How an answer is made

The agent beside the map reads the profile of each processed file (its fields, their types and values) and then works in steps. In each step it picks one of its tools, runs it against your data, and reads the result before deciding what to do next. The numbers in its answer are the results of those steps, never its own estimate.

It doesn't write SQL. It writes a small filter, like "class is hotspot and power loss above 50", which we check against the file's real fields and turn into a safe query. When it names a field that doesn't exist, it gets back "Did you mean power_loss_w?" and corrects itself; you'll sometimes see that step in the trace.

What it can do

Step	Used for
Read the file profile	Which fields exist, their values, what each class means.
Count	"How many tables have hotspots?", with any filter.
Calculate	Sum, average, minimum, maximum, median or 90th percentile of a field, overall or per group.
Search	Finding features by meaning, when the words differ from the data ("cracks" for "micro-fracture").
Look up features	Reading specific features and their values.
Check distances	Features within a distance of others, or touching them.
Draw on the map	Highlighting the features an answer is about, and moving the map to them.
Imagery	Showing an index layer and measuring it over an area (see Imagery and indices).
Actions	Exporting, sharing a view, making a task or a follow-up, when you ask. Anything that leaves your workspace takes your click.

Checking an answer

The steps under each answer show every tool call, what it was asked and what it returned. Open one to see the exact filter and the result.
Sources list the features the answer rests on. Each has a button that zooms the map to it.
The highlight on the map is the same set of features the numbers came from, not a separate search.

Questions that work well

Name the measure and the group: "average power loss per string", "total area of soiling by block".
Say where: "in view" uses the part of the map you're looking at, shape by shape (not just its rectangle).
Build on the last answer: "of those, which are still open?" filters the features you were just shown.
Use your own words. The agent sees every class value in the file, so it can match "hotspots" to whatever your data calls them.

After each answer you get a few suggested next questions. Stop ends an answer after its current step; the steps it already made stay in the trace, and are metered.

Conversations

Each project keeps its conversations, and the Chats page lists yours. A conversation remembers the model it started with. If that model is no longer available, the conversation becomes read-only and Continue in a new chat carries the last answers over to a fresh one.

Limits

Up to six steps per question. A question that needs more gets an answer asking you to narrow it.
Three answers can run at once in one organization; a fourth is asked to try again in a moment.
On Platform Models each answer costs credits for the model calls it made, shown with the answer. On your own key, your provider bills you instead.
An Owner can set a daily AI limit; when it's used up, questions wait until midnight UTC.