Knowledge base / Research and tools

Deep research

One message becomes a research task: Nymbot plans the questions to answer, searches and reads the web in several rounds, and writes a long report with sources you can open.

Starting a research task

Turn on the Research chip in the toolbar and send your question, or start the message with ?research. The chip turns itself off after one message, so the next thing you send is an ordinary reply again.

Research needs a Pro model pinned, because it makes many searches and model calls. Without one, the app says so and sends nothing. Before you send, the composer shows which model will do the work, what it will probably cost and the most it can cost.

What it does

1. Plans. The model breaks your question into the smaller questions a good answer has to cover.
2. Searches. It searches the web for them in up to four rounds, with different phrasings each time, including news searches for anything recent.
3. Reads. It opens the most promising pages, up to twenty in all, and notes what each one says against the question it answers.
4. Writes. It turns those notes into a structured report.

While it works, the reply shows each step as it happens: the questions it planned, what it searched for, which site it is reading and how many findings it has noted.

A long task can pause partway and carry on in a second step. The app picks it back up by itself, within the same maximum price.

The report

The report cites its sources with numbers, and the numbers match the sources list under it in order, so [3] is the third source. When there are more than two sources the list is folded into one row of site icons; tap it to open the numbered cards.

A report is only as good as what it found. Open the sources behind any claim that matters to you.

What it costs

Research is paid from your Pro balance and metered on the tokens it actually uses, across every step, like any other Pro reply. It never costs more than the maximum the composer showed.

• If your Pro balance cannot cover that maximum, it is refused before anything runs.
• A research task that fails is not charged.
• A chat's spending cap applies to it the same way it applies to any reply.

For bigger questions there is also Team mode, which splits the research between several models working at once.

The Tasks pane

The checklist button at the top of the chat opens its Tasks pane, which lists the chat's research as an outline, newest first: the questions it planned to answer, each search, the sources it read with their site icons, and the report it wrote, with the step it is on marked while it runs. Tap a step to jump to its place in the chat. The outline stays with the chat after the research finishes.