What the AI sees and what is sent to your provider

What goes in each request

The brief that Omne writes for the model contains:

• The name of the workspace and today's date.
• The connected accounts that are switched on for this chat, as a service and a handle, for example a service name and the account's username. The AI is told to use them to answer with real data, to say which account a number came from, and to read before it writes.
• A note on any account that could not be reached just now, so the AI can say so if your question needed it.
• The rule that a tool which changes something waits for your yes, and that if you decline, the AI must not try another way.
• An instruction to answer in your language, plainly and briefly, in Markdown.

After the brief come your messages and the AI's earlier replies in this chat. A chat can carry up to 80 messages, and each message can be up to 60,000 characters long. When the AI calls a tool, the tool's answer is added too. A long answer is cut at 24,000 characters, with a note that says how long it was, so one huge result cannot fill the whole conversation.

What stays out

Ask AI through a provider has no tools for your Library, your Planner, your posts or your store yet. The tools it has come from the accounts you connect. So a question such as "what did I save about pricing?" cannot be answered from your Library in a provider chat. The AI will tell you it cannot see it, or answer from general knowledge, so check where an answer came from when it matters.

An agent you run on your own computer is different. It gets tools to search and read your Library, save to it with your approval, and list your connected accounts. See What an agent can do in Omne.

Chats are not stored

Omne does not keep your conversations. The page holds the chat while it is open. Choosing New chat or leaving the page ends it, and you cannot bring an old chat back. Each message, though, goes to your provider, and what the provider does with it is set by your agreement with that provider. Check its settings on data use and training. Omne has no model of its own to train.

What Omne does remember is small and stays in your browser: the model you picked, the agent you picked, and which connectors are off, all per workspace on that device.

Tools and your accounts

When a tool runs, Omne opens a session with the service using the token you gave it, runs the call, and closes the session at the end of the turn. The service sees a request from your account, just as it would from any app you authorized. Tokens and pasted keys are sealed on Omne's servers and are not sent to the model. The model sees only the tool's name, its description, the details it chose, and the answer.

What the model can get wrong

A model can misread a tool's answer or invent a detail when the tool returned nothing. The brief tells it never to guess a number, but the habit is worth checking. Ask it which account an answer came from, and open that account when the number matters.

Content written by other people, such as the text of a post or an email returned by a tool, goes into the model as data. That is why changes always wait for your approval, even when the model seems sure.

Perplexity and tools

Perplexity answers from the web and takes no tools, so on Perplexity the brief and the conversation are all that is sent.