News

Load documents with AI, driven entirely by metadata

Typing data from documents is one of the oldest chores in business software. The Suite now has a component for it. You describe it in metadata, like any other page, and users get a complete import screen with no custom code.

How it works

  • You define the documents you expect as rows of the page: an invoice, a packing list, a certificate. Each row says which AI instructions read it, which query shows what was recognised, and which procedure loads it.
  • The user selects the files and presses Analyze. Every row comes back with a short summary and its totals, and the full data is one click away.
  • The user reviews, removes a row or analyzes it again, then presses Load. Each row is written and the final procedure runs, all in the same database transaction (Oracle and PostgreSQL). If anything fails, everything rolls back.

Original or scanned, one file or many

The same row reads a text PDF exported from another system and a scan of a paper document. You can give each row its own file, or one cumulative file for several rows, for example a single scan holding the invoice, the packing list and the summary. Each row reads only its own part, and any row can still receive a file of its own.

Real life is not tidy

  • A document that did not arrive does not block the others: a row without a file, or a file that does not contain that document, is skipped and clearly marked.
  • A row can be duplicated when the same kind of document comes twice, such as two statements from the same supplier, so both are loaded in the same run.

Your data, your rules

The loading procedures are your own SQL, and the AI instructions are configured per project. The summary comes back in the language of the user. The AI never writes to your database: only the Load button does.

We first used it on textile supply-chain paperwork: invoices, packing lists, yield statements, certificates. It is generic, so it fits any process that starts from a document.