What is sent to the model
Every question you ask goes to the model you chose, together with what the model needs to answer about your graph. For a cloud provider that means it leaves your computer; for your own server it goes only there. The same list stands in Settings ▸ Privacy ▸ What is sent to the model.
Sent with every question
- your message, and the earlier messages of the conversation that fit the context window;
- the graph: nodes, their inputs and outputs, links, positions (a very large graph in pages);
- what is selected and what is on screen;
- your recent actions on the graph;
- a summary of your catalog: which nodes and models you have;
- the app's instructions to the model and the descriptions of its tools;
- the values of the fields (prompts, seeds, model names), unless you turned Send field values to the model off in Settings ▸ Privacy. Values that look like keys are always hidden.
Sent when the assistant looks something up
- When it looks at a model: the description saved next to the model file, and what Civitai or Hugging Face said about it.
- What it read on the web. Pages are opened from this computer.
Never sent
"Never sent: model keys, the server access token, the model files themselves and the folders they lie in." Your settings file, your other projects and the conversation log are not sent either.
How to send less
- Turn off Send field values to the model.
- Turn off Allow web search and Allow reading web pages (Settings ▸ Assistant).
- Use your own server: nothing then leaves your computers.
- Choose no assistant: nothing is sent at all, and a question about an object is answered from the built-in knowledge in the Explain tab. See Working without a model.
- When the cloud provider's own web search is used, the provider sees the search words.
- Turn off Keep chat history in the project if you do not want the conversation stored.