Make any PDF accessible.

Tag untagged PDFs to the PDF/UA-1 standard — headings, paragraphs, lists, tables, figures and artifacts — with heuristics plus an optional vision model. Review each file, or tag a whole batch in one pass.

Documents Choose PDFs or drop them here Select one or more files — Ctrl/Cmd-click or Shift-click to multi-select

Form fields — optional, creates real fillable fields on flattened forms Create form fields — when a page is a flattened (non-fillable) form, a vision model locates its inputs and real, fillable form fields (text boxes, checkboxes, radio buttons, signature areas, …) are created at those positions and tagged as Form elements (works with any tagging engine below)

Tagging engine Tag documents — analyse the structure and write PDF/UA-1 tags (untick to only create form fields: the file is saved fillable but untagged)

Re-tag documents that already have tags (existing tags are stripped first) Tag immediately without reviewing (recommended for batches) Heuristics — built-in detectors, optionally refined by a vision model (below) OpenDataLoader — PDF structure parser: auto-tagging & reading order (needs Java 11+) YOLO — trained layout detector: element boxes & reading order from a vision model (needs PyTorch + Ultralytics) Docling — Docling-project layout model (heron-101): 17 element classes incl. page headers/footers & code (needs PyTorch + Transformers)

Vision model — optional, any OpenAI-compatible endpoint

Base URL Model

API key

Use a vision model to refine roles, reading order and tables

Use LLM only — let the model build the entire tag tree (structure, reading order, element boxes) from each page image, instead of refining heuristics Structure depth

Use a vision model to fix the reading order only (its own prompt — element roles, tables and everything else stay exactly as the engine detected them; mutually exclusive with the full refine above) Use a vision model to fix each table's cell grid (every detected table is sent to the model, which places its cells in the table's real rows and columns — this is what corrects the reading order INSIDE a table, which the option above cannot reach, since it sees a table as one region. Runs on its own or alongside it; a big table goes in bands. Nothing but the cell grid changes.) Generate short alt text for figures (its own prompt and pass — runs on its own or alongside either option above; needs only the endpoint)

The “Form fields” section above can share this endpoint — its detection runs independently of the refine option.

Document options

Language Title

LLM prompts — view & edit what each vision call is told; applies to the next job (GUI, API and CLI alike)

Each prompt below is the exact instruction text for one kind of vision call. Edits are saved to the server's prompts/ folder; the first save backs the shipped prompt up so “Reset to shipped” can always restore it.

Loading…