.map(r => ({
  sku:       r.SKU,
  title:     r.Name,
  priceCents: Math.round(Number(r.Price) * 100),
  stock:     Number(r.InStock),
}));
// body is the exact JSON your /products endpoint expects

Quick command-line check with jq

Before wiring anything, sanity-check the array shape and unique key set from the terminal.

# how many rows?
jq 'length' report.json

# what keys does the first object have?
jq '.[0] | keys' report.json

# drop repeated header rows and count again
jq '[.[] | select(.SKU != "SKU")] | length' report.json

The engine uses the first detected line of each page as the keys. If a report prints a title or date line above the actual column headers, that title line becomes your keys and the real header row becomes the first data object. Crop the title region out first, or post-process by promoting the second object's values to keys in your script.

Only the very first line of the document anchors the keys per page; on page 2+ the printed header line is emitted as a data row. This is expected given the page-by-page grouping. Filter rows where a key equals its own header text (r.SKU !== 'SKU') before sending to your API.

There's no text layer to read, so PDF.js returns no text items and the output array is empty. Run PDF OCR first to add a searchable text layer, then extract. A blank result almost always means the page is an image, not a bug.

The runner checks size and page count before extracting. On the free tier a file larger than 2 MB or a PDF with more than 50 pages is blocked with an upgrade prompt. Pro raises this to 50 MB / 500 pages. Split a large PDF with PDF Split or extract just the pages you need with PDF Extract Pages first.

There is no type inference — "12.50" stays a string, (123) stays "(123)". This is intentional so the tool never mangles an ID, a code, or a leading zero. Cast in your own code with Number(), a regex, or a schema's z.coerce.

Rows are defined by vertical position, so a long description that wraps to a second line is read as a separate row with mostly empty cells. Merge wrapped continuation rows in post-processing (a row whose key column is empty usually belongs to the row above).

If a later page has a different column order or count, its rows are still appended to the same flat array but keyed by that page's own first line. Mixed schemas in one array can break a strict validator. Split such documents and extract per section, or normalise keys after extraction.

Page counting tolerates light encryption, but the text-layer read can fail on a password-protected file. Remove the password with PDF Remove Password (you'll need the password) or PDF Unlock first, then extract.

The engine treats a whole page as one table region — it groups every line on the page. Two distinct tables on the same page merge into one block of rows with one set of keys (the first table's header). Separate the tables onto different pages, or split the array on a recognisable boundary row afterward.