Research over the files you choose

Upload folders of documents to a project's corpus, pick any mix of folders and files, and ask a question that only those files answer. Every claim in the report is checked against the files you picked, and every citation opens at the place it quotes.

This page covers what you can upload, how folders and selection work, what "only these files" does and does not guarantee, and what you see when your files cannot answer. It applies to the desktop app (Apple Silicon Macs) and the server alike.

What you can upload

Anything else is refused by name, with the list above, rather than read as text: a format whose quotes cannot be located has no place in a corpus whose citations are checked.

What is read, exactly. Spreadsheets are read as the values a spreadsheet program shows, not their formulas; hidden sheets are read and marked hidden. Word documents are read with tracked changes accepted; footnotes, comments, headers and footers are not read. Notebook outputs keep their text and say how many images were left out. Scanned PDFs need text recognition first; a PDF with no text is refused, and pages without text are listed.

Tables, slides and cells become text. A spreadsheet row is read as Row 14 — Item: Staff; Cost: 1200, with its header on every row, so a quote from it makes sense on its own. Quotes from these files are marked as coming from a text rendering of the file.

Files are limited to 25 MB. Office files that would expand past 200 MB, or that are password-protected, are refused with the reason.

Folders

Choose Choose folder, or drop a folder onto the Corpus page, and its subfolders come with it. Files upload one at a time; Stop uploading lets the file being sent finish and sends nothing after it. The same file in two folders is stored once and listed in both. A subfolder your computer does not let the app read is left out, and the app is not told it was there: compare the upload list with the folder if something is missing.

Each folder in Choose files to research has Rename, Move and Delete. Renaming or moving a folder changes nothing that cites its files.

Deleting a folder removes its subfolders and the files filed only there, with every version; a file also in another folder stays. Before anything is deleted, the confirmation says how many files will go. Research that cited a deleted file keeps the quotes it gathered, but the file itself can no longer be opened from it.

Each project has its own corpus, on the desktop app as on a server. Documents uploaded to the desktop app before projects had separate corpora were moved to a project named Documents from before 3.1, and the original store is kept as corpus.legacy-3.0.sqlite in the app's data folder.

Choosing files and asking

Tick folders and files in Choose files to research — a folder's box selects everything beneath it — then Ask about these files. The research form opens with your selection attached and Only the files you selected switched on.

The selection is fixed when the research starts: the exact version of each file is recorded, and a file you replace, delete or add afterwards does not change what that research may read. The run page lists the files it used.

Two guarantees, and what each one means

Only the files you selected limits the evidence. Every search, every passage read and every citation comes from the selected files; nothing from the web, another project or an auto-saved report can become evidence, and the report's own record and its exported bundle say which files were selected. It does not keep your question private: the question and excerpts from your files are sent to whichever model provider each agent is set to use.

Offline limits where the models run. Turn it on and the research refuses to start unless every agent uses a model at an address on this computer — Ollama, or a custom endpoint on this machine — and names any agent that does not. Offline research never reads the web.

The check is on the address, because that is all the app can see. A custom endpoint on this computer that is itself a router to cloud providers (OmniRoute, LiteLLM and the like) passes the check while your question and excerpts leave the machine through it. Only turn on Offline for such an endpoint if it forwards to models on this computer.

The two are independent: you can have either without the other.

How the answer is checked

Every sentence of a report on selected files is split into the individual claims it makes, and each claim is checked against the passages its citations point to:

• numbers and dates must appear in those passages;
• a citation that does not support its claim is removed from it;
• a conclusion must follow from the findings it rests on.

A passage holding a claim your files do not support is rewritten from what they do support, and checked again. A passage that still fails is removed, with a note saying so where it was. The run page shows how many claims were supported, kept out and removed.

When your files cannot answer, the research ends with "Your selected files don't answer this question. Nothing was answered from outside them." — not with an answer from the model's own knowledge.

When a claim could not be checked — the checking model was unavailable, or the spending limit ran out part-way — the research ends without a report rather than presenting unchecked claims as checked. The checking calls count against the same spending limit as the research.

Opening a citation

In a report's sources, Open where it's cited shows the file's text with the cited passage highlighted and named by its page, sheet and row, slide, cell or line. For a PDF, Open the PDF at page N opens the original file at the cited page, in a viewer that supports jumping to a page; any other viewer opens it at the first page.