Documents and scope

• From the composer's + menu, under Attach: Images, Files, or a whole Folder.
• From the documents pane's ADD menu: Files, or From URL to fetch a web page.
• A long paste into the composer becomes a .txt document of its own.

Accepted: PDF, Word (.docx), Excel (.xlsx), PowerPoint (.pptx); their OpenDocument twins (.odt, .ods, .odp); RTF, CSV, TSV, JSON, XML, text and Markdown; web pages (.html), email (.eml) and subtitles (.vtt); and pictures (PNG, JPEG, WebP, AVIF, TIFF, BMP, GIF). A file can be up to 64 MB unless your workspace sets another limit.

A file the corpus cannot read is refused before it uploads, with what to do instead: a .heic photo can be exported as JPEG, and an old .doc, .xls or .ppt saved in the newer format. A dropped folder's system files, such as .DS_Store, are skipped.

Each document shows one bar from 0 to 100, and every number on it is measured. Uploading fills the first part as the file's bytes are sent. Then it waits for a reader, with its place in line; the reader opens it and reads it page by page, slide by slide or sheet by sheet; splits it into passages and saves them; indexes the passages for search; and the bar reaches 100 when it is ready.

• Under the bar, five steps: Upload, Read, Split, Index and Ready. The Read step carries a scan mark once OCR has read a page.
• Beside them, what is being counted: “Page 4 of 12 · 2 by OCR · about 20 s left”, or “192 of 640 passages”. A time left appears once there is a measured rate to work it out from.
• The upload summary at the top of the pane reads “2 of 6 ready” with the average of the bars, and keeps a file that was refused, with the reason, until you dismiss it.
• To stop a document while it is being read, delete it: the reader stops at its next page.

A document that failed keeps its bar where it stopped, and says why, what that means, and what to do; when trying again can help it offers REPROCESS, which starts a new attempt from the beginning.

Email addresses, phone numbers, card numbers, national identifiers and people's names are removed before any passage reaches a model.

A scanned PDF page with no text layer, a Word file that is only pictures of pages, and a picture (a photo of a receipt, a whiteboard, a screenshot) are read with OCR. Printed text reads best; handwriting is not promised.

• While it reads, the row counts the pages OCR has read: “Page 6 of 12 · 6 by OCR”.
• Once ready, the row's details say how many pages OCR read, such as OCR 12p, and its answers cite those pages like any other.
• A picture with no words in it is kept and shown, but answers do not draw on it, and the row says so.

The pane's header counts your documents and how many are in scope. With nothing ticked, a question reads every document. Tick some and only those are read.

• SELECT has ALL, NONE and INVERT.
• With documents ticked, the ⋯ menu offers Tag…, Save as selection…, Move to folder… and Delete.
• Type @ in the composer to point a question at one document, note or sheet.

• One document: DELETE, then confirm. A document an answer cites is kept, and CITED BY shows which.
• Several, or Clear all…: they go at once with an UNDO for ten seconds. Answers that cited them keep their citations.