Charts

Charts are one of the three visual types the Streya agent produces, alongside cards and tables. They appear in both conversations and dashboards, and in both cases they’re built the same way: you describe the chart in plain language, the agent builds it.

This guide has two halves. The first is organized by the question you’re trying to answer — comparing categories, following a trend, finding outliers — with a live example of each chart shape. The second covers everything you can customize on a chart once it exists: colors, axes, labels, legends, and more. Every example is live: hover over the marks to see the same tooltips you’d get in Streya. (The data behind them is a sample online retailer.)

What makes a good chart request

You don’t need chart-making vocabulary — describe the question and the agent picks a sensible shape. But the requests that land right on the first pass name three things:

The measure — the number being shown: revenue, order count, average margin.
The breakdown — what each bar, point, or line stands for: by month, per product category, one point per customer.
The shape (optional) — say “as a line chart” or “as a donut” if you already have one in mind. If you don’t, skip it.

Bar chart of revenue by product category, highest first.

Everything else — colors, axis formats, legends, tooltips — has good defaults, and all of it can be changed by asking. Precision matters most in dashboards: a conversation lets you correct course turn by turn, but a well-described dashboard chart comes out right the first time.

Choose the right chart

Each section below is a common analytical job, the prompt that asks for it, and the chart it produces.

Compare categories

To rank things — products, regions, teams — ask for a bar chart. Horizontal bars keep long category names readable; say how to sort and how many to show.

You ask “Which product categories drove the most revenue in 2025? Sort them highest first.”

Variations to ask for: “only the top three”, “make the bars vertical”, “group everything under $10,000 into ‘Other’.”

Show a trend over time

To see how a number moves, ask for a line chart and name the time grain — daily, weekly, monthly, by quarter. Here the two years are split into separate lines so the seasonal shape is easy to compare.

You ask “Show monthly sales as a line chart, one line per year.”

Variations to ask for: “show it weekly instead — monthly hides too much”, “just 2025, one line”, “smooth the line out.”

Compare two periods

When the question is “how does this year stack up against last year?”, ask for the periods side by side. This is the same data as the line chart above — bars emphasize the month-by-month gap, lines emphasize the overall shape. Ask for whichever reads better, or ask for both and keep one.

You ask “Show me monthly sales from 2025 compared to 2024”

Variations to ask for: “show the year-over-year change instead, one bar per month”, “make it percentage growth rather than dollars.”

Show a running total

To see progress accumulate — revenue against a yearly goal, signups since launch — ask for a cumulative or running total view. An area chart makes the magnitude visible.

You ask “Show cumulative revenue through 2025 as an area chart.”

Variations to ask for: “add a horizontal line at our $120k target”, “one line per year so I can compare the pace.”

Break a total into parts

To see what a total is made of and how that mix changes over time, ask for a stacked bar chart: one bar per period, split by the category.

You ask “Break monthly 2025 revenue down by sales channel, stacked.”

Variations to ask for: “unstack them — channels side by side”, “show each month as percentages, so the mix is comparable.”

Show shares of a whole

For a single snapshot of how a total divides up, ask for a donut or pie. These work best with a handful of categories — beyond five or six, a sorted bar chart is easier to read.

You ask “What share of 2025 revenue came from each channel? Show it as a donut.”

Variations to ask for: “add the percentage to each label”, “make it a regular pie”, “show this as a bar chart instead.”

Spot relationships and outliers

To see how two numbers relate — and who breaks the pattern — ask for a scatter plot with one point per entity: customer, product, store. In this one, the lone point far to the right is a customer with 168 orders; that’s the kind of thing a scatter surfaces instantly and a table buries.

You ask “Plot each customer's number of orders against their total revenue, colored by segment.”

Variations to ask for: “size each point by total revenue”, “label the outliers”, “only the loyal clients.”

See patterns across two dimensions

When the question has two “by”s — by weekday and by hour, by region and by product — ask for a heatmap: one axis per dimension, the number as color.

You ask “When do orders come in? Weekday on the y-axis, hour of day on the x-axis, number of orders as color.”

Variations to ask for: “print the count in each cell”, “only weekdays, only business hours.”

Customize every part of it

A chart is never final. The agent edits the existing chart and re-renders it — you never start over — so ask for changes the way you’d mark up a printout: small, single-purpose, in plain language. If a chart comes back unreadable, just describe what’s wrong (“too busy”, “can’t read the labels”); you don’t need to know the fix.

Everything below works on any chart. The quotes are examples of how to ask.

Colors

One color or many — “use a single color for all the bars” / “color them by region.”
Specific colors — “2024 in grey, 2025 in gold.”
Highlight one thing — “highlight November, mute everything else.”
Color by rule — “red when negative, green when positive”, “orange for anything below target.”
Numeric color scales — “light-to-dark blue”, “a diverging scale around zero”, “reverse the scale.”
Transparency — “make the fill more transparent.”

Bars, lines, points, and areas

Bars — rounded or square corners; thicker or thinner; more or less gap between them; horizontal or vertical.
Lines — smooth curves instead of straight segments; stepped; dashed or dotted; thicker or thinner; “add a dot at each data point.”
Areas — more or less fill; with or without an outline.
Points — bigger or smaller; a different shape per series (circles, squares, triangles); filled or hollow.
Stacking — stacked, side by side, or “100% stacked so each bar shows shares.”
Missing data — “leave a gap where there’s no data” / “connect through it.”

Axes and scales

Axis titles — rename them or remove them.
Where the axis starts — “start the y-axis at zero” / “zoom into the 80–100 range.”
Log scale — when values span orders of magnitude.
Dual axis — “revenue as bars on the left axis, order count as a line on the right.”
Flip — “swap the axes.”
Gridlines and ticks — show or hide gridlines; fewer ticks; rotate or hide the labels.

Number and date formats

These apply anywhere a value appears — axis, data label, or tooltip:

Numbers — currency, percentages, how many decimals.
Abbreviations — “show $12k and $1.2M instead of full numbers.”
Dates — “just ‘Jan’, not the full date”, “label by quarter”, “use weekday names.”

Sorting, top N, and filtering

Sort order — by value (ascending or descending), alphabetically, or a custom order: “sort by our pipeline stages, not alphabetically.” Months and weekdays come back in calendar order by default.
Top N — “top 8, group the rest into ‘Other’.”
Leave things out — “drop the Unknown category”, “exclude internal accounts.”

Labels and annotations

Value labels — on every bar or point, or selectively: “only label the highest and lowest”, “label just the last point of each line.”
Totals — “put the total above each stacked bar.”
Reference lines — “a dashed line at the $50k target”, “a line at the average.”
Reference bands — “shade the acceptable range.”
Callouts — “annotate August — that’s when we launched.”

Legend

Show or hide it.
Position — top, bottom, left, or right; laid out horizontally or vertically.
Rename its title, or remove the title.

Titles, text, and tooltips

Title and subtitle — add, rewrite, or remove; align left or center.
Text styling — “make the axis labels darker and a bit bigger”, “shrink the legend text.”
Tooltips — choose which fields show on hover and how they’re formatted, or turn tooltips off.

When a chart isn’t the right visual

A single headline number — with or without a comparison (“Q2 revenue, % vs. Q1”) — is a job for a card.
Detail and lookup — one row per account with several columns — is a job for a table.

Ask for the form you want by name; the agent also picks sensibly on its own.

Export and reuse

Copy as image — copy any chart as a PNG, ready to paste into slides, docs, or chat.
Download CSV — export the data behind any chart.

Under the hood, every chart is a specification written in Vega-Lite, a grammar that can express far more than the single-panel charts above. You never need to read or write it — but if you know it, you can use it:

Borrow its vocabulary in prompts — marks (bar, line, area, point, rect) and encodings (x, y, color, size, shape) — for surgical precision.
Ask for layered charts (e.g. a line with a shaded target band behind it) or faceted small multiples (the same chart repeated per region). If Vega-Lite can express it, the agent can build it.

The Vega-Lite example gallery is a good catalog of what’s possible. This guide sticks to single-panel charts because they cover the overwhelming majority of dashboards — reach for the advanced forms when a simple chart genuinely can’t say it.

Limits to know

A chart draws from one query result at a time. The query itself can join many datasets — but if you want two unrelated datasets on one chart, ask the agent to combine them into one analysis first.