Skip to content

PandasAI Guide

PandasAI lets you ask questions about a DataFrame in plain English and get back an answer plus the pandas code that produced it. It does this by sending a description of your data (column names and types — not your rows) to a large language model, which writes pandas code that PandasAI then runs.

Why it matters in this course

It is the fastest way to prototype an analysis ("average tip by day of week?") and the clearest illustration of the CLO-3 lesson: the model generates code — you are responsible for verifying it. It is a drafting tool, never a substitute for knowing pandas.

Setup (once)

pip install pandasai
# And have a local model available (keeps your data on your machine):
ollama pull llama3.2     # see the LLM & Ollama Guide

Your first query

import seaborn as sns
from pandasai import Agent
from pandasai.llm import OllamaLLM

tips = sns.load_dataset("tips")

# Point PandasAI at your LOCAL model — data stays on your machine
agent = Agent(tips, config={"llm": OllamaLLM(model="llama3.2")})
answer = agent.chat("What is the average tip by day of the week?")
print("ANSWER:", answer)

# Audit step: look at the code the model generated
print("GENERATED CODE:", getattr(agent, "last_code_generated", None))
No model installed? You can still see how it works

Run Ollama first (ollama pull llama3.2). If PandasAI or Ollama is not available, the cell should print a friendly message instead of crashing — guard your code with try/except exactly like the course notebooks do (see Notebook 14).

The verification protocol (non-negotiable)

Every answer gets three checks before you may use it:

  1. Read the generated code — does it answer the question you actually asked?
  2. Verify one number yourself with plain pandas:
import pandas as pd
expected = tips.groupby("day")["tip"].mean()
print(expected)                 # compare with what PandasAI printed
  1. Log it — in your README/report: tool + version, the prompt, the generated code, and your verification line.

Common errors

Error Cause / fix
No module named pandasai pip install pandasai (reinstall after a venv change)
connection refused on port 11434 Ollama is not running — start it, then ollama pull llama3.2
Answer differs from your hand-computed number The model wrote reasonable-looking but wrong code — that is why step 2 exists; rerun with a clearer prompt
"Code generated, but not executed" style errors Ask simpler questions; long multi-step requests trip up small local models
Your data "leaked" You pointed PandasAI at a cloud model by default — always set an Ollama LLM for non-public data

Limitations (be honest about these)

  • The model sees the schema, not the rows — it cannot reason about data quality you haven't told it about.
  • Small local models make mistakes on multi-step aggregations and date logic.
  • PandasAI's API changes between versions (e.g. Agent vs SmartDataframe); code from tutorials may need small adjustments.
  • It is not a replacement for learning pandas — you cannot verify what you cannot read.

Responsible AI considerations

  • Privacy: prefer the local-Ollama path for any non-public data; never paste sensitive data into a cloud LLM.
  • Disclosure: record in your README that PandasAI assisted, with tool version and prompt.
  • Don't delegate judgment: the model suggests an aggregation; you decide whether it answers the actual question.

Go deeper