Document Processing Pipelines

Every organization has a pile: vendor invoices in a shared inbox, signed contracts in a drive, intake forms scanned at the front desk, receipts photographed on phones. Somebody types the important parts into a system, one document at a time. It's slow, it's tedious, and it's exactly the work a vision model can now do in seconds per document.

The hard part isn't getting a model to read a document. As you saw in Vision Agents, that's one API call. The hard part is doing it a thousand times without quietly corrupting your data, because a misread total looks just as confident as a correct one. In this lesson you'll build a pipeline that extracts invoices into a strict schema, checks every result in code, sends anything doubtful to a human, and measures its own accuracy field by field.

The shape of a document pipeline#

A reliable pipeline has five stages, and only one of them is the model:

inbox ──► 1. INTAKE ──► 2. EXTRACT ──► 3. VALIDATE ──► 4. ROUTE ──┬──► accepted ──► 5. EXPORT
           file type,     model +        code checks     clean?    │
           dedupe,        strict schema  arithmetic,               └──► review queue ──► human
           classify                      formats, rules                  corrections ──► eval set

The guiding principle: the model extracts, code validates, and humans decide the hard cases. The model is excellent at reading messy layouts and finding the right number on the page. Code is excellent at checking that numbers add up and dates make sense. Humans are needed only where those two disagree or can't tell. Notice this is a workflow, not an agent: the steps are known in advance, so your code sequences them and the model fills in one well-defined step.