A self-serve API for developers to parse, extract, evaluate, correct, and deploy document AI applications.
Split functionality finds where one document ends and the next begins, then creates a separate child document for each one, with its own page range and PDF. Every cut comes with a confidence score, so you know which ones to trust.
Real intake is rarely one file, one document. A mortgage package, a claims bundle, or a scanned mailroom batch holds many documents in a single file, and anything that assumes one document per file breaks on day one. Split looks at the page images, finds where each document starts and ends, and cuts the file into separate child documents. Each child has its own page range, its own PDF and a confidence score for the cut.
A loan package arrives as one PDF holding the application, pay stubs, bank statements, W-2s, tax returns etc. Split cuts it into individual documents, and each one goes to the right schema.
A claims bundle puts forms, reports, invoices and correspondence in one file. Split separates them so each document can be classified and extracted on its own.
A bill of lading, commercial invoice, packing list and customs forms arrive stapled together. Split separates them so each is classified and extracted against its own schema.
Classifying a whole package gives a meaningless answer. Split first, then Classify labels each real document.
Classify and Extract run once per child document, and parent to child lineage means every extracted value traces back to its page range in the original upload.
Only the cuts the model was unsure about go to a reviewer, who drags pages between document lanes on a visual canvas. The rest flow straight through.
Split runs on a vision model that reads page images. Pick any vision model in the builder, set the reasoning effort for reasoning models, and change it later without writing code.
Scans and photos, yes. OCR runs on page images. Handwriting quality depends on the engine you choose (Azure Document Intelligence or a vision LLM). There is no separate handwriting mode.
Both. You get JSON with whole-document Markdown, per-page Markdown and typed blocks. Post Processing can also render CSV.
Yes. Tables are detected as table regions and output as Markdown tables.
Yes. Every block has a bounding box, expressed as percentages of the page.
Yes. Choose semantic, page or fixed-token chunking, with configurable size and overlap.
Azure Document Intelligence by default. You can select a vision LLM per pipeline.
There is no hard page limit. The file size limit is 100 MB.
Yes, for the LLM-based sub-stages. Azure Document Intelligence usage is billed in credits.
Jobs, results and documents are kept for 24 hours after submission, then deleted.
