Conversational BI

Conversational BI turns a question into a chart. You ask an assistant to visualize something — "show me customer count by region as a bar chart" — and it puts a real chart card on your dashboard, built from a live query against your data. The numbers on the card are the actual result of that query, not a picture the model drew from memory.

Overview

When you ask to see data as a chart, the assistant first works out the query it needs and checks the result, then hands it off to build the card. The chart appears on a canvas next to the chat, and you can keep refining it by talking to the assistant. Everything on the card traces back to a query you can inspect.

Why the numbers are real

Charts only ever come from data that was actually queried. This is what makes a Conversational BI card something you can trust in a report.

The system runs your query for real, and only draws the chart if it succeeds. The assistant can't hand you a chart of made-up numbers — every value on the card is a real result from your connected data.

The query is preserved exactly. What the assistant asks for is what runs and what's recorded with the chart, character for character — including non-English column names. Nothing is trimmed or altered on the way to the card.
Same rules as any query. Charts read your data through the same read-only, per-tenant, billed path as every other question. There's no back door; a chart can never write to or overload your data.

When it's available

Conversational BI works whenever your assistant is connected to at least one database it can query. If a model has no database attached, there's nothing to chart, and the option simply doesn't appear. When several databases are connected, the chart is built from the first one the assistant queries.

Your databases and their synced copies may speak slightly different SQL dialects. The assistant is told which one applies before it writes a query, so charts come out correct whether they read from a live source or a synced copy.

How a Chart is Generated

You ask for a chart

In plain language: "chart revenue by month" or "add this to my BI dashboard as a pie."

The assistant writes and checks a query

It explores your tables, writes a query, and runs it to confirm the result and the column names before drawing anything.

The chart is built from that exact query

The system runs the query once more to be sure, and only on success does the card appear on your canvas — carrying the real rows behind it.

If something's wrong, it fixes it

If the query can't run, no chart is drawn; the assistant sees what went wrong and corrects it, rather than showing you a broken or empty card.

Chart Types

Pick the shape that fits the question:

Type	Use for
Line	Trends over time
Bar	Comparing categories
Area	Cumulative or stacked amounts over an axis
Pie	Part-to-whole composition
Stat	A single headline number (KPI)

Very large results are capped so a card stays readable — the chart shows the leading rows and marks itself as trimmed, while still reporting the true total row count.

Chart cards live on a canvas beside the chat, and they stay there between sessions. To change a card — "add product to this chart" or "make it a line chart" — just ask. The assistant edits the existing card in place instead of opening a duplicate, so your dashboard stays tidy as you refine it.

Safety & Failure Modes

Situation	What happens
A query can't run (syntax, permission, timeout)	No chart is drawn; the assistant sees the error and can try again
A request that isn't a plain read	Refused before anything runs — charts are read-only
The assistant gets stuck retrying	It's stopped and wraps up rather than looping
A result is larger than the cap	The chart is trimmed and marked as such; the true total is still reported

A chart is only ever created from a query that ran successfully against your data. If the query fails, you get no chart — never a chart with invented numbers.