Search and fetch when the answer needs primary sources. Pick Web from the sidebar; off by default so chats stay private.
Web is the interceptor for questions that need fresh sources. Pick it from the sidebar and the deep-thinking layer gains two tools: web_search for finding candidate sources and fetch_url for reading them in full. The model decides when to use them.
It's off by default. Berges AI is a private chat first; browsing on every turn would compromise that. Web is the explicit opt-in for the turns when current information actually matters.
web_search returns title + URL + snippet for a query. fetch_url returns the main text of a specific page. The model typically searches first, then fetches the best result.
A single message can include multiple searches and fetches. The model loops through tool calls until it has what it needs, then writes the final answer with the URLs it used.
Every search and fetch shows up as a badge in your message bubble ("Searched: X", "Fetched Y"). You can tell at a glance whether the answer came from the web or just the model.
Tools never run on the fast layer. Easy questions stay snappy and offline; Web only kicks in once the cascade has escalated.
When Web is active, your search queries are sent to a third-party search provider (currently Brave Search). They process the query to return results and do not retain it for training. See the Privacy Policy for details.
Yes, when it actually uses the tools. Each search and fetch adds a round trip. A typical Web-assisted answer takes a few extra seconds compared to a non-Web answer.
Almost. We block private, loopback, and link-local IP ranges at the network level so the model can't poke at internal services. Public URLs work; size and time limits apply.
No. Results are fed to the model for the turn and discarded. Only your message and the final assistant reply are persisted (encrypted at rest, same as everything else).