Transcribing a document

The Transcription Studio is the focused mode for writing and correcting a document's text: viewer on the left, text in the middle, tools on the right. It follows the same idea as Translation, applied to transcription.

Creating a document

From the Transcriptions area, choose "New document" and give it a title. You can also link it right away to a work already in the Library, searching by title in the same dialog — optional: without it, the document has no viewer, a single block of text. From a work's page in the Library you can instead create a transcription directly from that digitization, with the title already filled in. If the work has both images and a PDF on your computer, choose which to start from; with only one copy available there's no choice to make.

From a workspace Overview, the Transcriptions section shows only the documents assigned to that workspace. Open one from the list or use the + command to create one already assigned there.

The viewer on the left, one text per page

A document created from a work's page in the Library shows the linked digitization's page on the left, with zoom and panning — change page in the viewer, and the text on the right changes with it: each page has its own block of text and its own history. A document created from scratch in Transcriptions has no digitization to show: a notice appears in place of the viewer, not an error, and it stays a single block of text.

At the top, a document tied to a work presents it as the Library page does — author, year, place, and printer above, title below — with the link out to the library's site. The three-dot command offers only "Remove transcription": downloading, verifying, or archiving the work stay commands of the Library page, not of the Studio.

Writing and saving

Text saves after 30 seconds without changes. The indicator in the top right distinguishes unsaved text from a save in progress, completed, or failed, with a command to retry on failure. Changing page or leaving the Studio normally saves pending text immediately. A forced shutdown before a save can lose the latest edits.

When the text is ready, mark it as verified with the lock next to the "Page N" title: the text becomes locked, so an already-checked transcription doesn't get overwritten by accident. You can return it to draft at any time with the same command.

While the viewer is still opening the chosen page, a veil covers the text and history with a spinner in the middle: writing or restoring stay disabled until the page is actually shown. If opening fails, a warning appears in place of the spinner, with the same block.

Switching source, images or PDF

If the work has both copies, two commands in the viewer's bar switch between them. If the two declare the same number of pages, the switch is smooth: same numbering, the text follows. If the number doesn't match, switching to the secondary copy detaches the viewer from the text — browse it freely to find what you need, while the text pages through its own arrows, next to the "Page N" title. Switching back to the main copy restores the link on its own.

A third command detaches the link regardless of page counts, even while staying on the main copy: handy for glancing at a different page without moving where you're writing.

Automatic page reading (OCR)

In the OCR tab, on the right, you can have the text of the open page read by a connected AI model — the same mechanism already used for Translation, with providers and keys already configured. The command appears disabled until the conditions are met: the page must belong to a digitization linked to the document, and a provider and model that can read images must be chosen, either for the document or inherited from the workspace settings. The reason a command is disabled is always in its tooltip.

The text that comes back enters the history as a normal revision, marked "Automatic recognition": it never overwrites what's there, it stays editable like any other version, and a later manual correction simply creates the next revision.

The prompt that guides the reading belongs to the document: you edit it from any page and it applies to every page of that document, but to no other. A new document starts from the prompt chosen in the workspace settings, under OCR, where you can also load one from the prompt library. To reuse a prompt in another document, save it to the library from the edit command, then load it there. The reset command brings the document back to the workspace prompt.

The model receives the page image and the prompt, without the page number: the library's printed numbering rarely matches the position in the scan and would only confuse the reading.

Under the model, two small circles choose which image is sent:

• optimised: reduced to the long side chosen in Settings, Transcriptions tab (1500, 2000, 2500 or 3000 pixels), and recompressed. A smaller image is not enlarged;
• copy on this computer: the same image as the viewer, unchanged — the downloaded book's page, or the one saved in the cache while browsing online. Online it can be smaller than the optimised one, because the viewer asks the library for a ready-made size. If the copy is missing (cache emptied by its space limit, book only partly downloaded) the optimised one is sent.

The starting choice is in Settings; in the Studio you change it for the session, without it being saved in the document. With the right panel closed, the reading command stays under the reopen command.

While a page is being read its sheet is veiled and stays read-only, even after a cancellation request, until the job actually stops. Text you are still editing on another page is not replaced when reading finishes. A pill in the top row of the text column names the page being read, and stays visible even if you page ahead in the meantime.

The reading starts in the queue, like a download: you'll find it in the jobs panel while it runs, with the option to pause or cancel it the same way. A passing problem — the service asking you to slow down, a connection dropping for a moment — is retried on its own; a wrong key, a missing model or an empty answer is not, because retrying would give the same answer. An empty answer usually means the model found no text on the page.

Open the bottom panel and choose Transcription log — it only shows up here, inside a transcription document, mirroring the Translation log you see inside a project. Every reading leaves four rows: the start with provider and model, the image that was sent with its real size, weight and origin (downloaded book, cache or library), the full prompt that was sent (expandable row by row), and the outcome with duration, tokens, estimated cost and the number of the revision created. On top there are search, filters by row type and by level, and grouping by page.

History, summary and metadata

Every save stays in that page's history, in the panel on the right: it shows who wrote that version — manual correction, automatic recognition, or import — and when. The command on each history entry brings that version's text back as a new save, without overwriting earlier versions, even after a restore. Changing page changes the history shown, too. You can consolidate any version and give it a name, which you can edit later. Consolidated versions appear above ordinary saves without duplicating their text. Remove a name to move a version back to ordinary history. You can delete older versions one by one, or use Clear history to delete older ordinary saves. Consolidated, current, and verified versions survive the cleanup. The current and verified versions cannot be deleted. Deleting a version removes the ability to restore its text.

The Summary tab follows the Translation summary layout: it shows pages with text, word count, verified pages, and progress. It also shows completed OCR readings, tokens, and estimated cost.

The Metadata tab, next to History, shows the raw data saved for the current page: position, label, status, and revision count. It's there to show what's recorded today; how it's presented will change.

Current limits

Automatic recognition reads one page at a time, picked by hand in the Studio: there's no command yet to read a range of pages in one pass, or to read the whole document. It only works on digitizations linked through an IIIF manifest — a single-file work (PDF) isn't supported yet. The model doesn't yet receive the text of nearby pages as continuity context — only the current page's image and the configured prompt. The viewer doesn't yet have visual filters (brightness, contrast, inversion).