Skip to content

Examples gallery

Complete examples you can copy: each one has the workflow graph, the code to create and run it in Python, TypeScript and curl, and the real response of the server, produced by a script of the repository with the fake model and synthetic documents. make docs runs every example again and fails if an answer changes.

Example What it shows
Invoice with line items A parse → extract → validate graph.
Bank statement Every movement of a statement as a typed table (completeness: all_rows makes the extraction check it did not skip rows), and a rule that the opening balance plus the movements gives the closing balance.
Contract with rules Key terms of a services agreement and two business rules.
Receipt The smallest useful graph, parse → extract, on a till receipt.
CV Nested data from a CV: a list of small objects (skills) and a list of objects (experience).
Split a mixed scan One PDF holds a two-page invoice, a receipt and a one-page invoice, scanned back to back.
Classify and route classify reads the document and sends it down one edge per category; only the extraction on that edge runs, with the schema that fits.
Supplier lookup in a reference table The invoice says who the supplier is, but your ERP wants the supplier code.
Questions with citations over a corpus A knowledge node adds every document the workflow processes to a searchable corpus.
Fill in a form The fill_form node (called Edit in the Studio) finds the blanks of a PDF form, matches them with your instructions and writes the values in, returning the filled PDF as a run artifact.

Every example needs a running server: set DX_BASE_URL and DX_API_KEY (see the quickstart). Agents can do the same through MCP: connect your agent.

The answers are the fake model's

The documents are synthetic and the model is the repository's scripted fake, so the values are exactly what the example text says. With a real model on your own documents, expect the same shape (typed data, citations, checks), not the same values.