Marvin is a lightweight AI toolkit for building natural language interfaces that are reliable, scalable, and easy to trust.

Each of Marvin's tools is simple and self-documenting, using AI to solve common but complex challenges like entity extraction, classification, and generating synthetic data. Each tool is independent and incrementally adoptable, so you can use them on their own or in combination with any other library. Marvin is also multi-modal, supporting both image and audio generation as well using images as inputs for extraction and classification.

Marvin is for developers who care more about using AI than building AI, and we are focused on creating an exceptional developer experience. Marvin users should feel empowered to bring tightly-scoped "AI magic" into any traditional software project with just a few extra lines of code.

Marvin aims to merge the best practices for building dependable, observable software with the best practices for building with generative AI into a single, easy-to-use library. It's a serious tool, but we hope you have fun with it.

Marvin is open-source, free to use, and made with 💙 by the team at Prefect.

Install the latest version with pip:

pip install marvin -U

To verify your installation, run marvin version in your terminal.

Marvin consists of a variety of useful tools, all designed to be used independently. Each one represents a common LLM use case, and packages that power into a simple, self-documenting interface.

🦾 Write custom AI-powered functions without source code

🏷️ Classify text into categories

🔍 Extract structured entities from text

🪄 Transform text into structured data

✨ Generate synthetic data from a schema

🖼️ Create images from text or functions

📝 Describe images with natural language

🏷️ Classify images into categories

🔍 Extract structured entities from images

🪄 Transform images into structured data

🎙️ Generate speech from text or functions

🤖 Chat with assistants and use custom tools

🧭 Build applications that manage persistent state

Here's a whirlwind tour of a few of Marvin's main features. For more information, check the docs!

Marvin can classify text using a set of labels:

import marvin

marvin.classify(
    "Marvin is so easy to use!",
    labels=["positive", "negative"],
)

#  "positive"

Learn more about classification here.

Marvin can extract structured entities from text:

import pydantic


class Location(pydantic.BaseModel):
    city: str
    state: str


marvin.extract("I moved from NY to CHI", target=Location)

# [
#     Location(city="New York", state="New York"),
#     Location(city="Chcago", state="Illinois")
# ]

Almost all Marvin functions can be given instructions for more control. Here we extract only monetary values:

marvin.extract(
    "I paid $10 for 3 tacos and got a dollar and 25 cents back.",
    target=float,
    instructions="Only extract money"
)

#  [10.0, 1.25]

Learn more about entity extraction here.

Marvin can generate synthetic data for you, following instructions and an optional schema:

class Location(pydantic.BaseModel):
    city: str
    state: str


marvin.generate(
    n=4,
    target=Location,
    instructions="cities in the United States named after presidents"
)

# [
#     Location(city='Washington', state='District of Columbia'),
#     Location(city='Jackson', state='Mississippi'),
#     Location(city='Cleveland', state='Ohio'),
#     Location(city='Lincoln', state='Nebraska'),
# ]

Learn more about data generation here.

Marvin can cast arbitrary text to any Python type:

marvin.cast("one two three", list[int])

#  [1, 2, 3]

This is useful for standardizing text inputs or matching natural language to a schema:

class Location(pydantic.BaseModel):
    city: str
    state: str


marvin.cast("The Big Apple", Location)

# Location(city="New York", state="New York")

For a class-based approach, Marvin's @model decorator can be applied to any Pydantic model to let it be instantiated from text:

@marvin.model
class Location(pydantic.BaseModel):
    city: str
    state: str


Location("The Big Apple")

# Location(city="New York", state="New York")

Learn more about casting to types here.

Marvin functions let you combine any inputs, instructions, and output types to create custom AI-powered behaviors... without source code. These functions can can go well beyond the capabilities of extract or classify, and are ideal for complex natural language processing or mapping combinations of inputs to outputs.

@marvin.fn
def sentiment(text: str) -> float:
    """
    Returns a sentiment score for `text`
    between -1 (negative) and 1 (positive).
    """

sentiment("I love working with Marvin!") # 0.8
sentiment("These examples could use some work...") # -0.2

Marvin functions look exactly like regular Python functions, except that you don't have to write any source code. When these functions are called, an AI interprets their description and inputs and generates the output.

Note that Marvin does NOT work by generating or executing source code, which would be unsafe for most use cases. Instead, it uses the LLM itself as a "runtime" to predict function outputs. That's actually the source of its power: Marvin functions can handle complex use cases that would be difficult or impossible to express as code.

You can learn more about functions here.

Marvin can paint images from text:

marvin.paint("a simple cup of coffee, still warm")

Learn more about image generation here.

In addition to text, Marvin has beta support for captioning, classifying, transforming, and extracting entities from images using the GPT-4 vision model:

marvin.beta.classify(
    marvin.Image("docs/images/coffee.png"),
    labels=["drink", "food"],
)

# "drink"

💡 Feature idea? share it in the #development channel in our Discord.

🐛 Found a bug? feel free to open an issue.

👷 Feedback? Marvin is under active development, and we'd love to hear it.