Web search
Off by default. When enabled, it lets a reply draw on live results instead of only what the model learned during training.
This is one of only two features that can send anything off your phone. It is worth knowing exactly what it sends before you turn it on.
What gets sent
A derived search query — not your conversation, and not your message verbatim. Before it leaves the device, the query is stripped of:
- email addresses
- phone numbers
- long strings of digits
The query goes to the search provider whose API key you supplied. That provider is a third party you have chosen and contracted with directly; what they do with the query is governed by their privacy policy. We never see it, because there is no server of ours in the path.
It does not run on every message
A router running on your phone decides, per message, whether live results are actually needed. A question about your own documents, or one the model can answer from what it already knows, does not trigger a search.
So enabling the feature is not the same as sending every message to a search engine.
Setting it up
In Settings → Web search:
- Turn the feature on.
- Paste your API key. It is stored in the system keychain, marked so that it does not sync to your other devices and does not travel in a backup.
- Read the note describing what leaves the device — it says the same as this page, at the moment it matters.
Documents take priority
If you have documents indexed and your question is about them, those are used and no search runs. Both features answer the same question — what should this reply be grounded in — and only one prompt can be sent, so the tie breaks towards your own files. They are more likely to hold what a question about them wants, and grounding in them costs nothing off the device.
Web search still runs for anything your documents do not cover.
Turning it off
Switch it off and nothing is sent, on any message. The key stays in the keychain until you remove it.