Airow

AI-powered DataFrame processing made simple

Airow is a Python library that combines pandas or Polars DataFrames with AI models to process structured data at scale. Built on top of pydantic-ai, it provides type-safe, async processing of DataFrames using any AI model.

Features

• 🚀 Async processing with batch support for high performance
• 🔒 Type-safe outputs using Pydantic models
• 📊 Progress tracking with built-in progress bars
• 🔄 Automatic retries with configurable retry logic
• 🤖 Flexible AI models - works with OpenAI, Ollama, Anthropic, and more
• ⚡ Parallel processing within batches for maximum throughput
• 📝 Structured outputs with defined schemas and validation

Installation

# Core library without a DataFrame backend
pip install airow

# pandas only
pip install "airow[pandas]"

# Polars only
pip install "airow[polars]"

# pandas and Polars
pip install "airow[all]"

Pandas example

Install Airow with the pandas backend:

pip install "airow[pandas]"

The examples use Pydantic AI's openai:gpt-5 model string, so configure the corresponding provider credentials before running them.

import asyncio

import pandas as pd

from airow import Airow, OutputColumn


async def main():
    df = pd.DataFrame(
        {
            "description": [
                "Bright citrus flavors with a crisp finish.",
                "Rich dark fruit with firm tannins.",
            ]
        }
    )

    airow = Airow(
        model="openai:gpt-5",
        system_prompt="You are an expert in wine tasting and selection.",
        batch_size=2,
    )

    output_columns = [
        OutputColumn(
            name="summary",
            type=str,
            description="A concise summary of the wine",
        ),
        OutputColumn(
            name="style",
            type=str,
            description="The inferred wine style",
        ),
    ]

    result_df = await airow.run(
        df,
        prompt="Analyze this wine description.",
        input_columns=["description"],
        output_columns=output_columns,
        show_progress=True,
    )

    # result_df is a pandas.DataFrame; df is unchanged.
    print(result_df.head())


if __name__ == "__main__":
    asyncio.run(main())

Airow detects pandas automatically and returns a new pandas.DataFrame.

Polars example

Install Airow with the Polars backend:

pip install "airow[polars]"

import asyncio

import polars as pl

from airow import Airow, OutputColumn


async def main():
    df = pl.DataFrame(
        {
            "description": [
                "Bright citrus flavors with a crisp finish.",
                "Rich dark fruit with firm tannins.",
            ]
        }
    )

    airow = Airow(
        model="openai:gpt-5",
        system_prompt="You are an expert in wine tasting and selection.",
        batch_size=2,
    )

    output_columns = [
        OutputColumn(
            name="summary",
            type=str,
            description="A concise summary of the wine",
        ),
        OutputColumn(
            name="style",
            type=str,
            description="The inferred wine style",
        ),
    ]

    result_df = await airow.run(
        df,
        prompt="Analyze this wine description.",
        input_columns=["description"],
        output_columns=output_columns,
        show_progress=True,
    )

    # result_df is a polars.DataFrame; df is unchanged.
    print(result_df.head())


if __name__ == "__main__":
    asyncio.run(main())

Airow detects eager Polars DataFrames automatically and returns a new polars.DataFrame. LazyFrames are not currently supported.

Custom backends

Custom dataframe implementations can subclass DataFrameBackend and pass an instance explicitly:

from airow import Airow, DataFrameBackend

backend: DataFrameBackend = MyDataFrameBackend()
airow = Airow(
    model="openai:gpt-5",
    system_prompt="You are a data processing assistant.",
    backend=backend,
)