How Ask answers

Ask answers questions from the sources in one project's library. When you ask, Pangur searches the text of your sources for passages that bear on the question and hands those passages to a language model, which writes an answer and cites them with numbered markers such as [1] and [2]. Every citation leads back to a passage you can open at its page. The division of labour is worth keeping in mind as you read this page: Pangur finds and checks the evidence, and the model writes the prose.

Knowing how this works helps you judge an answer and predict what Ask will do with a question. The practical steps are on Ask a question.

What Ask can readCopy Anchor Link

Ask reads the text that Pangur holds for each item. That might be the text extracted from a PDF or another file, a transcription, or an imported edition, and how that text is chosen and indexed is set out in How Pangur reads your sources. A work whose file Pangur has not read yet, or which has no file at all, contributes only its catalogue details and abstract.

Ask will not answer about a work from its catalogue entry alone. If you ask about a work that has no full text, it tells you that the work is in your library but that no full text is held, and it offers to look for a copy online or asks you to add a PDF. A file you have only just added may still be being read, in which case asking again after a few minutes is usually enough.

Ask also starts from the library map, which is a short description of each work, written once from the work's whole text and stored. An entry records the work's subject, argument and structure, with page references. Pangur puts the map in front of the model before every question, so that the model knows what the library holds before it decides where to search. Map entries are written in the background by a model whenever Pangur reads a new or changed file, and that work counts against the project owner's daily model allowance. If the allowance is spent, the entries wait for the next day.

How Ask gathers evidenceCopy Anchor Link

When a model is available, Ask gathers evidence step by step rather than running a single search. The model sees your conversation, the library map, and a list of what else the project holds (analyses, corpora, summaries, drafts, exhibitions, arrangements and earlier conversations), and then chooses one action at a time. It might search the whole library or particular works for passages, read a named part of a work whole (an introduction, say, or a chapter), look at a work's record, its files or the works a PDF cites, read the records of a corpus filtered by person, place, type or period, open something you have made such as an analysis or a summary, or read whole works in the background when a question needs the whole of a book (see below).

The status line shows each step as it happens, and the finished answer keeps the record under How this answer was found.

These steps are bounded: six steps, 60 seconds in all and 36 passages for one question. If time runs out, or the model replies with something Pangur cannot use, Ask stops gathering and answers from what it has. If it has gathered nothing by then, Pangur runs one ordinary search instead. Because the model can only choose works that appear in the library map, it cannot cite a work you do not hold.

Answers without a modelCopy Anchor Link

When the server has no model to answer with, Pangur classifies the question with fixed rules and runs one search, then shows the matching passages with a note that the model is not switched on. In this mode the tab beside the home page's search field is labelled Find passages rather than Ask, and the iPhone app explains that Ask finds matching passages only. Which model Ask uses, and whether one is available at all, is decided by the administrator of your Pangur server.

Passages and citationsCopy Anchor Link

For a single search, Pangur considers up to 12 works and sends the model up to 12 passages. A PDF that nobody has opened before has to have its text extracted first, and Pangur does this for at most three such PDFs per question; the rest contribute their abstracts.

Pangur sends the sources to your screen before the model starts writing, so that if the model fails you still have a ranked list of passages to read. Once the answer is written, Pangur keeps only the passages the answer actually cites and numbers them in order.

Pangur also checks the answer against what it found. If you named a work or person that the cited passages do not account for, the answer says that it could not find those names and that it does not draw on them (names that an analysis has already recorded as people or places in your library are not reported as missing). If the answer claims that the library lacks something while works that matched went unquoted, Pangur adds a note naming those works. And How this answer was found lists the works that matched but were not quoted, works that were too long to read this time, and works that are still being read.

When nothing in the library comes close enough to the question, Ask says so instead of answering.

Whole works and summariesCopy Anchor Link

Some questions need a whole work read rather than passages from it. Ask recognises three such cases.

The first is a summary of one work. A request beginning "Summarise" followed by a title must name exactly one work, or Ask asks which one you mean. Pangur then reads the whole work in the background, which takes a few minutes, and the summary appears in the conversation and in the work's Outputs when it is done. For a work in several volumes, Pangur reads every volume; if today's allowance does not cover them all, it leaves some out and says which.

The second is a question about the whole of one or more works. Ask reads up to three of them in full in the background and assembles the answer from everything in them that bears on the question. The answer arrives in the conversation and in Outputs.

The third is a question about a named part of a work, such as chapter 3 of a book. Pangur finds that part in the stored text and reads it whole. If it cannot find the part, it says so and answers from the work as a whole.

While a background reading runs, the conversation shows its progress and updates itself when the answer is filed. These readings use the model allowance of the person who asked.

Manuscript leaves in answersCopy Anchor Link

When an answer cites a manuscript that has a facsimile, View the leaves opens the leaves beside the conversation. Ask may also offer to look at a particular leaf, saying what it would look for, with the buttons Read the leaf and Answer without it. Reading a leaf sends its images to a vision model and uses one of your daily manuscript readings (40 by default), which is why nothing is read until you choose Read the leaf. The limits are described in Reading pages with a model.

You can also ask Ask to transcribe a leaf, for example "Transcribe f. 3r of" followed by the shelfmark. This needs a facsimile whose images are ready, and it uses one of your daily two-pass transcriptions.

What Ask can make for youCopy Anchor Link

Ask can propose to make something: to analyse a source or a corpus, to create a corpus, a collection or a source group, to add works to a corpus or collection, to look online for a missing PDF, or to draft an exhibition. It always asks first, with Go ahead and Not now, and nothing happens until you agree. Ask only ever adds. It does not edit, rename, move or delete anything in your library, and if you ask it to, it tells you where you can do that yourself.

Some of these actions, such as an analysis, use your model allowance when they run, and the proposal says so. Drafting an exhibition needs publishing, which is not available on every account or deployment.

Limits and allowancesCopy Anchor Link

A question passes four checks before and during its answer. The figures below are Pangur's defaults; the administrator of your server can change them, and Plans, limits and usage is the reference for the figures that apply to you.

The plan check happens before the answer starts. The daily model allowance is counted in tokens, the units that models read and write. It covers Ask answers and the background readings described above, and summaries, Write's model tools and exhibition drafting draw on the same allowance. A full write-up uses a separate long-form allowance of its own.

Asking from an assistant or the command lineCopy Anchor Link

When an assistant such as Claude asks your library a question through Pangur's MCP tool ask_library, Pangur returns the passages and the library map, and the assistant's own model writes the answer. Pangur runs no model of its own for this, so the question counts against neither your plan's questions nor your model allowance. If the assistant then records its answer in one of your conversations with record_ask_turn, the answer appears in Ask labelled as written by the assistant, and it is not counted either.

On the command line, pangur ask asks exactly as the web page does and is subject to the same limits. pangur ask --materials returns the passages only, so no model runs and no allowance is spent. For the whole picture of who pays for what, see Models, assistants and who pays; for the tool itself, see The command line.

Conversations are personalCopy Anchor Link

A conversation belongs to the person who started it and to one project. Other members of a shared project cannot see your conversations, and you cannot see theirs. Each question runs its own search; what the conversation contributes is the earlier questions and answers and the works the last answer drew on, which is what lets a follow-up such as "what does she say about this?" be understood. Deleting a conversation deletes its questions and answers, and nothing in your library depends on it.