Every user-facing change, newest first. Fixes are listed with the same prominence as features — including the ones we found ourselves.
Retention enforced end to end, clearer paywalls, security page
Original files are now deleted the moment a job finishes, with a scheduled sweep as backstop — the one-hour deletion promise is enforced by machinery, not policy.
The security page shipped: where documents go, how accounts are protected, and a plain list of what we have not built yet.
Free accounts see the extraction paywall before choosing a file, with a sample extraction to show what a result looks like — previously the file uploaded first and was refused after.
Hitting a plan limit is no longer labelled as an error. “Paid capability” and “Out of credits” replaced “Could not extract”.
Password rules are shown at signup and ticked off as they are met; the same policy is enforced server-side.
Legal pages updated for live billing and AI extraction: refund terms for credit packs, tax wording, governing law, and an honest description of what AI output is and is not.
AI extraction, credit packs, and billing — the paid product
Structured extraction shipped: invoices, receipts, purchase orders, delivery notes and goods received notes come back as fields and line items, one AI credit per document up to 25 pages.
Every extraction checks its own arithmetic — line items against totals, quantity times price, received against ordered — and says plainly when a document does not add up. The model also names any field it could not read confidently.
When a document fails its checks, it is re-read once on a stronger model at no extra credit.
Credit packs went on sale: one-time purchases for OCR pages and AI credits, never a recurring charge. A first pack of $15 or more unlocks API keys without a subscription.
Paid plans went live through PayPal with receipts, invoice history and downloadable PDF invoices.
API keys shipped for paid plans: shown once, stored only as a hash, revocable within a minute.
Fixed same day: the extraction type picker silently used the invoice schema regardless of choice — found while benchmarking, fixed, and verified.
Accounts, dashboards, and the developer surface
Accounts launched: 100 OCR pages a month free, bigger files than the anonymous tools, and a usage history.
Sign-in with Google or email; welcome and verification emails carry the VisionParse layout, not a provider default.
The API documentation shipped, with every sample captured from a live call rather than written from memory.
The benchmark page published measured accuracy per engine — including the failure cases and the numbers we are not yet willing to advertise.
Fixed: compressed BMP uploads returned no text at all.
The free tools
Seven free OCR tools launched — image, PDF, JPG, PNG, screenshot, searchable PDF and table to CSV. No account, no watermark, files deleted within the hour.
PDFs with a real text layer are read exactly rather than re-recognised, and multi-page structure is preserved.
Downloads survive old Windows editors: UTF-8 with BOM, CRLF line endings.