The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external CSV files and DataFrames that are then queried using natural language. This creates a surface for indirect prompt injection where malicious instructions embedded in the data could influence the agent's behavior.
Ingestion points: The skill uses pd.read_csv() to load external data in the pipeline.py example and pd.DataFrame() for local data in multiple examples.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are used when passing data to the chat() method.
Capability inventory: The skill utilizes SmartDataframe.chat() and SmartDatalake.chat() which generate and execute Python code on the host system.
Sanitization: The provided code snippets do not implement custom sanitization or validation of the input data before processing.
[DYNAMIC_EXECUTION]: The skill utilizes the PandasAI library, which operates by generating Python code from natural language queries and executing that code at runtime.
Evidence: Use of SmartDataframe.chat() and SmartDatalake.chat() methods across all demonstrated examples including basic.py, multi-df.py, and cleaning.py.
Context: This is the core functionality of the library; the skill demonstrates the custom_whitelisted_dependencies configuration in config.py which can be used to restrict the execution environment.
[EXTERNAL_DOWNLOADS]: The skill documentation provides instructions to download and install packages from the official Python Package Index (PyPI).
Evidence: The installation section includes pip install pandasai, pip install pandasai[openai], and pip install pandasai[langchain] commands.

pandas-ai