Files

You can upload files and use them as context in chats.

Supported Formats

• Images: PNG, JPG, GIF, WebP
• Documents: PDF (text extraction)
• Text: MD, TXT, CSV, JSON, code files
• Office: DOCX, XLSX (text and table extraction)

Limits

• Max 10 MB per file
• Max 5 files per message (files and research documents combined)
• Of these, at most 3 research documents per message

Usage

• Upload via the paperclip in the composer or Cmd+U.
• Preview in the composer (thumbnails for images, badges for documents).
• Files and research documents can be attached together.
• Project assignment is possible; files appear in the sidebar.
• Context for the response is built from the extracted text.

How files become context

On upload, the content of each file is extracted server-side: text and code are read directly, while text and tables are pulled out of PDF, DOCX, and XLSX. When you attach a file to a message, that extracted text flows into the chat as context. Images are the exception: they are not extracted as text but passed to a vision-capable model as an image (see Vision).

Attached file context applies to chat responses. Agents and research do not use it directly; their results are stored as research documents that you can later attach in a chat.

Project files and attached files

• Project files are assigned to a project, appear in the sidebar, and can be attached again and again in that project's chats.
• Directly attached files are added to a single message as you write it.

In both cases, a file only becomes context once it is attached to the message in question. Files and research documents can be attached together (see Research): files provide raw material, research documents provide already-prepared, reusable knowledge from earlier runs.

Best practice

Attach the relevant files specifically, not entire archives. Very long extracted text is truncated; a few focused files give better context than a large, unspecific pile.

Security

• Validation via MIME type and magic bytes (protection against misdeclared files).
• Storage in private object storage (private bucket, delivery via short-lived signed URLs).
• Access is restricted to your own account (Row-Level Security).

Images & Vision

Uploaded images can additionally be analyzed by the model, see Vision.