What you can do today

  • Choose an already processed native or English OCR PDF and define typed fields with literal labels.
  • A registered five-step DAG runs reading, field extraction, evidence checks, owner review and JSON export.
  • Saved progress and a durable review wait let the worker resume after an interruption.
  • Runs pin source and extraction versions, show step status and retain review decisions. Changed or deleted sources prevent reuse of their outputs.
  • Review is required before export. Missing required fields and unavailable pages block approval; OCR warnings need explicit acknowledgment.

A reviewed field export that can resume

Select a processed invoice, define the invoice number and total, then start the workflow. Inspect the extracted values and their source pages, resolve blocking input problems in a new run, and approve a qualified result for JSON export.

Supported scope and limits

  • This is one registered document DAG. A general workflow builder, schedules, model steps and business-system writes are not available.
  • There are two active runs and 100 retained runs per workspace. The runtime is hosted on a single server; it is not a highly available service.
  • Literal-label extraction does not interpret arbitrary layouts. A repeatable script can consistently repeat a wrong rule, so source review still matters.
  • Document deletion revokes access to derived workflow outputs. Immutable audit and artifact records can remain stored; immediate physical erasure is not provided by this workflow.

Planned extensions

  • Extend registered operations to document matching, explicit unknown decision paths and reusable reviewed rules.
  • Add qualified model steps with bounded spending and evidence checks.
  • Add organization retention controls and coordinated recovery across source storage, workflow state and artifacts.
  • Require destination-specific duplicate prevention, delivery receipts and reconciliation before enabling external writes.

How to check the result

  • Interrupt a run before and after each durable step and verify the recorded result and recovery path.
  • Test duplicate events, expired approvals, changed source versions and unknown decision inputs.
  • Verify a retried destination operation creates one intended effect where its contract supports this; otherwise require reconciliation.
  • Measure complete accepted outcomes, exceptions, review time and total cost rather than completed steps alone.