pandas-gpt

Power up your data science workflow with LLMs.

pandas-gpt is a Python library for doing almost anything with a pandas DataFrame using ChatGPT or any other Large Language Model (LLM).

Installation

pip install pandas-gpt[openai]

You may also want to install the optional openai and/or litellm dependencies.

Next, set the OPENAI_API_KEY environment variable to your OpenAI API key, or use the following code snippet:

import openai
openai.api_key = '<API Key>'

If you're looking for a free alternative to the OpenAI API, we encourage using Google Gemini for code completion:

pip install pandas-gpt[litellm]

import pandas_gpt
pandas_gpt.completer = pandas_gpt.LiteLLM('gemini/gemini-1.5-pro', api_key='...')

Examples

Setup and usage examples are available in this Google Colab notebook.

import pandas as pd
import pandas_gpt

df = pd.DataFrame('https://gist.githubusercontent.com/bluecoconut/9ce2135aafb5c6ab2dc1d60ac595646e/raw/c93c3500a1f7fae469cba716f09358cfddea6343/sales_demo_with_pii_and_all_states.csv')

# Data transformation
df = df.ask('drop purchases from Laurenchester, NY')
df = df.ask('add a new Category column with values "cheap", "regular", or "expensive"')

# Queries
weekday = df.ask('which day of the week had the largest number of orders?')
top_10 = df.ask('what are the top 10 most popular products, as a table')

# Plotting
df.ask('plot monthly and hourly sales')
top_10.ask('horizontal bar plot with pastel colors')

# Allow changes to original dataset
df.ask('do something interesting', mutable=True)

# Show source code before running
df.ask('convert prices from USD to GBP', verbose=True)

Custom Language Models

It's possible to use a different language model with the completer config option:

import pandas_gpt

# Global default
pandas_gpt.completer = pandas_gpt.OpenAI('gpt-3.5-turbo')

# Custom completer for a specific request
df.ask('Do something interesting with the data', completer=pandas_gpt.LiteLLM('gemini/gemini-1.5-pro'))

By default, API keys are picked up from environment variables such as OPENAI_API_KEY. It's also possible to specify an API key for a particular call:

df.ask('Do something important with the data', completer=pandas_gpt.OpenAI('gpt-4o', api_key='...'))

OpenAI

pandas_gpt.completer = pandas_gpt.OpenAI('gpt-4o')

LiteLLM

pandas_gpt.completer = pandas_gpt.LiteLLM('gemini/gemini-1.5-pro')

Local (Huggingface)

pandas_gpt.completer = pandas_gpt.LiteLLM('huggingface/meta-llama/Meta-Llama-3.1-8B-Instruct')

OpenRouter

pandas_gpt.completer = pandas_gpt.OpenRouter('anthropic/claude-3.5-sonnet')

Azure

If you want to use the Azure OpenAI Service, you can globally configure the openai and pandas-gpt packages:

import openai
openai.api_type = 'azure'
openai.api_base = '<Endpoint>'
openai.api_version = '<Version>'
openai.api_key = '<API Key>'

import pandas_gpt
pandas_gpt.completer = pandas_gpt.AzureOpenAI(
  model='gpt-3.5-turbo',
  engine='<Engine>',
  deployment_id='<Deployment ID>',
)

Custom

It's also possible to use fully custom code generation:

def my_custom_completer(prompt: str) -> str:
  # Use an LLM or any other method to create a `process()` function that
  # takes a pandas DataFrame as a single argument, does some operations on it,
  # and return a DataFrame.
  return 'def process(df): ...'

pandas_gpt.completer = my_custom_completer

Alternatives

• Sketch: AI-powered data summarization and code suggestions (works without an API key)

Disclaimer

Please note that the limitations of ChatGPT also apply to this library. I would recommend using pandas-gpt in a sandboxed environment such as Google Colab, Kaggle, or GitPod.