Generative UI that tells you when it is guessing.
Point it at a JSON file. It picks a component, picks the fields, and renders a self-contained HTML page. Every one of those decisions is a typed Jev question over a fixed registry, so the model never writes markup and structurally cannot ask for a component or a field that does not exist.
The part that is different: when Jev is not sure, jevui does not pick. It renders both candidates side by side and says so.
$ jevui examples/revenue.json -o out.html
line_chart p=1.00 c=1.00
$ jevui examples/two-points.json -o out.html
bar_chart p=0.47 c=0.34 ambiguous, also rendering line_chart (p=0.46)Every generative-UI demo picks something. That is fine when the answer is obvious and quietly wrong when it is not, because a confident-looking render is indistinguishable from a coin flip.
Jev returns a confidence alongside the answer, and on this task it separates cleanly. Measured live against the registry below:
| Payload | Choice | Confidence |
|---|---|---|
| 12 months of revenue | line_chart |
1.00 |
| User records with roles | table |
1.00 |
| Three headline numbers | stat_cards |
1.00 |
| 20 categories | bar_chart |
0.63 |
| Time series with five fields | table 0.50 / line_chart 0.48 |
0.38 |
| A two-point series | bar_chart 0.47 / line_chart 0.46 |
0.34 |
The two it is unsure about are the two a designer would argue about. Two points is not enough for a line, and a time series with five fields per row is genuinely a table-or-chart question. So the uncertainty is not noise, and it is worth showing.
curl -fsSL https://raw.githubusercontent.com/anup-a/jevui/main/install.sh | shOr directly:
uv tool install git+https://github.com/anup-a/jevuiJEVUI_API_KEY=... # or JEVGREP_API_KEY, or TYPESAFE_API_KEYjevui [OPTIONS] FILE
| Flag | Meaning |
|---|---|
-o, --out PATH |
Write the HTML here (default jevui.html) |
--title TEXT |
Page heading |
--min-confidence F |
Below this, render both candidates (default 0.7) |
--explain |
Print the decisions and their probabilities |
--json |
Print the decisions as JSON instead of rendering |
Exit codes: 0 rendered, 2 error.
Jev answers a choice question over exactly these keys, so adding a component here is the only way for one to appear.
| Key | Used for |
|---|---|
line_chart |
A value changing over an ordered time axis |
bar_chart |
Comparing a numeric value across a few named categories |
table |
Rows of records with several fields each |
stat_cards |
A handful of headline numbers |
list |
A sequence of short text items |
For charts it then asks which field is the label and which is the value, again as a choice over the field names actually present. A hallucinated column is not representable.
Not your whole file. Row count, field names, which fields are numeric, and the first five rows. That keeps a 10,000-row file as fast and as cheap as a 10-row one.
All of them share the same instinct: when the model is not sure, say so rather than guess.
MIT.