What you can index, and exactly how retrieval behaves. A knowledge base belongs to a project. Documents have their text extracted, then are chunked, embedded and indexed; workflows and Chat with Data retrieve from it under the asking user's permissions.
The upload dialog still lets you pick some refused types (images, .pptx, .doc, .msg, .epub). They are rejected during processing, not at selection.
source names a storage download action — Google Drive download_file, Cloud Storage get_object, Azure Blob get_blob or Microsoft Graph download_drive_item — with its connection and parameters. The file's bytes never pass through a step output. The Document → Knowledge base template is built on this.
Treat answer as untrusted model output: validate it with a condition or a human step before using it to post, write or query.
This site needs JavaScript for the full experience. Every page above is reachable without it.