PDF agents that can cite the page

Give AI agents structured PDF context instead of a raw file upload: page text, tables, bounding boxes, confidence, and links back to the exact source region.

Connect an agent

A PDF agent needs tools, not just files

Most agents treat PDFs as opaque uploads. They read too much, miss tables, and answer without proof. okraPDF exposes PDFs as structured, queryable sources so an agent can resolve a document, inspect its layout, ask for the relevant section, and cite a page region instead of guessing from a blob.

What the agent gets

Structured blocks for headings, paragraphs, tables, figures, and page regions.
Bounding boxes for cited values, so answers point back to the exact place in the PDF.
Confidence scores for routing uncertain extraction results to review.
Hosted document views and deterministic source URLs that an agent can hand to a human.
MCP, CLI, SDK, and API paths over the same document substrate.

Agent workflow

Resolve a PDF URL or uploaded document into an okraPDF source.
Read structure before parsing the whole document.
Ask for the page, section, or table needed for the task.
Return JSON plus cited page evidence.
Open a web review surface when a human decision is required.

Example agent pattern

okra context resolve https://example.com/report.pdf --json
okra context get src_123 --section tables --json
okra context ask src_123 "Which table contains Q3 gross margin?" --citations

Frequently asked questions

What is a PDF agent?

A PDF agent is an AI workflow that can inspect, query, transform, or route a PDF with tools. okraPDF gives that agent structured document context and cited evidence instead of making it reason over an opaque upload.

How does okraPDF keep agent answers grounded?

Every parsed value can carry page and bounding-box evidence. The agent can return a claim with a source link and region, so a person can verify it against the original PDF.

Can a human review the agent result?

Yes. okraPDF workflows include review surfaces for low-confidence extraction, accessibility attestation, and publish permissions, so agents can hand decisions to a human instead of silently guessing.