---

name: pandas-ai description: >- PandasAI lets you ask questions about pandas DataFrames and CSV files in plain English, using an LLM that writes and runs the analysis code. Use when a user wants to chat with a DataFrame, query several tables together, generate charts from a prompt, or run PandasAI 3 with OpenAI or a local Ollama model. license: Apache-2.0 compatibility: 'Python 3.8-3.11 (pandasai 3.0.0 does not install on 3.12+), any OS, an LLM API key or a local Ollama model' metadata: author: terminal-skills version: 1.1.0 category: data-ai repository: https://github.com/sinaptik-ai/pandas-ai tags: - pandas-ai - pandas - llm - natural-language - data-analysis

PandasAI

Overview

PandasAI adds a chat() method to DataFrames. It sends your question and the column layout (plus a few sample rows) to an LLM, which returns Python or SQL; PandasAI runs it and gives back a string, number, DataFrame or chart. The current release is 3.0.0 (October 2025). Version 3 changed the API: the old SmartDataframe(df, config={...}) and from pandasai.llm import OpenAI style still appears in tutorials, but SmartDataframe is deprecated and the pandasai.llm module only exposes the base LLM class. The model comes from the separate pandasai-litellm package. Checked against the package source and docs.pandas-ai.com/v3.

Instructions

Step 1: Install (Python 3.11 or older)

python3.11 -m venv .venv && source .venv/bin/activate
pip install pandasai pandasai-litellm

pandasai 3.0.0 declares python <3.12, so on 3.12 or 3.13 pip refuses to install it; use a 3.11 environment. There are no pandasai[openai] or [langchain] extras in v3.

Step 2: Configure an LLM once, globally

import os
import pandasai as pai
from pandasai_litellm.litellm import LiteLLM

llm = LiteLLM(model="gpt-4.1-mini", api_key=os.environ["OPENAI_API_KEY"])
pai.config.set({"llm": llm, "verbose": False, "max_retries": 3})

LiteLLM model strings select the provider (for example gpt-4.1-mini or ollama/llama3.1; see the LiteLLM provider list). Config keys in v3 are llm, save_logs, verbose, max_retries and file_manager; old keys such as enable_cache, conversational, save_charts and custom_whitelisted_dependencies are gone.

Step 3: Load data and chat

df = pai.read_csv("data/companies.csv")          # or pai.DataFrame({...}), pai.read_excel(...)
response = df.chat("What is the average revenue by region?")
print(response)
print(response.last_code_executed)               # the code the LLM generated
df.follow_up("And only for 2025?")               # continues the same conversation

Step 4: Several DataFrames, charts, local models

employees = pai.read_csv("data/employees.csv")
departments = pai.read_csv("data/departments.csv")
pai.chat("Average salary per department name?", employees, departments)   # joins across frames

chart = df.chat("Plot a bar chart of revenue by region")
chart.save("exports/revenue_by_region.png")      # chart answers are ChartResponse objects

local = LiteLLM(model="ollama/llama3.1", api_base="http://localhost:11434")
pai.config.set({"llm": local})

Examples

Example 1: "Which country has the highest GDP in my CSV?"

import pandasai as pai
countries = pai.DataFrame({
    "country": ["USA", "UK", "France", "Germany", "Japan"],
    "gdp_billion": [25460, 3070, 2780, 4070, 4230],
})
print(countries.chat("Which country has the highest GDP?"))

Result: prints USA, and response.last_code_executed shows the pandas expression behind it.

Example 2: "Compare orders against customers and chart it"

orders = pai.read_csv("data/orders.csv")
customers = pai.read_csv("data/customers.csv")
answer = pai.chat("Show total order value per customer country as a bar chart", orders, customers)
answer.save("exports/order_value_by_country.png")

Result: a PNG of order value per country; check the generated code before trusting the join.

Guidelines

• The LLM sees column names and sample rows, so do not point it at tables with personal or secret data unless that provider is acceptable; use a local Ollama model when data must stay on the machine.
• By default the generated code is executed in your own Python process. Treat prompts and data as untrusted input, and pass a Sandbox implementation via sandbox= for anything exposed to other users.
• Answers can be wrong. Print last_code_executed, spot-check numbers against plain pandas, and do not use it where exactness matters without review.
• df.chat() returns analysis results, not a promise to edit the DataFrame in place; do cleaning in plain pandas when you need a reproducible pipeline.
• Keep API keys in environment variables, never in source. Small local models often fail on multi-table questions.
• For production reporting, ask the model once, then freeze the generated code as normal pandas.