datasium

The powerhouse for intuitive data processing — a visual data-workbench built on Polars (lazy pipelines) and NiceGUI (browser/desktop UI).

uv run datasium

Open http://localhost:8080 (or the native window if run with --native), load a dataset (try sample.csv) via the upload, and inspect its columns, types, and shape in the detail panel. Multiple datasets can be loaded and switched between in the sidebar.

Per active dataset you can:

• Select columns to project (leave empty = all columns).
• Define row filters via a reusable filter builder: pick a column, an operator and a value; combine rows with Match ALL (and) or Match ANY (or). Filters compile to a Polars expression used with lf.filter. Supported operators include:
  • Numeric: equals, not equals, >, >=, <, <=, between, not between, is in, is null, is not null, is NaN, is not NaN.
  • String: equals, not equals, contains, not contains, starts with, ends with, matches regex, is in, length equals / > / <, is null, is not null.
  • Boolean: is true, is false, is null, is not null.
  • Temporal: equals, not equals, >, >=, <, <=, between, not between, is null, is not null.
• Preview the result as either Selected columns × selected rows or All columns × selected rows, with a row/column count. The preview is computed on demand — click Preview to run it.
• Run SQL against the active LazyFrame via the Query tab (table name self, e.g. SELECT * FROM self WHERE age > 20). The query runs lazily and is only collected when you click Run. Executed queries are kept in a history list shown below the input, each with its result (or error) and a re-run button.
• Transform the dataset via the Transform tab:
  • Sort by one or more columns (ascending or descending per column).
  • Rename a column.
  • Computed columns — create a new column from an expression:
    • Arithmetic: A + B, A − B, A × B, A ÷ B, A mod B, A ** B, A // B (floor div) — between two columns or a column and a scalar.
    • Aggregation broadcast: sum, mean, min, max, median, std, count of a column written to every row.
    • Cumulative: running sum, mean, min, max, count.
    • String: UPPERCASE, lowercase, strip whitespace, string length, Title Case, reverse.
    • Rank / index: dense rank by a column, 0-based row index.
    • Conditional: when col > / < / == X then A else B, or when col is null then A else B.
  • Group-by aggregation — group rows by one or more columns and aggregate (mean, sum, min, max, median, std, count, first, last, n unique). The result is saved as a new parallel dataset in the registry.
  • Sort, rename, and computed columns modify the active dataset in place; group-by creates a new dataset you can switch to in the sidebar.
• Plot the active dataset with Plotly: choose a plot type (scatter, line, bar, histogram, box, violin), X / Y / color columns, and (for bar) an aggregation statistic (mean / sum / min / max / median / count). Plot from the entire dataset or just the current selection (the rows that pass the Select-tab filters); click Plot to render the figure.

Datasets are held as Polars LazyFrames end to end, so nothing is materialised until you ask for it. Expensive work — rendering a result table, running a statistic, building a plot, or executing a SQL query — only happens when you click the corresponding button (Preview, Calculate, Plot, Run). Editing filters or column selections just stages the change; it never triggers a recomputation on its own.

uv run pytest