Map + chart + chat

AI
Charts
Maps & geospatial
A map, a bar chart, and an LLM chat all read the SAME ipyleaflet bounds() accessor, so panning, asking, and a chat-driven fly_to tool call always agree with each other.

249 real car-share pod locations around Montreal, shown on an ipyleaflet map with a plotly bar chart of what’s currently in view. A chat panel (via shinychat + chatlas) gives an LLM two tools: summarize_view (reads the exact same map.bounds() accessor the chart uses) and fly_to (moves the map to one of five named places, validated against a fixed allow-list). Because the tool and the chart share one accessor, the chat’s answer and the chart can never disagree.

Map + chart + chat

Run it

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

Source: app.py · Open in Shinylive

Try this

  • Pan or zoom the map yourself and watch the bar chart recompute.
  • Ask the chat “what’s visible on the map right now?” – it calls summarize_view, which reads map.bounds() directly, and its counts match the chart exactly.
  • Ask it to “fly to the west island” (or downtown, old port, plateau, overview) – the fly_to tool moves the map in place (map.widget.center = ...), and a follow-up question shows the chart and chat agreeing on the new, much smaller, visible set.

Accessors demonstrated

  • map.bounds() – the map’s current ((south, west), (north, east)) viewport. Read by the chart on every render, and again inside the summarize_view tool via reactive.isolate() (so the tool doesn’t create a reactive dependency on the map).

Read on demand

summarize_view() is a plain function – not a @render.*/@reactive.calc – called from outside Shiny’s reactive graph whenever the LLM decides to invoke the tool. It still gets an accurate map.bounds(): every accessor attaches its browser-side listener as soon as the widget renders and keeps recording interactions whether or not anything has read it yet, so there’s no warm-up step and no risk of missing a gesture made before the first chat turn.

Verification

Server boots clean with --with chatlas --with anthropic (both ephemeral overlays; neither is in the project’s locked venv). Driven with Playwright + shinychat.playwright.ChatController, making real calls to claude-haiku-5-5 via ANTHROPIC_API_KEY:

  • “what’s visible on the map right now?” -> “Used summarize_view. There are 249 car-share pods visible on the map right now. By time of day, that breaks down to: Night: 127, Morning: 38, Afternoon: 33, Evening: 51…” – matches the bar chart exactly.
  • “fly to the west island” -> “Used fly_to. The map is now centered on the West Island at zoom 12…” – map recentered, and the bar chart dropped to a single “evening” bar, confirmed in img/after_fly_to.webp.

Source: promoted from docs/widget-guides/articles/ai-exploration/examples/04_map_chart_chat_combined.py (see NOTES.md).