Web search
Perch searches the public web when a question needs current or outside facts, opens the pages it finds, and keeps web sources labeled apart from your own documents.
Perch answers from your own material first. When a question needs something your files do not hold, a current fact, a price, a release date, a page you have not read, Perch searches the public web and reads what it finds.
The short version
On Perch AI Web
Web chat has search built in, and it decides for you whether a message needs it.
- The globe beside the message box is Always search the web. Off, Perch searches only when the message needs it. On, every message searches.
- Left off, Perch still searches when the message reads like a lookup. A fast text match catches the obvious cases with no delay, and anything it misses goes to a small model that reads every language Perch speaks and turns your message into a standalone query.
That last step matters more than it sounds. What you type in a chat is not a search query. "so is it open now" is a reasonable thing to say and a useless thing to hand a search engine, so Perch resolves what "it" refers to from the conversation before searching.
In Perch Terminal and the CLI
Here search is one of the agent's tools, and the agent decides when to call it inside a turn. Two tools do the work:
- Search returns ranked results with their titles, addresses, and a snippet from each page. Perch asks for five results by default and can ask for up to ten.
- Fetch opens one address and reads the page as text.
Perch typically searches, picks the results that look relevant, and fetches those pages before it answers. Recent searches and page reads are cached for a few minutes, so a turn that comes back to the same source does not fetch it twice.
A page that returns nothing readable, because it is blocked or draws itself with JavaScript, falls back to the agent browser on the desktop app, which loads the page properly and reads it from there. That fallback is silent and rare.
For academic work Perch can also search a scholarly index, which returns papers with their authors, venue, year, and DOI rather than web pages.
Sources come back with the answer
Perch cites a web source the same way it cites one of your documents, and the two are labeled apart, so you can tell at a glance whether a claim rests on your material or on a page Perch found. See Citation verification.
One limit is worth knowing: the draft checker checks claims against the sources in your workspace, not against the web. A web citation tells you where Perch read something. It is not the same as the checker confirming it.
When one search is not enough
A question that needs several passes belongs on deep research instead. It plans the question into parts, searches your files and the web across rounds, and comes back with a cited report and a claim ledger. See Deep research.
Retrieval
How Perch grounds an answer in your documents, what it has learned about you, and the public web when a question needs it.
Memory
Perch keeps two kinds of standing context: instruction files you own in a folder, and account memory that follows you across chats and devices. Here is what each one does, how notes get saved and corrected, and how you see what a turn used.

