Reading pages with a model

Transcribing a page image involves two separate jobs, and Pangur does them with different software. First, something has to find where the lines of writing are on the image. Then something has to read each line. Pangur finds the lines with a line-detection program that runs on its own server. It reads the lines with a vision model, which is a language model that accepts images, hosted by an outside provider that the administrator of your Pangur server chooses. Only the reading counts towards a daily limit.

A connected AI assistant can do the reading instead, with its own model. In that case Pangur still finds the lines, but none of Pangur's models is used and nothing counts against your limits.

Line detectionCopy Anchor Link

Line detection finds the baselines and outlines of the lines of text on a leaf. Pangur uses Kraken, an open-source program for historical documents, running on Pangur's server. Because Kraken is not a language model and reads no words, line detection has no daily limit and does not count against your plan. You start it with Detect lines on the transcription desk, and a model reading starts it on its own when a leaf has no lines yet.

If the server has no line detection set up, Pangur says so when you try, and model reading on the desk is unavailable as well, because it needs the detected lines.

Transcription with a modelCopy Anchor Link

Transcribe with AI on the transcription desk reads a leaf in two passes. For each detected line, Pangur cuts a strip of the image at full resolution and sends the strips to the vision model along with a reduced view of the whole leaf. In the first pass the model reads each line, and in the second it checks each line against its image. Dense leaves are sent in several requests of up to 20 line images each.

Pangur also gives the model whatever context it has: up to three lines you have already corrected on this manuscript, as examples of the hand; the language and period of the writing, taken from the manuscript's catalogue description or the item's record; and your reading rules together with the guidance you gave for each text type.

For a printed book (a book item, or one with an ESTC number), Pangur tells the model to read the page as type, to keep the long s and ligatures, and to file catchwords, signatures and running titles as page furniture.

The model may offer up to two alternative readings of a line, which you can choose between in the markup menu. Lines you have corrected are not overwritten by a later reading. Every reading is a draft for you to check against the image, since a model can misread a hand, and the desk shows each line beside its region so that you can.

If the model returns nothing usable, Pangur says so and suggests trying again or choosing another model. If it can read none of the lines, Pangur suggests trying again or supplying a transcription to align.

Reading one leaf with a questionCopy Anchor Link

The facsimile viewer in an item's Read tab offers a second, simpler way to use the vision model: Read this leaf with the vision model. Type a question about the leaf, or leave the field blank to have the leaf transcribed, and choose Ask of this leaf or Transcribe. Pangur cuts the leaf into tiles at full resolution and sends up to six of them, in reading order from the top left, in one pass. A large leaf may therefore need two readings, and the result shows where the first one stopped. The reading appears below the viewer and is not saved as a transcription. It reports the tokens used and the time taken.

Ask uses this same reading when an answer needs to look at a manuscript leaf, and only after you agree (see How Ask answers).

Daily limitsCopy Anchor Link

The two kinds of reading have separate limits, each counted per person over 24 hours:

The limits are kept small because each reading sends a whole leaf's images to the model. When you reach one, Pangur tells you which limit you have met. The limits are the same on every plan and are separate from your plan's Ask questions. The figures above are the defaults, which the administrator of your Pangur server can change, and every limit is listed in Plans, limits and usage.

The limit belongs to the person who starts the reading, so in a shared project a reading you start counts against your own limits, not the project owner's.

When reading is switched offCopy Anchor Link

The administrator of your Pangur server decides whether reading with a model is available at all, and which model is used. When it is not available, the desk's button says that reading with AI is not switched on for this server, the one-leaf reading is hidden, and a request is refused with a similar message. Line detection, transcription by hand and supplied transcriptions still work.

Reading by an assistantCopy Anchor Link

An assistant such as Claude, connected to Pangur over MCP, can transcribe leaves with its own model. Pangur never runs a vision model for an assistant, so this uses neither your daily limits nor your plan. The assistant works in two steps:

1. transcribe_manuscript_leaf prepares one leaf. Pangur detects its lines if necessary and returns each line's identifier and region, its current text, whether an editor has corrected it, and instructions for reading it. The assistant then looks at the leaf's images with manuscript_leaf or manuscript_leaf_tiles.
2. save_manuscript_leaf_reading saves the assistant's reading onto the detected lines, with a diplomatic and an expanded reading for each.

If the leaf already had text, Pangur first saves a version named "Before agent reading of leaf" followed by the leaf's number, so that you can restore the earlier text. Lines an editor has corrected are kept. The saved lines are marked as model readings, and the text is indexed for Ask straight away.

The command-line tool works differently. pangur manuscript transcribe ITEM CANVAS asks Pangur's server to do the reading with Pangur's own vision model, exactly as the desk's button does, and it counts against the same daily limit of 20. For a view of which actions use Pangur's models across the whole product, see Models, assistants and who pays; connecting an assistant is covered in Assistants and Pangur.