TL;DR
Agentic Document Extraction (ADE) is built from six composable operations. Learn these six words and you can describe almost any document workflow:
- Parse — make the document computable.
- Extract — request structured fields as output.
- Classify — determine the document type.
- Split — break large documents into components.
- Section — generate a hierarchical table of contents.
- Grounding — show exactly where every value came from.
Video Walkthrough
A short tour of the six concepts and how they fit together:
The Vocabulary: Six Words, One Foundation
ADE isn't a single monolithic call. It's a small set of primitives that compose. Here's what each one does and when to reach for it.
Parse — Make the document computable
Parse reads a raw file and turns it into something a program can work with. No layout pre-training or fine-tuning is required. It returns markdown and JSON outputs, and includes layout detection and visual grounding out of the box. Parse is the foundation — most other operations run on top of its output.
Extract — Request structured fields as output
Extract pulls specific fields from a document. Use it after Parse. It takes a schema (which supports infinitely long, nested definitions), returns clean JSON for each key–value pair, and includes traceability — a bounding-box reference for every extracted value.
Classify — Determine the document type
Classify assigns each page in a document to a type using its visual characteristics. It reads the raw file — no Parse needed — so use it before Parse for fast routing. Given a four-page file, Classify might label pages 1–2 as Invoice and pages 3–4 as Packing Slip, giving you a per-page class before you spend compute parsing.
Split — Break large documents into components
Split separates sub-documents out of Parse's output, and can separate multiple instances of the same type. Use it after Parse.
Split vs. Classify:
- Classify labels each page on the raw file, before Parse — fast routing.
- Split separates sub-documents from Parse's output — more precise.
- Rule of thumb: Classify to route in; Split to divide precisely.
Section — Generate a hierarchical table of contents
Section returns a reading-order list of sections and subsections, and lets you control how sections are grouped. Use it after Parse when you need document structure — a table of contents, chunk boundaries, or navigation.
Grounding — Show the location
Grounding tells you exactly where a value lives:
- Page number — which page the block sits on.
- Bounding box — its exact rectangle on that page.
- Line-level box — one box per rendered line.
So a value like $4,250.00 traces straight back to its spot on the page — enabling audits, human-in-the-loop review, and visual citations.
The Mental Model: How They Compose
The six words aren't a menu you pick one from — they stack:
- Classify to filter or route pages before parsing.
- Parse to read the document into blocks and grounding.
- Then Split, Section, or Extract on top.
- Grounding traces every result back to its source location.
Learn the vocabulary and the pipelines write themselves.
Getting Started
- Read the documentation — full reference for each ADE operation.
- Try the playground — run your own document and see Parse, Extract, and Grounding in action.