Web search

Every message is read by a set of deliberately narrow rules before it is answered. They are tuned so a needless search is rare — a search spends one of your daily look-ups and a moment of your time — while the questions that genuinely cannot be answered from training data get one without you asking.

What makes Vresk search on its own:

• You say so — “search the web for…”, “check online”.
• An instruction to go and look: a message that opens with “research…”, “look up…” or “find out about…”.
• Market questions — prices, tickers, exchange rates — and live events such as news, weather, scores and elections, when the phrasing is anchored to now.
• A window measured backwards from today. “The biggest releases of the past six months” has a different correct answer every week, so it cannot be answered from a fixed snapshot.
• A measured quantity pinned to the present. “The current population of Tokyo” answered from memory would be a confident wrong number, so it is looked up.

There is no search switch. Say what you want in the message and Vresk does it:

• A quick look-up — “search the web for…”, “check online”, or a question anchored to now.
• Proper research — “research … thoroughly”, “write me a deep research report on…”: a few minutes of real work, ending in a sourced report. More on deep research.
• A calculation — “what is 18% of 4,350”, or a short script to run: worked out, not guessed.
• A designed deck or document — the one thing you add by hand when your words did not say so: Design, under the Add menu beside the composer (the + button on a phone). On the free models you see the plan before anything is made; on the other models the model designs the deck itself.

Adding Design pins a small pill in the composer. It applies to your next message only, and tapping the pill puts the decision back on automatic. There is no standing mode to remember to turn off.

Your words win. Asking for a search in so many words always searches, whatever the automatic rules think, and a question that does not need the web gets an ordinary answer.

When a search feeds an answer, the sources appear as chips: a numbered marker in the text resolving to a small globe, the site’s domain and the page title. A chip opens the page in a new tab.

Chips are minted on our servers from what the search actually returned — the model only writes the marker. A web address the model types out in its own words is deliberately not a clickable link. That asymmetry is a safety property: every link you can click traces to a page a search really found, never to an address a model composed from memory.

The small card above the answer says what actually happened, and never describes a search that did not run:

Ask the same thing again quickly and the card may add “reused from a search in the last 15 minutes”: the sources were chosen for the earlier search and will not change until that window passes, even if you regenerate the answer.

Look-ups are counted per person over a rolling 24 hours: 20 as a guest, 100 with an account. Your limits does not show this count; you will know you have reached it because the search card tells you.

Running out never blocks an answer. Vresk keeps answering from what it knows, and the search card reads “Daily web searches used up” so you know the reply was not checked against the web. The other allowances, and the promise that none of them shrink silently, are on the limits page.