AI Widgets

How DuckViz turns your data into chart suggestions without ever seeing your row values.

DuckViz's AI watches the shape of your data — column names, types, and a light statistical fingerprint — and suggests charts that fit. It never sees the values themselves. SQL executes locally against DuckDB-WASM in your browser, and the result populates the chart.

What you'll see

After ingesting a file, the right panel shows widget cards as the AI works through your schema:

Title — e.g. "Revenue by Category"
Chart — rendered live against your local DuckDB
Description — what the chart shows
SQL query — expandable for audit or edit
Add (+) — sends the widget to an open dashboard

Widgets appear progressively, not all at once: the data domain comes first, then a chart suggestion, then the finished widget. There's no spinner-and-wait — events show up in your sidebar one at a time.

Why it's fast (and usually free)

Most schemas have been seen before. DuckViz answers from cache or from a curated template catalog whenever it can — only true LLM calls are billed. Cache hits typically respond in under 200 ms; cold-schema runs that hit the LLM finish in 1.5–3 s.

When a recommendation can't be answered fast, DuckViz transparently falls back to a slower multi-step path. You don't pick which path runs — the system does.

How recommendations are built

For each widget, three things get decided:

Data domain — one of 51 domains across 5 categories (Infrastructure, Security, Application, Data & Analytics, Business). The domain steers chart-type selection.
Chart type — picked from DuckViz's registry of 80+ pre-built charts — bar, line, area, treemap, sankey, heatmap, candlestick, and so on.
SQL query — a DuckDB query that produces the right shape for the chosen chart. The query is shown on each widget card and runs entirely in your browser.

Privacy guarantee

The AI receives column names, data types, date-sample strings (so it can pick the right date format), and the detected domain. It never sees actual row values.

Auto-fix on bad queries

If a generated query errors against your data, DuckViz retries automatically up to three times with the error text. Most failures clear on the first retry. If all three fail, the widget is dropped from the timeline so the rest of your dashboard stays clean.

Asking for a specific chart

The chat input in the right panel takes free-form requests:

"Show a monthly revenue trend"
"Heatmap of user activity by hour and day"
"Grouped bar comparing sales across regions"

Ambiguous requests open a clarification drawer first — DuckViz asks one or two questions before spending credits.

Log-aware tailoring

When the detected domain is log-family (e.g. "Network & Firewall", "Application Logs"), chart picks favor timelines, error-level breakdowns, top-N sources, and calendar heatmaps. If DuckViz also knows your log format, format-specific columns flow into the suggestions automatically.