How Iris works

Follow one document through Iris, step by step, with every API call.

This is the technical view. To just convert a document, use the demo.

Follow a document

Iris turns a PDF into accessible HTML. Each step is marked by who does it. AI: models read and write. Code: prepares their input, checks their output and throws away replies that lose content. If asked, iris-pdf then tags the original PDF, using code only. Each step below says what it does, how this deployment sets it up, and, once you run a document, what happened and which API calls it made.

How a document moves through Iris and iris-pdfThree lanes of steps. The numbered steps below say the same in words.1 Iris, one page at a time2 Iris, the whole document3 iris-pdf, only if a tagged PDF is asked forYourdocumentCodeTurn each pageinto an image150 DPICodeRead the form'sfields (iris-pdf)if it is a formAIPage agent writesthe page's HTMLone call per pageAICheck HTML againstthe image (+ Code rules)a second modelpage imagefield namespage HTMLproblems: fix onceevery page's HTMLCodeJoin pages, fix ids,axe-core checkaccessibility checkAIReader listsproblemsreads text onlyAICopy editor fixes(Code refuses losses)a few rounds at mostCodeReadable names,unsolved list addedwhen you download itdocumentproblemsfixesdocumentAccessibleHTMLHTML + original PDFCodeRead the page'swords + positionsOCR for a scanCodeMatch Iris's wordsto the page'swords, not coordinatesCodeTag text in place,add form fieldsstructure treeCodeCheck pixels andwords, then saverefuses if changedwordsmatchestagsTaggedPDFAIOptional review:does it say what it shows?

A PDF, or page images in order. A real conversion: a few minutes.

Its session id, starting ses_. Nothing is converted again.

1. Upload

Code: Each page becomes an image, and a form's fields are read. The document joins a queue and starts when a slot is free.

2. Read and check each page

AI: The page agent writes each page's HTML from its image. AI: A second model checks it against the image, and Code: rules check links, image descriptions and ids. A page that fails is fixed once.

3. Join and check

Code: The pages are joined into one document, its ids are made unique, and axe-core checks it.

4. Review

AI: A reader reads the whole document as text, as a screen reader gets it, and lists problems. AI: A copy editor fixes them, looking at the page images. Code: A fix that loses content is thrown away. This repeats for a few rounds at most.

5. Result

Code: Readable names are added, and any problem left open is listed in the HTML. Feedback starts a new run of the same document.

6. Tagged PDF (optional)

Code: Only when asked: iris-pdf matches the HTML's words to the PDF's, tags the text in place, adds form fields, and checks every page before saving.

This deployment

From GET /v1/settings, /v1/limits and /v1/stats. Secrets are never shown. Each step above lists the settings that shape it.

Loading…

The API

Every endpoint, from GET /v1/openapi.json, the exact contract. A call this page made is shown in its step; try the others here.

Loading…

Run log

One line per thing Iris did. Each name links to its definition.

Words the run log uses
Page agent
The model call that writes one page's HTML from its image.
Verdict
The checking model's answer about one page: is it faithful to the image, and is it accessible.
Fragment
One page's HTML, before the pages are joined.
Block
One top-level part of the joined document. The copy editor changes the document block by block.
Round
One pass of the review: the reader lists problems, then the copy editor fixes them.
Declaration
The page agent saying a page is blank. Iris can refuse it, for example for a page that is only too dark to read.
Run
One conversion of a document. Feedback starts a new run of the same document.

No document yet.