pandasai

Conversational data analysis using natural language queries on pandas DataFrames. Use when you want to ask plain-English questions about data, generate charts, explain transformations, or build explor

SKILL.md

GitHub

PandasAI Skill

Chat with your data using natural language. Ask questions about DataFrames and get insights, visualizations, and explanations powered by LLMs.

When to Use

USE when:

  • Exploring an unfamiliar dataset with open-ended natural language questions
  • Generating quick visualizations from descriptive prompts
  • Explaining complex data transformations to stakeholders
  • Building conversational data exploration interfaces (Streamlit, Jupyter, FastAPI)
  • Rapid prototyping of data analysis workflows

DON'T USE when:

  • Production pipelines requiring deterministic, version-controlled outputs
  • Processing highly sensitive PII without anonymization (use privacy mode)
  • Performance-critical paths with very large DataFrames (>100K rows)
  • Simple queries that direct pandas operations would handle faster

Install

uv add pandasai                          # core
uv add pandasai openai                   # + OpenAI backend
uv add pandasai matplotlib seaborn plotly # + visualization

export OPENAI_API_KEY="sk-..."
# or ANTHROPIC_API_KEY / GOOGLE_API_KEY for other backends

Quick Start

import pandas as pd
from pandasai import SmartDataframe
from pandasai.llm import OpenAI

df = pd.read_csv("data.csv")
smart_df = SmartDataframe(df, config={"llm": OpenAI(model="gpt-4", temperature=0)})

result = smart_df.chat("What is the total revenue by region?")
smart_df.chat("Plot a bar chart of monthly sales")   # saves chart to ./charts/
print(smart_df.last_code_generated)                  # inspect generated code

Core Capabilities

Capability API Notes
Natural language query smart_df.chat(question) Returns value, DataFrame, or chart
Multi-table queries SmartDatalake([df1, df2]).chat(q) Assign df.name before passing
Inspect code smart_df.last_code_generated Verify correctness
Enable caching config={"enable_cache": True} Avoids repeat LLM calls
Privacy mode config={"enforce_privacy": True} Anonymize sensitive columns first
Chart saving config={"save_charts": True, "save_charts_path": "./charts"} Auto-saves matplotlib figures

Supported LLM Backends

Provider Import class Env var
OpenAI from pandasai.llm import OpenAI OPENAI_API_KEY
Anthropic from pandasai.llm import Anthropic ANTHROPIC_API_KEY
Google Gemini from pandasai.llm import GoogleGemini GOOGLE_API_KEY
Azure OpenAI from pandasai.llm import AzureOpenAI AZURE_OPENAI_API_KEY
Local (Ollama) from pandasai.llm import LocalLLM —

Switch backends by passing a different llm= object to config. For fallback logic, see MultiBackendAnalyzer in references/api-reference.md.

Privacy Pattern (Quick)

import hashlib

safe_df = df.copy()
for col in ["name", "email", "ssn"]:
    safe_df[col] = safe_df[col].apply(lambda x: hashlib.md5(str(x).encode()).hexdigest()[:8])

smart_df = SmartDataframe(safe_df, config={"llm": llm, "enforce_privacy": True, "enable_cache": False})

See references/api-reference.md for full PrivacyAwareAnalyzer with audit logging.

Key Gotchas

  • Determinism: Set temperature=0 for reproducible results.
  • Large DataFrames: Sample to ≤10K rows before wrapping to avoid context overflow.
  • Cache invalidation: Cache keys are per-question string; data changes don't auto-invalidate.
  • Chart display: In headless environments set matplotlib backend to Agg before import.
  • Multi-table: Assign df.name = "table_name" before passing list to SmartDatalake.

References

  • Full code examples (all 6 capabilities): references/api-reference.md
  • Integration patterns (Streamlit, FastAPI, Jupyter): references/api-reference.md
  • Best practices, caching, cost management: references/api-reference.md
  • Troubleshooting table: references/api-reference.md
  • Upstream docs: https://docs.pandas-ai.com/
  • GitHub: https://github.com/gventuri/pandas-ai

Version History

  • 1.0.0 (2026-01-17): Initial release — NL queries, chart generation, multi-backend, privacy modes