Core concepts

What contains what

An organisation holds the credit balance, the people and the tier. Inside it are workspaces. A workspace holds documents and tables, and it is the level at which access is granted — sharing a workspace shares everything in it.

A table is a set of questions asked of a set of documents. Its rows are documents, or bundles of documents that belong together. Its columns are the questions. Where a row meets a column there is a cell, and a cell holds an answer, the citations it was read from, and a confidence score.

What a table actually is

The shape is worth looking at, because it is the whole product. Documents down the side, questions across the top, an answer at every intersection with a page reference attached:

Everything else in the application is in service of that grid: uploading fills the rows, writing columns fills the headers, running fills the cells, and the citation is how you check one without reading the document again.

Idea one: reading is paid for once

Uploading a document reads it, splits it into pages and indexes it. That is charged once, per document — not per table, and not again when you add a column six months later. Asking is cheap; reading is the expensive part, and it has already happened.

This is why the order of work matters less than it looks, and why it pays to upload everything relevant up front rather than only what the first table needs. See how credits work.

Idea two: nothing is recomputed behind your back

Columns carry a version, and every answer records the version it was produced under. Editing a column’s prompt or type bumps that version and marks the answers it produced stale — they still show, they are just answering a question you no longer asked. Rerunning is your decision, because rerunning spends credits and because a silent rewrite is not something you could audit.

The same principle covers your own edits. Typing over a cell does not delete the model’s answer; it records yours alongside it, with who changed it and when. See corrections and running and reruns.

Vocabulary

The words used throughout these docs, and in the application. Where the interface and the underlying data model disagree, the interface’s word is the one used here.