Document intake

Invoices and contracts arrive as PDFs and scans. Millstream fetches each document, has an agent extract its fields, checks them in plain code, and sends only the exceptions to an accounts team, with every correction recorded against the step that made the mistake.

The pipeline

flowchart LR
  W(["webhook: upload"]) --> F[fetch document] --> X["extract (agent)"] --> V[validate]
  V -->|no issues| S[/spot check/]
  V -->|issues| E[/exception review/]
  S & E --> P[post to ERP]

pipeline: invoice-intake
description: Extract, validate and post supplier invoices

vars: { minConfidence: 0.9, autoApproveMaxEur: 5000 }

triggers:
  upload: { webhook: { rate: 120/min } }

streams:
  docs.extracted: { retention: 400d }

steps:
  fetch:
    action: ./actions/fetch-document
    in: trigger.upload
    out: docs.fetched
  extract:
    action: lab/extract-invoice@1
    in: docs.fetched
    out: docs.extracted
    concurrency: 6
  validate:
    action: ./actions/validate-invoice
    in: docs.extracted
    out:
      - { when: "size(item.issues) == 0 && item.confidence >= vars.minConfidence", out: docs.clean }
      - { when: "true", out: docs.exceptions }
  spot-check:
    action: review
    in: docs.clean
    out: docs.approved
    config:
      inbox: accounts
      timeout: 72h
      onTimeout: { escalate: accounts-leads }
      autoApprove: "item.totalEur <= vars.autoApproveMaxEur"
      sampleRate: 0.05
  exception-review:
    action: review
    in: docs.exceptions
    out: docs.approved
    config:
      inbox: accounts
      timeout: 48h
      onTimeout: { escalate: accounts-leads }
      editable: [/supplier, /poNumber, /totalEur, /lines]
      reasons: [duplicate, wrongSupplier, missingPo, unreadable]
  post:
    action: lab/erp-post@1
    in: docs.approved

budget: { dayUsd: 40 }

What Millstream does for it

Documents as artifacts

• fetch downloads the file and attaches it to its output as an artifact, stored by content hash in a local directory or S3, never inlined in the item.
• The item view previews the PDF beside the extracted fields, the agent's transcript and its cost.

Extraction and validation

• extract answers the action's output schema: supplier, PO number, totals, lines and a confidence.
• validate is ordinary code: totals add up, the tax matches, the PO exists in the reference data shipped in the package's assets/.
• Each output goes to the first route whose condition holds, so the rule for "clean" lives in the pipeline file.

People on the exceptions

• Exceptions wait in the accounts inbox; the reviewer corrects the listed fields and approves, or rejects with a reason code.
• Clean invoices under 5,000 EUR are approved by rule, and 5% of them still reach a person as review samples, so the rule is measured.
• millstream check warns that agent output reaches post, an external effect, through a review that auto-approves. The warning is advisory: here the validation code and the samples are the chosen control, and the file says so in plain view.

Failures

• An unreadable scan that fails every retry becomes a dead letter, listed by error in the UI, and returns to its run with millstream replay --from dlq --run <id> once the action is fixed.

How it learns

A correction to /totalEur is attributed by walking the item's lineage to the nearest step whose output changed that field. That step is extract; validate passed the value through unchanged, so it is not blamed. Three or more such edits on one field become a lessons proposal for extract, with a rule and the corrected examples. When sampled rule approvals start disagreeing with people, the reviewPattern insight raises a warning below 90% agreement.

Figures to watch

Next

Human review · Observability · Security