Workflows - anyformat

You compose the graph from a small set of node types, wire them with edges, and run any document through the whole thing as a single call. Each workflow targets one kind of document and one shape of output. Need a different output? Build a different workflow.

The five node types

A workflow’s graph is built from five node types. Each one does one thing.

Node	What it produces	Typical use
Parse	Structured markdown of the document (text + tables + figures, with bounding boxes and per-block confidence)	Always required; exactly one per workflow
Classify	A category verdict for the document	”Is this an invoice or a receipt?”
Splitter	Document segments (one logical document broken into multiple)	“This PDF contains four invoices stapled together”
Extract	Typed fields (data_type) pulled from upstream parsed content	The classic “give me invoice_number, total, date” use case
Validate	Pass/fail checks on extracted values	”Total must be ≥ 0 and within 1% of line-item sum”

Edges (source → target) wire nodes into a pipeline. Branching nodes (classify and splitter) need a branch on each outgoing edge to say which category or split rule the downstream node handles. The full topology rules — exactly one parse node, no cycles, who can connect to whom — live on the Create workflow endpoint reference.

Three common shapes

You’ll see these three over and over. They’re the same five node types arranged differently.

Parse-only

One parse node, no edges. Produces markdown — useful when you want anyformat’s parsed output to feed your own pipeline (RAG, custom LLM, search index).

[parse]

Linear: parse → extract

A parse node into an extract node. The default for “give me structured fields from this document.”

[parse] → [extract]

Branched: classify → extract-per-type

A classifier picks the document type, then routes to an extract node tailored to that type. Same idea works for splitters.

                       ┌──> [extract_invoice]
[parse] → [classify] ──┤
                       └──> [extract_receipt]

See the recipes for end-to-end examples of each shape.

How to think about workflows

Most users follow the same lifecycle:

Create

Define the graph — which nodes you need, how they connect

Refine

Run sample documents through it, inspect the output, tighten the schema or instructions

Publish

Mark it ready for production use

Run at scale

Apply it to many documents — manually, via API, or via cloud-storage integration

In the web platform you usually build workflows visually from the home screen. Programmatically you submit a JSON graph to POST /v2/workflows/.

What’s next?

Runs & results

What happens when you run a workflow, and how the output of each node type is returned

Field types

The data_type values an extract node’s schema can use

Build your first workflow

Linear parse → extract walkthrough — UI, curl, and Python

Create workflow (API)

The full graph schema, all five node types, and topology rules

Documentation Index

​The five node types

​Three common shapes

​Parse-only

​Linear: parse → extract

​Branched: classify → extract-per-type

​How to think about workflows

​What’s next?