A PDF lands in your flow, and getting anything usable out of it means three separate steps: OCR it, parse the result into something structured, then figure out where to put the images so they don't just vanish as base64 blobs.

The new Agentic Doc Extraction node does all three in one place. Give it a document, tell it what you want back, and you get clean Markdown, structured JSON, and every image, stored and ready to use, from a single step.

What you actually get back

• Markdown, page by page or joined into one document ready to drop into a RAG pipeline as-is.

• Structured JSON, shaped to a schema you write yourself. No prompt tuning; you describe what a field means once, and the model finds it on every document you run through.

• Every image in the document, automatically stored with a signed URL you can use right away and a durable key for later. No more inline base64 bloating your flow's output.

• Bounding boxes and per-image annotations, if you want the model to read a chart or figure the same way it reads a page.

How it works

Drop the node into a flow, point it at a document URL, and pick your OCR provider: Mistral's public API, or your own private Azure AI Foundry endpoint if that's where your data needs to stay.

Under the hood, each run gets its own isolated container. It splits the document into pages, OCRs them in parallel, and stores any images it finds.

The part that actually matters is the Output Schema. You're not writing a prompt and hoping you're describing each field in plain language, and the model reads the document specifically for that field. Change the description, change what comes back. No trial and error.

If you're coming from Doc Extractor

Doc Extractor is retired, but nothing you've already built breaks. If you're moving a flow over, three things behave differently and are worth knowing before you switch:

• Gemini isn't supported here. Only Mistral and Azure AI Foundry. If you were on Gemini, you'll need to pick a different provider.

• extractedText is now a list of pages, not one long string.

• structuredData is now a list too, one entry per page, instead of a single object.

The docs have a full field-by-field mapping if you want to check before you touch a production flow.

[Agentic Doc Extraction→]