Data Analysis for Copilot

Executes Python code generated by Copilot in a Pyodide WebAssembly-based sandbox environment, enabling the LLM to perform computational tasks it cannot execute natively. The extension intercepts code generation requests from the Copilot chat interface, routes them to the Pyodide runtime, captures execution results (stdout, stderr, return values), and streams outputs back to the chat context. This architecture isolates untrusted LLM-generated code from the host system while providing a Python 3.x-compatible execution environment.

Integrates CSV files as first-class context objects within the Copilot chat interface, allowing users to reference files via natural language (e.g., 'Analyze the file #filename.csv') and enabling the LLM to access file metadata, schema, and sample data. The extension parses CSV headers, infers data types, and provides row counts and column statistics to the LLM without requiring manual copy-paste of file contents. This context is maintained across multiple chat turns, allowing iterative refinement of analyses.

When Python code execution fails in the Pyodide sandbox, the extension captures the error (exception type, message, stack trace) and feeds it back to Copilot with context about the original code and input data. The LLM then generates corrected code based on the error, which is automatically re-executed. The mechanism for 'smart' retry is not documented, but likely involves prompt engineering to guide the LLM toward common fixes (type errors, missing imports, logic errors). This creates a feedback loop where the LLM iteratively refines code until execution succeeds.

Copilot generates Python visualization code (using matplotlib, plotly, or other Pyodide-compatible libraries) based on natural language requests like 'create a bar chart of sales by region'. The extension executes this code in the Pyodide sandbox and renders the resulting visualization (image or interactive chart) directly in the chat interface or as an exportable artifact. The visualization code is also made available for export to Jupyter notebooks or standalone Python files, enabling users to refine or reuse visualizations outside the chat context.

Copilot generates Python code for statistical analysis and predictive modeling tasks (e.g., 'build a linear regression model to predict sales') based on natural language requests and CSV data context. The extension executes this code in the Pyodide sandbox, capturing model outputs (coefficients, R-squared, predictions) and making them available in chat. Specific model types and algorithms supported are not documented, but likely include regression, classification, and clustering models from scikit-learn or similar libraries. Generated code is exportable for use in Jupyter notebooks or production pipelines.

Copilot generates Python code for common data cleaning tasks (handling missing values, removing duplicates, type conversion, filtering, aggregation) based on natural language descriptions of desired transformations. The extension executes this code in the Pyodide sandbox on the loaded CSV data, displaying the transformed dataset and making the transformation code available for export. This enables users to clean and prepare data for analysis without writing pandas code manually, with immediate feedback on the results of each transformation.

The extension captures all Python code generated and executed during a chat session (data cleaning, analysis, visualization, modeling) and makes it available for export as a Jupyter notebook (.ipynb) or standalone Python script (.py). This enables users to take exploratory work done in chat and convert it into reproducible, shareable artifacts. The exported code includes markdown cells with explanations (likely generated by Copilot) and preserves the logical flow of the analysis.

The extension registers a right-click context menu option on CSV files in the VS Code file explorer, allowing users to trigger data analysis workflows directly from the file tree without opening the file first. Selecting this option likely opens the Copilot chat interface with the CSV file pre-loaded as context, enabling immediate natural language analysis requests. This integration reduces friction for users who want to analyze files without navigating to the editor first.

When a CSV file is open in the VS Code editor, the extension displays an icon (likely in the editor toolbar or gutter) that users can click to trigger data analysis workflows. Clicking this icon opens the Copilot chat interface with the current CSV file pre-loaded as context, enabling immediate analysis requests. This provides an alternative to right-click context menus and makes the feature more discoverable to users actively editing CSV files.

The extension maintains CSV file context and execution state across multiple chat turns, allowing users to ask follow-up questions and request modifications without re-loading the file or repeating context. The LLM retains awareness of previous analyses, generated code, and execution results, enabling iterative refinement of analyses through natural conversation. This context includes the current state of the data (after transformations), previously generated visualizations, and model outputs.