How analysis works

Analysis reads a primary source, such as a letter book, a cartulary, a register or a volume of sermons, one division at a time. For each division it records the people, places, organisations, works and events the text names, the role each one plays, and the exact words that support each finding. Readings of several sources are gathered into a corpus, so that you can follow a person, a place or a date through all of them. You find Analysis under Analysis in the navigation and on each library item's Analysis tab.

Analysis suits sources whose value lies in who, where and when. For questions about what your secondary literature argues, Ask is usually the better tool.

What a reading containsCopy Anchor Link

A reading is one analysis of one source, kept as a record of what was read and what was found. It is made up of four kinds of thing.

Units are the divisions the source was read in: a letter, a charter, an entry, a sermon. In the corpus views a unit is also called a record. A mention is one occurrence of an entity in a unit, and it carries the entity, its role in that unit (for a letter, the roles are sender, recipient, bearer or mentioned) and a verbatim quotation. An event is something the text says happened, with its date as the source gives it, its place, its participants and a quotation. A relation is a stated connection between two entities, such as kinship or patronage, again with the quotation that states it.

An entity is the person, place, organisation, work, concept, event, object or other subject that mentions refer to. Entities belong to the project rather than to a reading or a corpus, so one person named in three sources is one entity, whichever corpus or reading the mentions came from.

How Pangur reads a sourceCopy Anchor Link

Pangur's server reads a source in five stages.

1. Divide. Pangur divides the text into units without a model, so this stage costs nothing. It uses the document's own headings and numbering where they exist. If you have given a marker, such as "To the same", it starts a new unit at each line that begins with those words. If neither applies, it reads the text in even sections of about 6,000 characters. A unit longer than 24,000 characters is read in parts, labelled "part 1 of 3" and so on, and a document can have at most 1,200 units.
2. Recognise the kind of document. One short model call decides what the document is, such as a letter book or a cartulary. The kind determines the roles the reading uses: a charter has granter, grantee, witness, superior and notary; a testament has testator, executor, legatee, witness and cautioner; a court record has pursuer, defender, judge, procurator, witness and assize.
3. Extract. The model reads the units in small batches and returns mentions, events and relations, each with a quotation.
4. Check every quotation. Pangur looks for each quotation in the stored text of its unit. A finding whose words cannot be found in the document is kept but marked: it is flagged for review and left off the timeline and the map.
5. Join entities. Pangur joins a mention to an existing entity only when the names match exactly. The model is also given the names your project already holds and may file a mention under one of them, but only when the unit's own words settle the question; these filings are listed for checking under Identified as read in a corpus's Review view. Looser matches, such as two spellings of one surname, are left for you to decide.

Each unit can hold at most 40 mentions, 25 relations and 20 events.

A reading is a hypothesis you reviewCopy Anchor Link

The model's findings are a draft. A person reviews a reading on the review desk, confirming, rejecting or correcting mentions, adding what the model missed, and deciding which names belong to one person. Each correction is recorded against the passage it concerns rather than against a particular reading, which is what allows Pangur to apply your corrections again when the source is read a second time. The desk itself is described in Review a reading.

Every mention and every exported row records whether it came from the model or from a person, and a mention you add yourself is labelled "added by hand".

Corpora are groupingsCopy Anchor Link

A corpus is a named set of sources that you read and explore together. It holds a name, a description, the sources in it and the reading guidance you give. The readings themselves belong to the sources and the entities belong to the project, so deleting a corpus removes only the grouping: every reading and every correction stays, and a source can be in several corpora at once. The day-to-day tasks are in Work with corpora.

Reading a source againCopy Anchor Link

You can read a source again, for example after correcting its transcription or choosing a different way to divide it. Pangur makes a new reading from scratch and then applies your earlier corrections to it. Anything new that the second reading finds is unreviewed. Until the new reading is finished, the corpus views go on showing the previous one.

If the source's text changes after it was read, its Analysis tab tells you that the text has changed and invites you to read it again. The reading itself stays pinned to the exact text that was priced and read, so adding a file later does not alter what an existing reading contains.

Who pays for a readingCopy Anchor Link

Analysis runs a model on Pangur's side, so a reading counts against the daily model allowance of the person who starts it. It is the most expensive action in Pangur per document. A 300-page volume comes to about 220,000 input tokens, which is more than half of the default daily allowance of 400,000 tokens. Pangur shows the model and an estimated cost before you start, and the item shows the actual cost afterwards.

If the allowance runs out part-way through, Pangur keeps what it has read and marks the reading as partial, saying how many batches were read ("12 of 30 batches were read before the daily allowance ran out."). You can read the source again the next day, after the allowance resets at midnight UTC. Limits differ between plans and an administrator can change them, so check the figures that apply to you under Plans, limits and usage. The wider picture of what uses a model and whose allowance it draws on is in Models, assistants and who pays.

Two pieces of corpus work are charged to the project's owner rather than to the person looking: the background corpus digest, and the test of a concept you suggest. Both are described in Explore a corpus.

An assistant such as Claude can also analyse a source with its own model and submit the results to Pangur, which costs nothing against your Pangur allowance (Source inventories and assistant-submitted analysis).

When analysis is not availableCopy Anchor Link

Pangur refuses to read a source in a few situations. An administrator can switch Analysis off for a deployment, in which case Pangur says that analysis is not enabled. If no model is configured, the reading fails with a message to that effect. An item needs text to be read: Analysis uses an item's extracted text or its transcription, and an item with neither is marked as having no readable text yet, so you need to transcribe its pages or add a file with text (How Pangur reads your sources). If the project is limited to one collection, items outside that collection are marked as outside the project's collection and cannot be read. And in an organisation's project, a member with the viewer role can read readings, explore corpora and export results, but cannot start a reading, change a corpus or correct a reading (Project members and roles).

Analysis on other surfacesCopy Anchor Link

The web client and the Mac app, which shows the same pages, have all of Analysis. The iPhone and iPad app has its own Analysis screens, including corpora, the ESTC import, the review desk and the corpus views (Pangur for iPhone and iPad). Assistants connected over MCP and the command line can read corpora and their records, and can submit analysis made with the assistant's own model, but they cannot start one of Pangur's own readings.