Document Reader

What changed in Document Reader recently

1. New · Release 1.44 · 29 October 2026

   Read with, in the Document Reader: Automatic reads with the model chosen for reading documents, unclear handwriting re-read on the strongest — or choose any model your live providers offer that reads pictures: every Claude, ChatGPT and Gemini model that Admin → AI → Pull all models lists, new releases included, each with what it is good at (the hardest handwriting, speed, cost).

2. New · Release 1.43 · 29 October 2026

   What is on the page, in the Document Reader: a photo or PDF is read as a page — each heading, field (its printed label kept apart from the value), table, picture or drawing, stamp, signature, tick box and handwritten note, in reading order, with what it means and what it is for, where it sits, whether it is handwritten and how sure the reader was. The photo shows every part outlined where it sits, and the text is the page read in order. Ask’s data input reads photos the same way.

3. New · Release 1.42 · 29 October 2026

   The best model for each kind of work, chosen by the platform from every model your live providers offer, by quality, understanding, speed and price: answers (Ask, chats, WhatsApp and e-mail replies), reading documents (the Document Reader — it must read images and PDFs), hard calls (deeper analyses, settling answers that disagree), quick reads (sorting a floor update, placing data input) and speech. Each choice says why and names the runner-up; previews, experimental models and moving aliases are never chosen on their own. Choose any model by hand and it is kept while the provider still offers it; one button hands it back to the automatic choice.

4. New · Release 1.41 · 29 October 2026

   Answer together (Admin → AI, off until the owner switches it on): Claude, ChatGPT and Gemini answer the same question at once, with the same records. When they agree, the answer comes straight back; when they differ, one of them reconciles the answers against the records and lists where they differed. Ask, data input and the Document Reader say who answered, any model that timed out or failed, and whether they agreed or were reconciled — never “agreed” unless the platform found the answers the same. It costs one call per live model, plus one when they differ.

5. New · Release 1.40 · 29 October 2026

   Paperless copy in the Document Reader: as soon as a photo, scan, PDF or file is read, keep it as Excel (the details and fields, any table found in it on its own sheet, the full text line by line), Word (the same, with the photo of the paper) or PDF (a print-ready A4 page your browser saves as PDF, with searchable text). Each file uses the text and fields as they are on screen — your corrections included — says when it was read and by what, gives the AI model’s certainty only where a model read the fields, and is marked as an extracted copy not yet checked against the order or filed. Arabic documents come out right to left, with Arabic shaped correctly.

6. Improved · Release 1.38 · 27 October 2026

   The Document Reader reads and places a document in one step and shows the result as a review: what the document is (its type is now read from it — choose one only if it gets it wrong) and what it is for, which order or plant it belongs to and the references that put it there, what it says, every field it will fill with its value now and after, the screens it updates (Procurement, Suppliers, Inventory, Decision Center…) and the waiting decisions it answers. When it is right, “Confirm and file” is the only step; when it is not, correct the text or a field and check again, or place it on another project.