"""Example: AI chat that plots matplotlib figures in a sandboxed Docker container. The chat agent is wired to a ``PlotPanel`` (on the right of the chat) via five ``BaseTool`` wrappers: ``plot_by_code``, ``run_code``, ``load_data_from_csv``, ``attach_current_plot_to_osw_page``, and ``document_current_evaluation``. When the required OSW environment variables are set, an additional eight OSW-connector tools (``get_page_html``, ``download_osl_file``, ``sparql_search``, ``find_out_everything_about``, ``get_topic_taxonomy``, ``get_instances``, ``get_file_header``, ``get_website_html``) are registered; otherwise they are silently omitted and the example still runs with the three core plotting tools. Environment variables --------------------- LLM — one of the following provider set-ups (see ``src/panelini/panels/ai/default_config.yml``): ANTHROPIC_API_KEY OR AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_API_VERSION OSW (optional, enables 8 extra tools + OSW upload of the current plot): OSW_DOMAIN OSW_USER OSW_PASSWORD BLAZEGRAPH_ENDPOINT BLAZEGRAPH_USER BLAZEGRAPH_PASSWORD The three OSW auth variables (``OSW_DOMAIN``, ``OSW_USER``, ``OSW_PASSWORD``) are *required together* when any OSW tool is invoked. Credentials stay in memory — no ``accounts.pwd.yaml`` is written, and the ``input()`` / ``getpass`` prompt that ``osw.express`` normally shows on a fresh machine is bypassed. Missing any of the three produces a ``RuntimeError`` with the missing names. Panelini (optional): PANELINI_AI_CONFIG_PATH # override the default AI config file Install hints ------------- With uv (from a checkout of this repo): uv sync --extra ai --extra ai-llm-sandbox # or, with OSW connector: uv sync --extra ai --extra ai-llm-sandbox --extra ai-osw # Run the example: uv run python examples/panels/ai/plot_by_code.py With uv (as a dependency in another project): uv add 'panelini[ai,ai-llm-sandbox]' # or, with OSW connector: uv add 'panelini[ai,ai-llm-sandbox,ai-osw]' With pip: pip install 'panelini[ai,ai-llm-sandbox]' # or, with OSW connector: pip install 'panelini[ai,ai-llm-sandbox,ai-osw]' Runtime requirements -------------------- * Docker daemon running (``plot_by_code`` / ``run_code`` spin up ``python:3.12-slim`` sandbox containers). * A ``.env`` file next to this script (or exported env vars) with the LLM credentials; ``python-dotenv`` is loaded below. OSW connector behaviour ----------------------- The OSW tools are registered only when *all six* OSW env vars are set (three auth vars + three Blazegraph vars). Otherwise they are silently omitted from the tool list, and the example runs with ``plot_by_code``, ``run_code``, and ``load_data_from_csv`` only — plus the two OSW-bound plot tools, which themselves degrade gracefully (returning an error string instead of raising) when ``osw`` is unavailable at runtime. Right sidebar — plot model override ----------------------------------- The right sidebar exposes the last plot's Python code, a model picker (default: *Claude Sonnet 4.6*), and a *Regenerate plot* button. Clicking *Regenerate* asks the selected model to rewrite the current plot script according to a free-text "what should change?" instruction and re-runs it through the same sandbox. The main chat agent is unchanged and still drives the primary plotting flow. """ from __future__ import annotations import panel as pn from dotenv import load_dotenv from panelini import Panelini from panelini.panels.ai import AiChat from panelini.panels.ai.plot import ( PlotPanel, build_plot_context_sidebar, make_osw_tools, make_plot_tools, ) load_dotenv() SYSTEM_MESSAGE = ( "You are a helpful assistant with access to tools. " "ALWAYS call tools directly to fulfill the user's request — never describe " "what a tool call would look like or output JSON of a hypothetical call. " "If you have the information needed to call a tool, call it immediately.\n\n" "When the user asks to download files and run code:\n" "1. Download the files using the download tool (if available).\n" "2. Pass the returned file paths directly to plot_by_code or run_code via " "the file_paths parameter.\n" "3. Include all required pip packages in the libraries list.\n\n" "plot_by_code / run_code sandbox rules:\n" "* plot_by_code MUST save its figure to '/sandbox/output.png'.\n" "* Files passed via file_paths are available at '/sandbox/'.\n\n" "MICRESS / micpy usage (pip package: micress-micpy):\n" "Geometry files (.geof) are automatically paired with binary files in the " "sandbox — do NOT set geometry manually. Correct API:\n" " from micpy import bin\n" " with bin.File('/sandbox/FILENAME.conc1') as f:\n" " field = f.read(-1) # NOT read_field()\n" " fig, ax, cbar = field.plot() # NOT bin.plot()\n" " fig.savefig('/sandbox/output.png')\n\n" "If the user wants to create something in OSW, first try to find the " "category page for the given topic/keyword, then create the instance." ) plot_panel = PlotPanel() tools = [*make_plot_tools(plot_panel), *make_osw_tools()] chat = AiChat(system_message=SYSTEM_MESSAGE, tools=tools) app = Panelini( title="AI + Plot by Code", sidebar_enabled=True, sidebar_right_enabled=True, sidebar_right_visible=True, ) app.sidebar_set(objects=chat.sidebar_objects) app.sidebar_right_set(objects=build_plot_context_sidebar(plot_panel)) chat_card = pn.Card( chat.chat_interface, title="Chat", collapsible=False, sizing_mode="stretch_both", min_height=600, styles={"padding": "15px", "margin-right": "10px"}, ) plot_card = pn.Card( plot_panel.plot_panel, title="Plot", collapsible=False, sizing_mode="stretch_both", min_height=600, styles={"padding": "15px", "margin-left": "10px"}, ) app.main_set(objects=[pn.Row(chat_card, plot_card, sizing_mode="stretch_both")]) if __name__ == "__main__": pn.serve(app.servable(), title="AI + Plot by Code", port=5008)