Access Control

​

Permission Inheritance

Omni respects the permissions from your source systems. You only see content you already have access to in the original application - if you can’t view a document in Google Drive, you won’t see it in Omni either.

​

How It Works

Every document indexed by Omni carries a permissions record alongside its content. The record has three fields: public, a list of authorized user emails, and a list of authorized group identifiers. Group memberships are synced from each connector and stored against your user profile. When you search, Omni:

1. Finds matching content across all sources
2. Filters results to documents where any of the following holds:
   • The document is marked public
   • Your email is in the document’s user list
   • Your email domain matches a group in the document’s group list
   • You belong to one of the document’s listed groups
3. Returns only content you’re authorized to view

The same permission filter applies to the AI assistant’s search tool — it cannot retrieve or cite any document you wouldn’t see in Search. People-directory lookups and chat-history lookups are separate capabilities with the scopes described below.

​

People Directory

The people directory is separate from indexed documents. The search_people capability searches active directory records and can return organization-visible profile fields supplied by connectors, such as a person’s title, department, manager, or office. It does not use a document’s public/users/groups permission record.

​

Chat History

search_chats and read_chat expose only the current user’s previous Omni chats. A user cannot use them to read another user’s conversations, and chat content is treated as untrusted context rather than instructions.

​

Source-Specific Permissions

Most connectors mirror a source’s own access model. A few behaviors are worth calling out: Users without a valid per-user credential for an MCP-backed connector simply cannot use its MCP actions until they authorize the provider.

​

Projects

Projects are personal workspaces, not a sharing mechanism. Project instructions only shape the model’s behavior, and context documents must be documents you can already read. The assistant’s search inside a project is still permission-filtered, so a project never widens access beyond what you already have. See Projects.

​

FAQ