The Shiny issue atlas

AI
Data exploration
6,088 GitHub issues from 9 Shiny-ecosystem repos, embedded and projected with embedding-atlas – select a region to see a Python issue sitting next to the older R issue that asked the same thing, or ask Claude what a selection shares.

The map needs a WebGPU-capable browser, and the Summarize selection button makes a live Claude call (ANTHROPIC_API_KEY required; the button explains itself when the key is unset). There is no Shinylive link: WebGPU and a live model call don’t belong in a browser tab.

6,088 GitHub issues from 9 Shiny-ecosystem repos (R and Python Shiny, bslib, htmltools, shinywidgets, shinychat, shinytest2, shinydashboard, Shiny Server), embedded with all-MiniLM-L6-v2 and projected to 2D with embedding-atlas. Issues that ask for the same thing cluster together – including across languages, so a 2024 py-shiny issue can sit right next to the 2015 R Shiny issue that asked first. The sidebar’s “Asked twice” list surfaces those pairs for whatever region you select, and a Summarize selection button asks Claude, live, what the selected issues have in common and what single change would resolve most of them.

The Shiny issue atlas

Run it

git clone --depth 1 https://github.com/posit-dev/py-shinywidgets
cd py-shinywidgets/website/gallery/showcases/issue-atlas
uv run --with chatlas --with anthropic shiny run app.py

On first run the app downloads issues.parquet (~3 MB) from py-shinywidgets-data; prep_data.py (needs the GitHub CLI) and embed.py rebuild it from the GitHub Issues API:

uv run python prep_data.py
uv run --no-project --python 3.12 --with embedding-atlas --with sentence-transformers \
    --with umap-learn --with pyarrow python embed.py

embed.py downloads the embedding model on first run and caches projections under data/.cache/.

Source: app.py

Try this

  • Drag a box over a dense cluster (shift + drag, or toggle rectangle-select mode in the map’s own toolbar). The sidebar narrows to the selection and lists any “asked twice” pairs inside it.
  • Click Summarize selection: Claude reads the selected issues’ titles and says what they have in common and what single change would resolve the most of them.
  • Clear the selection (the atlas’s own “Clear filters” control) to go back to all 6,088 issues.

Accessors demonstrated

  • reactive_read(atlas.widget, "_predicate") plus atlas.widget.selection() – reaching into the EmbeddingAtlasWidget directly for the __row_index__ positions of the currently-selected rows, read by the stats panel, the “Asked twice” list, and the summary_titles() call that feeds Claude – one accessor driving three different views of the same selection.
  • A reactive.extended_task wired to ui.bind_task_button – the live Claude call runs off the main reactive graph so the rest of the UI (stats, twin list) stays responsive while “Reading issues…” is showing, and the result is keyed by selection_key(sel) so a stale response for a since-changed selection is dropped instead of displayed.

Verification

Driven end-to-end with Playwright in headless Chromium (--enable-unsafe-webgpu --use-angle=metal): a shift-drag selection over a dense cluster narrowed 6,088 points to 1,492, the “Asked twice” list populated with real issue pairs (e.g. py-shiny #1770 next to the older shiny #2922 it echoes), and clicking Summarize selection returned a real, on-topic Claude summary – confirmed in img/after_summary.webp against img/hero.webp.

Source: promoted from website/gallery/showcases/issue-atlas/README.md.