How it works

Set up a document type once by pointing at the values. After that, any file of that type becomes JSON automatically.

You do this once per document type. It takes minutes, and you only ever point at things. There is no expression language to learn: the one step that involves arithmetic is built from dropdowns, and the tool proposes it for you.

1. Open a real example

Drag in a file your system has produced. It stays in your browser. Nothing is uploaded while you are setting things up.

No file to hand? Press Load demo instead. It opens a sample statement with the finished template already applied, so you can see all five steps at once and press Extract straight away.

Because these files were laid out for a printer, every value sits in the same place on every copy. The account code is always on the same line, starting at the same column. That reliability is what makes this approach work.

A customer statement displayed as a grid of characters with a numbered ruler. — A statement, exactly as it would have printed.

2. Point at the individual values

Choose Field and drag across an account number, a date, a total. Each one becomes a named value in the result, and you name it to suit whatever will receive it.

Two values highlighted with labels, and a list showing what each captured. — The panel shows what you actually picked up, so you can check it before going further. If a selection is a character out, **adjust** nudges it.

3. Point at the rows that repeat

Transaction lines are a Table. Drag a box around the rows, then drag across each column in turn. It stays ready for the next column, so you can do them all in one go.

A dashed box around transaction rows with three columns highlighted inside it. — The dashed box is the area that repeats; the solid blocks are its columns. A separate rule decides where one row ends and the next begins, which matters when a long description spills onto a second line.

4. Say how to recognise this document

Choose Detect and drag over whatever marks the document type: a code in the corner, a report title, a heading. This is what lets you send files without having to say what they are.

The letters sta outlined on the first line, with a matching rule and a green label reading matches this sample. — Statements from this system start with `sta`; invoices start with `inv`. The green label confirms it as you go. Open a different document type to check the rule is not too broad. A rule that claims everything is the mistake you cannot otherwise spot.

5. Say what must add up

This is the step that separates a parser from a parser you can trust. A check states something that should always be true of this document, and the usual one is that the transaction lines add up to the printed total.

Press Suggest from this file and the tool looks for arithmetic that already holds across every document you have loaded, then offers it. It proposes only relationships that are exactly true, with no rounding allowance, because one that merely nearly holds is a coincidence and would fail on the first real run.

The rules panel showing a check reading sum of transactions.value equals totalDue, labelled 2 pass. — Built from three dropdowns rather than typed: **sum of transactions.value**, **equals**, **totalDue**. The expression underneath is what actually runs, so it stays inspectable. **2 pass** means both statements in this file agree with their own printed totals.

Figures printed rounded to the penny drift by up to half a penny per row, so on a long report set allow a difference of rather than picking a tolerance that is too tight on big documents and too loose on small ones.

You can also tick fields under Must not be empty. What that catches is a layout change that silently empties one, which otherwise arrives as a blank rather than as an error.

Why this matters more than it sounds. A parser that stops is cheap. A parser that quietly returns four transactions where the document printed five is expensive, and nobody reads sixty-three statements to notice. The document already carries its own proof, because the lines add up to the total. Declare that once and the good statements pass silently while the one bad one raises its hand, with the difference named.

6. Check the result, then publish

Press Extract to see the JSON that comes out, exactly as the API will return it. When you are happy, press Publish. From that moment the document type can be processed automatically.

Results for two documents, a selector reading all 2 as an array, and a summary reading 2 ok. — One file in, one JSON object per customer out: here two, and on a full run sixty-three. The summary says how many came through cleanly, so a problem shows up immediately rather than three steps downstream.

One file, many documents

A statement run arrives as a single file containing every customer. Two settings turn it into one result per customer, and it is worth knowing which does what.

Setting	What it decides
Pagination	Where each printed page starts.
Split	Which pages belong to the same customer. Usually on key change, so a new document begins when the account code changes.

Splitting this way handles long statements without being told to: one that runs onto a second page still carries the same account code, so both pages stay together as a single result.

Worth knowing

Send files exactly as your system produced them. Opening a report in a text editor and saving it again can shift every column, and the result is wrong figures rather than an obvious error. Pass the original through untouched.

Changes take effect when you publish. Edit and re-test as much as you like; nothing running in the background changes until you press Publish. Earlier versions are kept, so a result from last month can always be reproduced.

Keep recognition rules distinct. If two document types could both claim the same file, we tell you rather than guessing. Making one rule more specific resolves it.