Skip to main content

Data flow

For the short version, read What is sent. This page adds the mechanism.

What leaves your computer with each request​

The local assistant sends a model request to the OfficeLabs cloud, which forwards it to the model you picked. A request contains:

  • your messages and the conversation so far
  • the list of tools the model may call, and any tool choice
  • the document context in the instructions: title, document type, cursor position, outline, the first 300 characters of your selection, and the first 3000 characters of the document
  • tool calls and their results, including any document text the model reads with a tool
  • a rendered image of the document, when the model uses a screenshot tool

Large older tool arguments are shortened before they are sent again.

Everything is sent over an encrypted connection. Sign-in and token refresh go to the same service.

What the service keeps​

For each call, the service keeps only the metering fields described in How usage is measured: account, model, token counts, cost, timing and status. It does not store your prompts or the model's answers. Errors are logged with the model name, not the content.

The backend uses no product analytics or crash-reporting services.

What stays on your computer​

  • Your documents. They are read and edited in place by the document bridge. Only the document context listed above and the parts the model reads with a tool travel in requests.
  • Your sign-in. See Sign-in.
  • Traces. Traces stay on your computer. They leave it only if a developer has configured an external tracing service; a normal install has none.

The model provider​

The model you pick receives the request from the OfficeLabs service. See Processors and Retention.