Documents: PDF, Word, slides and spreadsheets

Research papers, a brand book, the deck a client sent, a spec sheet — these used to need copying into plain text before anything could use them. Now they go straight into the media library and are parsed on your machine as they come in: open one and read the full text, let the agent read it, wire it into a board or a workflow, or turn it into an editable note in one click.

Importing#

Use Import in the media library or drag files into the window — the same way as video and images, several at once if you like. Recognised formats:

Kind	Formats	Split into
PDF	.pdf	Pages, each with a page image
Slides	.pptx (also .ppt with LibreOffice installed)	Slides: title, bullet points, speaker notes
Word	.docx (also .doc with LibreOffice installed)	Sections
Spreadsheets	.xlsx (also .xls with LibreOffice installed), .csv	Sheets
Other text	.md, .txt, .html, .epub	Chapters or the whole text

The library's type filter gains Document, and cards show the format and page count. Files the library doesn't take, such as archives, are refused at import with a note on what can be imported, instead of being filed as video.

Parsing on your machine#

Each import is parsed locally — nothing leaves your computer and nothing is billed. The result is Markdown split by page, slide, sheet or chapter, plus the document's embedded images; every PDF page also gets a rendered page image.

• With LibreOffice installed, two more things happen: slides and Word pages get layout images too (so the agent can see what a slide looks like), and the legacy doc / ppt / xls formats are converted and parsed. Without it you get text and embedded images, and the detail view says so.
• Very long documents get page images for the first 200 pages (all the text is still extracted); very large tables keep the first 2000 rows and 40 columns. Both are stated in the detail view rather than cut silently.
• Parsing queues in the background, so several PDFs imported together take turns. Stop halts a parse in progress; it is recorded as stopped, not failed, and you can parse again any time.
• Local parsing handles born-digital documents, not scans or complex layouts. When very little text comes out, the detail view suggests it may be scanned or image-only — that's the time to use another parser (below).

Reading a document#

Opening a document asset shows a three-column reader: the parsed text on the left, every page of the original in the middle (Original document · read-only), and metadata on the right. Click a section heading on the left and the middle column scrolls to that page. The header says which parser produced it and how many pages (slides, sheets, sections) it has.

Clicking a document attachment in the agent's composer, in a message or in a citation opens the same reader.

Parsing it again with another parser#

Scans, image-only PDFs, multi-column layouts and documents full of formulas and complex tables are hard for the local parser. With a document-parsing plugin installed (MinerU document parsing, which ships with the app, is one):

• For one document: press Parse again in the reader and pick an option such as Parse with MinerU. The new result is stored as another parse of the same document; the agent, boards and workflows use the latest successful one.
• As the default from now on: choose it under Settings → Capability providers → Document parsing.

Imports are still parsed locally first; sending a document to a cloud parser only happens when you name it — installing a plugin never uploads documents quietly. Setting up MinerU, its network option and its quota are covered in Using plugins.

Save as note#

To edit the text or have AI keep writing from it, use Save as note — in the library's right-click / ⋯ menu and in the reader. The full text becomes a new note, the document's images enter the library only at this point, and the note's sources point back to the original document. The document itself stays in the library unchanged.

On boards, in workflows and with the agent#

• Boards: drag a PDF / Word / slide deck / spreadsheet onto a board (or paste it, or pick one under Add → From library → Media) and you get a document item holding the parsed text. Wire it into a writing or generation item and that text is what the model reads (the backend reads it along the connection; one that is still parsing or failed to parse is reported, not skipped). Save as note on its action bar swaps it for an editable note you can ask AI to write in. See Creative boards.
• Workflows: the Document to Markdown node turns a document into Markdown text, optionally only some pages or slides (From section / To section), and can name the parser to use; left empty it uses the latest existing parse, or parses locally first.
• The agent: attach a PDF / Word / slide deck / spreadsheet / Markdown file right in the chat (📎, or drag it from Finder into the conversation). Short ones are handed over whole; long ones come with a table of contents and it reads section by section as needed, picking up big tables mid-section. When it needs to see a page's layout, a chart or a screenshot, it passes those page images to a vision model. See AI Studio & agents.