Ask Artifacts

An Ask Artifact is a visualization the Ask agent composes inline inside an answer. The agent selects building blocks as it writes, so an answer can contain none, one, or several, in any order, interleaved with prose.

Every artifact is populated from values the agent’s data tools returned for the grounded source. There are nine types.

Chart

A titled chart with a category or time axis and one or more named series.

Field	Renders as
type	line, area, or bar for series across the x-axis; donut or pie for share-of-whole
title	Heading above the chart
xKey	The row field used for the x-axis; for donut and pie, the category-label field
series	One entry per plotted series: data key, display label, and color
data	The plotted rows

For donut and pie, the first series’ key names the numeric value field and any further series are ignored. Y-axis and tooltip values are grouped below 10,000 and shown compactly (1.2M) above it. X-axis category labels longer than 14 characters truncate with an ellipsis. The chart carries an accessible description naming the chart type and its series labels.

Table

A sortable table.

Field	Renders as
columns	Header cells; each has a key, a label, an optional align (left, center, right), and an optional format
rows	Body rows, keyed by column key
sortBy / sortDir	The initial sort; without them the table renders unsorted
title	Optional heading

format is number, percent, or text. Numeric and percent columns are formatted to at most two decimal places, right-aligned by default, and sorted numerically — including numeric strings. Percent columns append %. Missing, null, or empty cells render as an em dash and sort last.

Clicking a header sorts by that column and toggles direction. At most 15 rows render; any beyond that are summarized below the table.

KPIs

A grid of stat tiles.

Field	Renders as
kpis[].label	The tile caption
kpis[].value	The figure — numbers are thousands-grouped, strings shown verbatim
kpis[].unit	Optional suffix beside the value
kpis[].delta	Optional signed change against the prior period, shown with an up, down, or flat arrow
kpis[].deltaDirectionGoodIs	up (default) or down — which direction counts as an improvement
title	Optional heading above the grid

A delta is colored and labelled by whether it improved: green and “better”, red and “worse”, or grey and “unchanged”. deltaDirectionGoodIs: "down" inverts that, so a falling figure reads as an improvement. The delta’s accessible name states the value, that it is versus the prior period, and the improvement word — colour is never the only signal.

Segments grid

A side-by-side comparison of the same question across groups. Each column carries:

• a title and an n=N badge giving how many responses the group contributed
• mini-stat tiles, each with a label, a value, an optional suffix, and an optional signed delta
• a stacked sentiment bar over positive, neutral, and negative
• theme rows, each with a name, a mention count, and a 0–100 bar fill
• an optional quote with its attribution

Up to three columns render side by side on desktop; more wrap. Each column renders at most eight theme rows, with the remainder summarized below.

Theme bars

Horizontal mention-rate bars, rendered under a Themes heading. Each bar has a label, a 0–100 percentage driving the fill width, and a formatted value string such as 5 of 8.

An optional tone sets the fill colour and adds a matching icon beside the label, with the tone word appended to the bar’s accessible name:

Tone	Fill	Icon meaning
default	Brand	—
ok	Green	Positive
warn	Amber	Warning
bad	Red	Negative

Percentages outside 0–100, and non-numeric ones, clamp to the range. Blank labels fall back to an em dash; long labels truncate with the full text on hover. At most 15 bars render, with a “+N more themes” note for the rest. An optional footer line sits below the bars, typically stating sample size or confidence. An empty payload renders “No data”.

Quote

A single verbatim quote, rendered as a block with a left border coloured by sentiment.

Field	Renders as
text	The quote body; blank falls back to an em dash
who	Attribution; blank falls back to “Anonymous”
role	Optional role shown in the meta row
sentiment	pos, neu, or neg — sets the border colour and a smiley, neutral, or frown icon
tags	Optional pills in the meta row

The sentiment icon carries a text label, so the colour is not the sole signal. At most six tags render; the rest collapse into a “+N more” pill.

Conflict Signal

A Conflict Signal is an amber callout surfacing a divergence the agent found — groups whose views differ on the same theme, presented rather than averaged away.

Field	Renders as
theme	The theme in question, shown as an “On:” caption
summary	A one-line plain-language statement of the disagreement
groups	One row per clustered group: its label and a one-line summary of its position
quote	Optional supporting quote, inlined beneath the groups with its own sentiment icon and label

Clustering happens server-side. An empty groups list leaves the summary standing alone; blank group labels fall back to an em dash.

Sources

The Grounded in panel: a row of numbered chips naming the responses an answer cites.

Each chip has a number, a label, and optionally the id of the response it points at. Chips carrying a response id are clickable; chips without one render but do nothing. Each chip’s accessible name is Source {n}: {label}, so the number is never the only cue. At most 12 chips render, with a “+N more” note beyond that. Long labels truncate with the full label on hover, blank labels fall back to an em dash, and an empty panel renders “No data”.

Trace

A Tool use panel listing the steps the agent took for that turn. Each step has a label and a status:

Status	Shown as
Done	Green check, labelled “Done”
In progress	Spinning indicator, labelled “In progress”
Queued	Ellipsis glyph, labelled “Queued”

Steps may carry a duration. The panel stays expanded while the turn is streaming and collapses once it completes; clicking the header toggles it, and a manual expand is not overridden afterwards. When the answer carries tool-call data, the trace is built from that data in preference to any trace the agent emitted itself.

Inline evidence citations

Within an answer’s prose, [Evidence N] markers render as small numbered chips. Clicking one scrolls to the matching chip in that answer’s Sources panel and briefly highlights it. Resolution is scoped to the answer it appears in, so source 2 in one answer never jumps to source 2 in another.

A marker with no matching Sources entry in the same answer still renders, but is inert and labelled “Source detail unavailable”.

Rendering states

State	Shown as
Payload still streaming	A “Rendering…” placeholder in the artifact’s position
Payload complete but unrenderable	An inert message naming the type, such as “Invalid chart data”

An unrenderable artifact never removes the surrounding prose — the rest of the answer renders normally.

Copy behaviour

The copy action under an answer serialises prose, headings, bullets, and markdown tables. Ask Artifacts are skipped rather than copied as their underlying data.