Marvin is a lightweight AI engineering framework for building natural language interfaces that are reliable, scalable, and easy to trust.
Sometimes the most challenging part of working with generative AI is remembering that it's not magic; it's software. It's new, it's nondeterministic, and it's incredibly powerful - but still software.
Marvin's goal is to bring the best practices for building dependable, observable software to generative AI. As the team behind Prefect, which does something very similar for data engineers, we've poured years of open-source developer tool experience and lessons into Marvin's design.
Marvin's docs are available at askmarvin.ai, including concepts, tutorials, and an API reference.
pip install marvin
Getting started? Head over to our setup guide.
To ask questions, share ideas, or just chat with like-minded developers, join us on Discord or Twitter!
Marvin's high-level abstractions are familiar Python interfaces that make it easy to leverage AI in your application. These interfaces aim to be simple and self-documenting, adding a touch of AI magic to everyday objects.
๐งฉ AI Models for structuring text into type-safe schemas
๐ท๏ธ AI Classifiers for bulletproof classification and routing
๐ช AI Functions for complex business logic and transformations
๐ค AI Applications for interactive use and persistent state
Marvin's most basic component is the AI Model, a drop-in replacement for Pydantic's BaseModel
. AI Models can be instantiated from any string, making them ideal for structuring data, entity extraction, and synthetic data generation.
You can learn more about AI models here.
from marvin import ai_model
from pydantic import BaseModel, Field
@ai_model
class Location(BaseModel):
city: str
state: str = Field(..., description="The two-letter state abbreviation")
Location("The Big Apple")
# Location(city='New York', state='NY')
AI Classifiers let you build multi-label classifiers with no code and no training data. Given user input, each classifier uses a clever logit bias trick to force an LLM to deductively choose the best option. It's bulletproof, cost-effective, and lets you build classifiers as quickly as you can write your classes.
You can learn more about AI Classifiers here.
from marvin import ai_classifier
from enum import Enum
@ai_classifier
class AppRoute(Enum):
"""Represents distinct routes command bar for a different application"""
USER_PROFILE = "/user-profile"
SEARCH = "/search"
NOTIFICATIONS = "/notifications"
SETTINGS = "/settings"
HELP = "/help"
CHAT = "/chat"
DOCS = "/docs"
PROJECTS = "/projects"
WORKSPACES = "/workspaces"
AppRoute("update my name")
# AppRoute.USER_PROFILE
AI Functions look like regular functions, but have no source code. Instead, an AI uses their description and inputs to generate their outputs, making them ideal for NLP applications like sentiment analysis.
You can learn more about AI Functions here.
from marvin import ai_fn
@ai_fn
def sentiment(text: str) -> float:
"""
Given `text`, returns a number between 1 (positive) and -1 (negative)
indicating its sentiment score.
"""
sentiment("I love working with Marvin!") # 0.8
sentiment("These examples could use some work...") # -0.2
AI Applications permit interactive use cases and are designed to be invoked multiple times. They maintain three forms of state: the application's own state
, the AI's plan
, and a history
of interactions. AI Applications can be used to implement many "classic" LLM use cases, such as chatbots, tool-using agents, developer assistants, and more. In addition, thanks to their persistent state and planning, they can implement applications that don't have a traditional chat UX, such as a ToDo app. Here's an example:
from datetime import datetime
from pydantic import BaseModel, Field
from marvin import AIApplication
# create models to represent the state of our ToDo app
class ToDo(BaseModel):
title: str
description: str = None
due_date: datetime = None
done: bool = False
class ToDoState(BaseModel):
todos: list[ToDo] = []
# create the app with an initial state and description
todo_app = AIApplication(
state=ToDoState(),
description=(
"A simple todo app. Users will provide instructions for creating and updating"
" their todo lists."
),
)
# invoke the application by adding a todo
response = todo_app("I need to go to the store tomorrow at 5pm")
print(f"Response: {response.content}\n")
# Response: Got it! I've added a new task to your to-do list. You need to go to the store tomorrow at 5pm.
print(f"App state: {todo_app.state.json(indent=2)}")
# App state: {
# "todos": [
# {
# "title": "Go to the store",
# "description": "Buy groceries",
# "due_date": "2023-07-12T17:00:00+00:00",
# "done": false
# }
# ]
# }
๐ท๏ธ Build bulletproof and lightning-fast classifiers
๐งฉ Extract structured & type-safe data from unstructured text
๐งช Generate synthetic data for your applications
๐ซก Solve complex deductive and inferential tasks at scale
๐ Scrape web data without custom scrapers
๐ Customize ChatGPT with system prompts and tools
๐ Extract relevant insights from your data
๐งโ๐ป Add a junior developer to your team
๐ฃ๏ธ Quickly add NLP to your app
๐ฑ AI applications with persistent state
๐ต๏ธ Autonomous agents with high-level planning
๐ฌ Text-to-application: generate stateful applications by describing them