Pipeline Factory
Versioned extraction operations
Limited access
Teach one known format—one page or a multi-page package. Let selected language models propose its page templates and fields, refine and test the Pipeline, then publish a version your systems run through the Pipeline endpoint.
- Teach
Create one page template per sample page
- Prove
Refine model proposals into one Pipeline field contract
- Test
Challenge page matching with fresh files
- Publish
Freeze an immutable Pipeline version
- Integrate
Call the endpoint and receive the defined fields
- One page template per sample page
- One combined Pipeline field contract
- A versioned integration endpoint
Pipeline Factory is available through limited access while authoring, disclosure, and product hardening continue.
Teach the Pipeline page by page
One format. One page or a package.
Upload the format as it really arrives. Full Document Extraction reads every sample page once, then selected language models propose a page template with anchors, fields, types, and source regions for each page.
- Use one page, a multi-page document, or a supported mixed sample set
- Create one page template per sample page
- Combine the enabled fields from all pages into one flat Pipeline field contract
Refine the Pipeline until the contract is right
Models propose. Your team decides.
Compare model proposals, adopt the strongest starting point, edit it directly, or send feedback so the same model reworks the same sample into a new proposal without erasing the earlier one.
- Review anchors, field names, types, output keys, and source regions
- Refine a proposal with feedback while preserving comparison and rollback
- Keep final approval with the reviewer—not an automatic model score
Test routing on fresh documents
Page order is not identity
Challenge the exact draft with a new document. Send the complete package or a subset of pages in a different order and inspect which pages the Pipeline matches and how it routes each match before publication.
- Match pages by anchors and relationships rather than page number
- See passed, partial, and mismatch outcomes
- Identify missing templates and unmatched pages, plus states and source evidence for fields on matched pages
Publish the Pipeline
The production boundary
Compile the reviewed page templates and field contract into an immutable version. Production keeps using that published version while new edits remain safely in draft.
- Activate the version your team approved
- Keep draft edits out of production traffic
- Pin a specific Pipeline version for controlled integrations
Call the Pipeline endpoint
Your integration starts here
Your system submits one document to the Pipeline endpoint. The published version routes every matched page to its template and returns one flat field response shaped by the reviewed contract.
- Send one document containing the complete page set or only the pages available, in any order
- Receive named, typed fields from every matched page
- Inspect missing or unmatched pages, plus explicit states and source-linked evidence for fields on matched pages while the result is retained