Reference files with @
Pull any uploaded file — or specific pages of it — into your next message.
What this is
Inside a workspace chat, typing @ opens an autocomplete of every file in the workspace's context
library. Pick one and Curie inserts a reference token that attaches the file to your message. The assistant
can then read the file when generating its reply.
You can reference a whole file, or only specific pages of a PDF — handy when you want to quote a contract clause or ask about chapter 3 of a manual without loading the entire document.
Step by step
- Open a workspace chat. Either open the workspace and pick a shared chat from the sidebar, or go to one of your chats and share it into a workspace first.
- Start typing in the composer and hit @. A popover appears with every file in the workspace. Keep typing to filter by filename.
-
Pick a file.
Curie inserts a reference token that looks like
@filename.pdfin the composer. Under the hood that's an@[filename.pdf](ctx:123)reference for a context file, or@[filename.pdf](doc:123)for a document from a knowledge base or a chat upload. -
(Optional) Limit to specific pages.
For long PDFs you can narrow the reference to a page range — e.g.
@[manual.pdf:pages=5-12](ctx:123)— so the model only reads those pages. - Send the message. The sent message shows a paperclip badge for each referenced file. Click a badge to download the file.
Good to know
- The autocomplete lists context files, notes and snippets (not URLs), plus every document the chat's agent can retrieve from: files uploaded to the chat and the files in any knowledge base attached to the agent profile. A document row shows its knowledge base after the filename and says "not indexed" while it is still being processed.
- Mentioning a document points the assistant at that exact file — useful when two knowledge bases hold a file with the same name. Page ranges apply to context files only.
- Referenced files count toward the token budget of the call, so big PDFs are cheaper when you scope them by page range.
- For jobs bigger than a single reply can handle, consider chunked PDF-to-Markdown conversion or multi-PDF extract instead.