Schema Builder

Tell Docxster exactly what to pull from a document, and how to trust it

Schema Builder is where you define what data Docxster should extract from a document type: the fields, the tables, and the rules each one has to follow. Draw a box on a sample document, tell it what that field is, and every document of that type gets extracted the same way from then on. If you've been hand-checking extracted data because there was no way to enforce what "correct" even looked like, this closes that gap.

How it works

A schema starts with fields, string, number, boolean, or date, and tables with their own columns underneath. You link each one to a document by drawing a box directly on a sample file. That works across multi-page documents too, and zoom and pan won't throw off the box coordinates once you've set them.

A schema can hold more than one sample instance, and each one is tracked separately for whether it's been tested. Run Extraction lets you check the real output against a real document before anything goes near production. Toggle a field or table off and it's skipped during extraction instead of deleted, and toggling one as required means a reviewer downstream can't submit with that value missing or blank.

Two controls sit directly on a field:

1. Use Constant Value – turn this on for a field and the model stops looking for it in the document. It just returns whatever fixed value you typed in, every time, which works well for something like a cost center code or a department tag that was never going to be in the document anyway.

2. Validation – attach a regex, either your own or a built-in preset (numbers only, all caps, a length range, an email format). If a reviewer enters something that doesn't match, they get a red border and an inline "doesn't match required format" message, with Approve blocked server-side until it's fixed. An empty value skips the regex check and falls back to whatever the Required toggle says instead.
   
   

If only a couple of fields on a document came back wrong, Selective Re-Run lets you rerun just those fields, or just that schema, instead of sending the whole document through extraction again. That saves time and the AI usage a full rerun would burn on fields that were already fine. Long documents get their own fix too: anything six pages or more used to be prone to timeouts and fields silently coming back as "No data found," and a chunking change now splits longer documents smarter at extraction time, without changing how shorter documents are processed.

Once you're working with more than a handful of document types, folders keep them organized. Create one, rename it, move a schema in or out, or delete it, and anything you haven't sorted sits in an "Uncategorized" bucket pinned at the top. Publishing a schema is what makes it available to flows, and every publish is versioned. The underlying active state for a field or table persists correctly through import and export too, so a disabled label doesn't quietly reactivate itself when a schema moves between environments.

Where this is useful

Freight and logistics operations

A BOL number, a PRO number, or a rate confirmation code has to follow an exact format, and a regex validation rule catches a malformed one before it moves downstream, instead of someone catching it after the shipment's already in motion. Mark fields like weight or ship date as required so a reviewer can't approve a load manifest that's missing one.

Customs brokers

An HS code or an entry number needs to follow an exact format, and regex validation catches a malformed one before it ever leaves review. A constant field can carry a fixed processing code that has nothing to do with what's actually printed on the commercial invoice or entry summary, so you're not hunting for a value that was never going to be there.

Lending and mortgage

A loan file has fields that can't go through blank: income figures, borrower names, dates. Mark those required on a loan application or income verification sheet, and a reviewer physically can't approve a file that's missing one, instead of catching it two steps later.

Manufacturing and procurement

Purchase orders and spec sheets often run long, and the chunking fix means a ten-page document doesn't come back with half its fields marked "No data found" the way it used to. Mark critical fields, like part numbers or quantities, as required so a bad extraction gets caught before it reaches a buyer.