How documents become answers you can trust.

Workspaces hold the work

A workspace is a folder for one body of work — a deal, a portfolio, a client. Documents and tables both live inside one, and access is granted per workspace, so sharing a workspace shares everything in it.

A workspace is private until you add a sharing rule. Rules grant Read, Read & Write, or Manage, to the whole organisation, a role, or one person.

Upload once, reuse everywhere

Dropping files into a workspace ingests them: each document is read, split into pages and indexed. That happens once, no matter how many tables later ask questions of it. PDF, DOC, DOCX, PPTX and XLSX are supported, as are JPEG, PNG and WebP images — an image is read as a single page.

This is why the order of work matters less than it looks. Adding a column to a table you built months ago costs no more than adding it on day one — the documents are already read. Ingestion is what page counts are charged on; asking more questions later is cheap. A document that fails to ingest shows up on your home page rather than going quiet.

A table is a question per column

Rows are documents. Columns are questions you write once and ask of every row. Each column has an output type — text, number, date, yes/no, a list, or one of a fixed set of options — so answers come back structured rather than as prose. A column can also answer with a picture: the page, or the part of a page, that shows what you asked for, cropped into the cell and citing where it came from.

Set the type to match what you will do with the answer, not how the document phrases it. A contract that says “this agreement shall continue for a period of thirty-six (36) months” is a number column, not a text one. Presets exist for common document types if you would rather start from a set of columns than a blank table.

Many documents, one row

A master contract and its amendments are one agreement, and they should answer as one subject. Group related documents into a bundle and the bundle becomes a single row.

A bundle holds up to 15 documents. The first three are included in the column’s rate; each one past the third adds to that cell’s cost, because each is read separately.

Every answer carries its source

Running a table answers each cell independently, so results arrive as they land rather than all at once. Every answer carries a citation back to the page it came from and a confidence score. Click any answer to open the cited page and read it in context.

Nothing here asks you to trust an answer you have not checked. The point of the citation and the confidence score is that spot-checking is fast and targeted rather than a full re-read. Reruns are scoped: select some rows, some columns, or a handful of cells, and only those run again.

Corrections are recorded, not overwritten

Type over any cell to override it. The original answer is kept underneath and the change is recorded, so a table stays auditable — you can always see what the model said, what a person changed it to, and who changed it.

Editing a column’s prompt marks every answer it produced as stale, ready to rerun. Home lists failed, stale and low-confidence cells, so none of them has to be hunted for.

For documents that keep arriving

Put a table in auto mode and it runs itself against each new document added to the workspace — useful when documents arrive continuously and you want the table to stay current without anyone remembering to press anything.