Multi-Table 2D Infinite Canvas Interactive Gadget

Launches an interactive Shiny gadget containing an infinite 2D canvas with mouse-wheel zoom, pan, draggable table cards, interactive column mapping, discrepancy recoding, and LLM-assisted tidyverse code generation.

Usage

sideview(
  ...,
  title = "Sideview Canvas",
  viewer = if (requireNamespace("rstudioapi", quietly = TRUE) &&
    rstudioapi::isAvailable()) {
     shiny::dialogViewer("Sideview Canvas", width =
    1400, height = 900)
 } else {
     shiny::paneViewer(minHeight = 500)
 },
  theme = "auto",
  state = NULL,
  env = parent.frame(),
  run = interactive()
)

Arguments

...: Arbitrary table objects (data.frame, tibble, matrix, table, gt, flextable, htmlwidgets), named or unnamed.
title: Title for the gadget viewer.
viewer: Shiny viewer gadget display mode (e.g. shiny::paneViewer(), shiny::dialogViewer(), shiny::browserViewer()), or a character shorthand ("browser", "dialog", or "pane"). Default is shiny::dialogViewer() (width 1400, height 900) when RStudio is available, otherwise shiny::paneViewer().
theme: Canvas theme: "auto" (default; follows the RStudio editor theme), "light", "dark", an sv_theme() or a bslib::bs_theme().
state: A saved workspace to restore: a sideview_state (see new_sideview_state()) or the path to a file written by the toolbar's Save button or sv_save_state(). Card positions and sizes, zoom and pan, sorts, filters, joins and the pipeline come back; cards are matched to tables by name.
env: Environment where Run in R evaluates the generated code and stores its result (as sv_result unless you name it). Defaults to the environment sideview() is called from, which is your workspace when you call it from the console.
run: Logical; if TRUE, launches the interactive gadget via shiny::runGadget(). If FALSE, returns the shiny.appobj definition without launching (useful for testing or embedding). Defaults to interactive().

Value

If run = TRUE, the generated code (invisibly) when you click Done, or NULL after Cancel or closing the window. If run = FALSE, returns a shiny.appobj with the normalized cards attached.

Big tables (arrow and duckdb)

Arrow tables, datasets and queries and duckdb tables (dplyr::tbl()) go on the canvas as cards that show a 50-row preview. When the duckdb, dbplyr and DBI packages are installed, sorting, filtering, scrolling, the join check, the column inspector and the codebook use the whole table: duckdb runs them and only summaries and row windows come back into R. The table is never collected. Without duckdb, or when the link to a table is lost, the card says "First 50 of N rows" and the checks use the preview.

Generated code stays lazy and ends in one collect(). Known differences from an in-memory card:

"Contains" on a number matches its text. The grid filter drops a trailing .0, but in generated code duckdb writes a double as "130.0" where R writes "130", and float noise or scientific notation can differ. The generated filter is a literal match, stringr::str_detect(col, "<escaped literal>").
POSIXct values at exactly midnight, and -0, can print differently.
Join keys of different types compare as text, and double keys compare exactly.
Arrow time-zone offsets are computed for 1900 to 2100; local times outside that range may be off. duckdb TIMESTAMPTZ columns show in UTC.
Joining arrow with duckdb copies the arrow side into a duckdb temporary table (arrow::to_duckdb() then dplyr::compute()), which costs memory; a data frame joined into duckdb is copied as sideview_<table>.
duckdb 1.4.3 can crash when it scans an arrow factor column that contains NA. sideview decodes such columns itself, but generated code that calls arrow::to_duckdb() on one may hit it: convert the factor to character first.

If Run in R fails on a lazy pipeline, the message suggests moving collect() above the step that needs R.

Single-key shortcuts work when the canvas has the keyboard and you are not typing: F fit all cards, 0 reset zoom, +/- zoom, M show or hide the minimap, D the drawer, T light/dark, / the AI prompt and ? the full list. Ctrl+K finds a card or column, Ctrl+Z/Ctrl+Y undo and redo a pipeline step, and Esc cancels a connection or closes a popover. Double-click a card title, a minimap rectangle or a search result to zoom to that card.

Options

sideview.session_grace: Seconds the gadget waits after its last browser session closes before stopping (default 5), so reloading the page does not end the app.
sideview.llm_backend: "ollama" (default, or the SIDEVIEW_LLM_BACKEND variable) or "openai" for any OpenAI-compatible server.
sideview.llm_model: Model name. Defaults to SIDEVIEW_OLLAMA_MODEL or "llama3.2" for Ollama, and SIDEVIEW_OPENAI_MODEL or "gpt-4o-mini" otherwise. sv_llm_models() lists what is available.
sideview.llm_endpoint: Base URL. Defaults to OLLAMA_HOST or http://localhost:11434, and OPENAI_BASE_URL or https://api.openai.com/v1.
sideview.llm_api_key: API key for OpenAI-compatible servers (default: the OPENAI_API_KEY variable).
sideview.llm_think: TRUE lets Ollama reasoning models "think" before answering. Default FALSE, so they answer promptly.

Only schema summaries (column names, types and simple profiles) are sent to the model, never raw rows.

Examples

if (interactive()) {
  sideview(mtcars, iris)
}