What is the Ask bar?

A box that turns a plain-English question into a study — but it shows you the method it intends to use before it computes a single number, and it needs an AI key.

The ask bar sits under the desk header and again at the top of every dataset report. Without an API key it renders a single line pointing at Settings → AI and nothing else; templates, the builder and every existing study work exactly as before without one.

With a key, the flow has three visible beats, and the middle one is the point:

• You ask. "After a green candle closing in the top quartile, how often does ES take the high?"
• The engine shows its planBefore any result. Which dataset it will use, which new columns it intends to create and the expression behind each, which outcomes it will measure, and the method it will use. You approve that plan with Run study, or Cancel it.
• You get a candidate study, tagged You asked, which you then keep or discard like any other.

Showing the plan first is deliberate. A number arriving straight out of a question is unauditable — you cannot tell whether it answered what you meant. The plan trace is where you find out that "green candle" got compiled into something you did not intend, while it is still free to fix.

Two things it does rather than guess:

• It asks back. If the question is genuinely ambiguous the engine asks one clarifying question instead of picking for you. One is the cap; a second round means the two of you are talking past each other, and it says so and asks you to rephrase.
• It lets you overrule the target. If it picked the wrong dataset, the plan trace has a dropdown to re-aim it, and the plan is re-validated against the new target before Run is offered — so an unrunnable target is refused at the plan rather than halfway through.

Inside a deep research note there is also Not what I meant, which discards the study and records the miss against the original question.