Deep research
Ask Vresk to research something properly and it does not answer from memory. It plans, searches in rounds, reads pages in full, mines the data it finds, and writes a report that cites its sources — a few minutes of real work. It starts when you ask for it in so many words, and the card in the thread shows every step as it runs.
A different thing from web search
Web search is Vresk checking a fact mid-answer: a query to our search providers, a handful of results, citation chips on the reply. Deep research is a project. Your question is split into sub-questions; each is searched, and the most promising pages are read in full rather than skimmed as result snippets. Tables and numeric series found along the way are mined out and kept. When the reading is done, everything is written up as one report — and the draft is checked against the sources it cites before you see it.
All of that takes minutes, not seconds. You are not left staring at a spinner while it happens: the card in the thread narrates every step as it runs, which is the next section.
It starts when you ask
Research starts from your message. Ask for it plainly — “research … thoroughly”, “write me a deep research report on…”, “a detailed document on…” — and the run begins at once. The sub-questions it chases come from what you named in the ask, so the ask is where you steer: name the two or three things you care about, and those become the questions.
When your message only hints at research — a broad question that could be answered either way — Vresk answers plainly first and offers “Research it properly” as a chip under the reply, with the time it takes. A tap runs the same question as research. Nothing that takes minutes starts on a guess.
Every run counts against your daily allowance (below). Over it, the run stops before searching and the reply says so in plain words — “You've used up today's research runs, so I stopped before searching. Nothing was spent on this one.” — rather than pretending to have researched anything.
Watching it work
While the run works, the card in the thread is a live timeline. Collapsed, it shows the step happening right now and how many steps the run has taken; expanded, the whole run so far.
| What the step says | What is happening |
|---|---|
| Searching the web (round 2) | A fresh set of queries. Runs go in rounds — later rounds chase what the earlier ones turned up. |
| Reading example.com (3 of 14) | Pages being read in full. The counts are real — this is the third of fourteen pages in the round. |
| Found data on example.com | A table or numeric series was mined from a page. The step carries a small preview of the rows, and a small chart when the numbers form a series. |
| Filling in gaps | The run noticed something missing from its reading and went back for it. |
| Writing the report | The reading becomes one document. |
| Checking the sources | The finished draft is read back against the pages it cites. |
Everything on the timeline is something that actually happened. There is no percentage bar and no estimated time remaining, because the run knows neither — and a step that failed shows as failed rather than quietly disappearing.
The report, and where it came from
The run ends in a report: a document with numbered citation markers through the prose and a Sources section at the end. Each source is listed with its site, its title, the page's own publication date when the page carries one, and the full address written out so you can always copy it. The list is stamped with a “Retrieved” date — the day the run gathered its pages — so you know exactly how fresh the reading is.
That list is compiled by Vresk from the run's own record of what it searched and fetched. The model does not write it and cannot add to it, and a citation number in the text that does not match a real source is removed rather than left pointing at nothing.
Not every source could be read
From report to designed document
A finished report carries a “Design this” button. It lays the same report out as a designed document — cover, headings, your brand — and it is layout, not a rewrite: not a word of the report changes. Pictures are a separate checkbox, off by default, with the cost stated in plain words before you tick it. More on designed documents.
What a run costs you
Every account gets 3 deep research runs a day, free. That is a declared floor, not a trial: if it ever moves, the change appears on the changelog before it takes effect, never silently. How all the allowances behave — which limits refuse and which degrade — is on the limits page.
A run that stops at the allowance is not a run. Over the day's allowance, a research ask stops before searching and the reply says so — “You've used up today's research runs, so I stopped before searching. Nothing was spent on this one.” Nothing is spent on a run that stops there.