How do I process documents whose layouts are all different?

It helps to name what you are actually dealing with, because the three cases have different economics.

Fixed forms are identical every time: a W-2, an ACORD certificate, a government application. The layout is defined by whoever publishes the form, and it changes only when they revise it. Semi-structured documents carry the same concepts in different arrangements: every invoice has a total, but no two vendors place it in the same spot. Free-form documents — contracts, letters, medical notes — have no reliable structure at all, only language.

Templates are defensible for fixed forms and a bad investment for the other two. If your corpus is mostly semi-structured, the number of templates you must maintain grows with your supplier list, and that curve never bends.

Count your distinct senders. Multiply by the time it takes to build and test one template. Then multiply the same number by an annual churn rate — vendors redesign their paperwork, acquire each other, and change billing systems — and you have the recurring cost of the template approach for your business. Compare that against writing one schema per document class and reviewing exceptions.

The trap is that templates feel accurate. On the documents they were built for, they are: a coordinate box either contains the number or it does not. The cost is invisible until the day a supplier moves their total two centimetres left and every downstream number silently changes.

With a field-first tool, your leverage is in the description text, because that is the instruction the reader follows. Descriptions written as labels are almost useless; descriptions written as disambiguation rules are what make a schema portable.

Compare "invoice number" with "the vendor’s own identifier for this bill, usually labelled Invoice No., Rechnung-Nr., or Document Number; never the purchase order or account number". The second one survives a German supplier, a document that prints both numbers side by side, and a template that calls it "Reference".

• Say what the value is, then say what it is not. Most extraction errors are confusions between two similar numbers on the same page.
• Name the synonyms you know, including other languages your suppliers use.
• State the format you want: date as YYYY-MM-DD, money as a number with no symbol, codes uppercased.
• For choice fields, use an enum with the exact allowed values instead of free text.
• Mention where the value usually is only as a hint, never as a rule — position is the thing you are trying to stop depending on.

When a document class genuinely splits — say your "invoice" schema now has to cover freight bills with per-container line items — the answer is a second schema, not a fork of the first. Keep each schema tied to a class of documents your workflow treats the same way.

Versioning matters more than teams expect. In Dokyumi, editing a schema’s fields bumps its version, and past extractions keep pointing at the version they ran under, so a change today does not silently reinterpret last quarter’s results. Whatever tool you use, make sure you can answer "which definition produced this row?" a year from now.