Vault Vision
Local detection of tables, figures, charts and equations — and honest limits on it.
In short
Vault Vision finds tables, figures, charts and numbered equations in a paper and reads their contents. Detection happens on your computer, with no model call; reading the contents uses the engine you choose. It is imperfect and it is slow, and the app says so.
1Upload a paper with a text layer
Detection runs in the background with per-page progress. A PDF with no text layer is refused — there is no OCR.
2See the elements
Overlays are coloured by type. Click one to see its crop, caption and content, with badges for estimated, vision-read or corroborated.
3Use or correct them
Cite an element, send a table to the notebook as a DataFrame with provenance, or edit a cell and keep the edit history.
For technical users
Detection uses pypdfium2 (200 dpi render + word boxes) and a Table Transformer exported to ONNX on onnxruntime, plus caption heuristics; IoU > 0.5 merges boxes and crops are padded 2%. Measured: detection 12–54 s per paper on the held-out set; extraction adds one model call per element (about 92 s for a 2-page, 3-element paper; 278 s for three charts). Chart values are always estimated with an axis note. Table cells come from the text layer when one exists, otherwise from vision. The model ONNX file is downloaded on first use and is MIT-licensed.