v0.19.0

What's Changed

• Publish docs on release or manually by @jackmpcollins in #162
• Bump pytest-cov from 4.1.0 to 5.0.0 by @dependabot in #165
• Add Structured Output docs by @jackmpcollins in #167
• Remove unneeded wrapper functions for openai and litellm completion by @jackmpcollins in #171
• Add AnthropicChatModel by @jackmpcollins in #174

Full Changelog: v0.18.1...v0.19.0

---

🎉 Native support for Anthropic using AnthropicChatModel 🎉

from magentic import prompt
from magentic.chat_model.anthropic_chat_model import AnthropicChatModel


@prompt(
    "Name three fruits. Be sure to return this in correct JSON format matching the schema.",
    model=AnthropicChatModel("claude-3-opus-20240229")
)
def say_fruit() -> list[str]: ...

say_fruit()

v0.19.0a0

Prerelease for testing PR #174

v0.18.1

What's Changed

• Fix: LiteLLM Anthropic async tool calls by @jackmpcollins in #155
• Make dependabot ignore patch version updates by @jackmpcollins in #157
• Add github workflow options to run existing live tests by @jackmpcollins in #159
• Load .env file for tests. Fix flaky return_dict test. by @jackmpcollins in #160
• Remove tool_call.type check to fix LiteLLM Mistral tool calling by @jackmpcollins in #161

Full Changelog: v0.18.0...v0.18.1

v0.18.0

What's Changed

• Add docs and classes for formatting objects in prompts by @jackmpcollins in #126
• Upgrade ruff. Format ... on same line by @jackmpcollins in #134
• poetry update by @jackmpcollins in #133
• Switch from deprecated functions param to tools by @jackmpcollins in #144
• Docs: Fix bytes vision example by @mnicstruwig in #146

New Contributors

• @mnicstruwig made their first contribution in #146

Full Changelog: v0.17.0...v0.18.0

---

Now fully migrated to tool calls 🎉 Which enables using the new Claude-3 models via litellm, as well as parallel function calls. See more at https://magentic.dev/function-calling/

@prompt(
    "Sum {a} and {b}. Also subtract {a} from {b}.",
    functions=[plus, minus],
)
def plus_and_minus(a: int, b: int) -> ParallelFunctionCall[int]: ...


output = plus_and_minus(2, 3)
print(list(output))
# > [FunctionCall(<function plus at 0x106b8f010>, 2, 3), FunctionCall(<function minus at 0x106b8ef80>, 3, 2)]
output()
# (5, 1)

---

Warning

Breaking Change: FunctionResultMessage now accepts FunctionCall as argument instead of a function

FunctionResultMessage now (again) takes a FunctionCall instance as its second argument. Further, when constructing a chat message history, the same FunctionCall instance must be passed to an AssistantMessage and the corresponding FunctionResultMessage. This so that the result can be correctly linked back to the function call that created it.

def plus(a: int, b: int) -> int:
    return a + b

plus_1_2 = FunctionCall(plus, 1, 2)

@chatprompt(
    UserMessage("Use the plus function to add 1 and 2."),
    AssistantMessage(plus_1_2),
    FunctionResultMessage(3, plus_1_2),
)
def do_math() -> str: ...

---

Example of using the new formatting classes to format a prompt. See more at https://magentic.dev/formatting/

from magentic import prompt
from magentic.formatting import NumberedList


@prompt("Continue the list:\n{items}")
def get_next_items(items: NumberedList[str]) -> list[str]: ...


items = NumberedList(["apple", "banana", "cherry"])
print(get_next_items.format(items=items))
# Continue the list:
# 1. apple
# 2. banana
# 3. cherry

v0.18.0a0

Prerelease for testing PR #144

v0.17.0

What's Changed

• Add support for GPT4 vision by @jackmpcollins in #116

Full Changelog: v0.16.0...v0.17.0

---

Add support for GPT4 Vision with new UserImageMessage message type. See https://magentic.dev/vision/

from magentic import chatprompt, OpenaiChatModel, Placeholder, UserMessage
from magentic.vision import UserImageMessage


IMAGE_URL_WOODEN_BOARDWALK = "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"


@chatprompt(
    UserMessage("Describe the following image in one sentence."),
    UserImageMessage(Placeholder(str, "image_url")),
    model=OpenaiChatModel("gpt-4-vision-preview", max_tokens=2000),
)
def describe_image(image_url: str) -> str:
    ...


describe_image(IMAGE_URL_WOODEN_BOARDWALK)
# 'A wooden boardwalk meanders through lush green wetlands under a partly cloudy blue sky.'

v0.16.0

What's Changed

• Enable registration of new FunctionSchemas by @jackmpcollins in #119
• Update pre-commit hooks by @jackmpcollins in #120
• Add docs site by @jackmpcollins in #124
• Add Placeholder class for templating by @jackmpcollins in #118
• poetry update by @jackmpcollins in #125

Full Changelog: v0.15.0...v0.16.0

---

New docs site 🎉 https://magentic.dev

---

Custom types can now be registered to be used as return type annotations with @prompt-decorated functions. For example, after registering a pandas DataFrame the following example works. See the docs page for how to do this: https://magentic.dev/examples/registering_custom_type/

import pandas as pd
from magentic import prompt


@prompt(
    "Create a table listing the ingredients needed to cook {dish}."
    "Include a column for the quantity of each ingredient."
    "Also include a column with alergy information."
)
def list_ingredients(dish: str) -> pd.DataFrame:
    ...


list_ingredients("lasagna")
# <DataFrame of ingredients>

---

The Placeholder class has been added to enable better templating of AssistantMessage and custom Message subclasses. More info in docs https://magentic.dev/chat-prompting/#placeholder . Sample usage:

from magentic import chatprompt, AssistantMessage, Placeholder, UserMessage
from pydantic import BaseModel


class Quote(BaseModel):
    quote: str
    character: str


@chatprompt(
    UserMessage("Tell me a quote from {movie}"),
    AssistantMessage(Placeholder(Quote, "quote")),
    UserMessage("What is a similar quote from the same movie?"),
)
def get_similar_quote(movie: str, quote: Quote) -> Quote:
    ...


get_similar_quote(
    movie="Star Wars",
    quote=Quote(quote="I am your father", character="Darth Vader"),
)
# Quote(quote='The Force will be with you, always.', character='Obi-Wan Kenobi')

v0.15.0

What's Changed

• Use Ellipses in functions with docstrings by @jackmpcollins in #109
• Use function instead of FunctionCall in FunctionResultMessage by @jackmpcollins in #110
• Add format method to Message. Allow registering new types with message_to_openai_message by @jackmpcollins in #114
• poetry update by @jackmpcollins in #115
• Add last_message property to Chat by @jackmpcollins in #117

Full Changelog: v0.14.1...v0.15.0

---

Message subclasses now have a format method to customize their value using the function arguments when used with the @chatprompt decorator. This allows for more complex templating than just the string templating that is currently supported. For a real use case with GPT-4-vision see in-progress PR #116

from typing import Any
from magentic import chatprompt
from magentic.chat_model.message import Message


class UppercaseMessage(Message[str]):
    def format(self, **kwargs: Any) -> "UppercaseMessage":
        upper_kwargs = {k: str(v).upper() for k, v in kwargs.items()}
        return UppercaseMessage(self.content.format(**upper_kwargs))


@chatprompt(
    UppercaseMessage("hello {x}"),
)
def say_hello(x: str) -> str:
    ...


say_hello.format("world")
# [UppercaseMessage('hello WORLD')]

In addition, the openai JSON serialization for custom Message subclasses can be added without modifying magentic itself! This uses the functools.singledispatch decorator. Together, these changes will allow users to create Message classes that they can template exactly as needed, which will be useful as the variety of types of inputs to LLMs increases.

from typing import Any

from magentic.chat_model.openai_chat_model import OpenaiMessageRole, message_to_openai_message
from magentic.chat_model.message import Message
from openai.types.chat import ChatCompletionMessageParam


class CustomMessage(Message[str]):
    def format(self, **kwargs: Any) -> "CustomMessage":
        return CustomMessage(self.content)


@message_to_openai_message.register
def _(message: CustomMessage) -> ChatCompletionMessageParam:
    return {"role": OpenaiMessageRole.USER.value, "content": message.content}


message_to_openai_message(CustomMessage("hello"))
# {'role': 'user', 'content': 'hello'}

PR #114

---

Warning

Breaking Change: FunctionResultMessage init param function_call: FunctionCall replaced by function: Callable

Only the name of the function is needed when serializing the FunctionResultMessage so we do not need the whole FunctionCall. This simplifies creating @chatprompt function where the chat contains a function call. e.g.

from magentic import (
    chatprompt,
    AssistantMessage,
    UserMessage,
    FunctionCall,
    FunctionResultMessage,
)


def plus(a: int, b: int) -> int:
    return a + b


@chatprompt(
    UserMessage("Use the plus function to add 1 and 2."),
    AssistantMessage(FunctionCall(plus, 1, 2)),
    FunctionResultMessage(3, plus),
)
def do_math() -> str:
    ...


do_math()
# 'The sum of 1 and 2 is 3.'

PR #110

v0.14.1

What's Changed

• Add dependabot config by @jackmpcollins in #85
• Run poetry update by @jackmpcollins in #98
• Add ruff to pyproject.toml. Update ruff. by @jackmpcollins in #107
• Handle functions with *args and **kwargs by @jackmpcollins in #106

Dependabot

• Bump jinja2 from 3.1.2 to 3.1.3 by @dependabot in #81
• Bump jupyter-lsp from 2.2.0 to 2.2.2 by @dependabot in #82
• Bump notebook from 7.0.0 to 7.0.7 by @dependabot in #83
• Bump jupyterlab from 4.0.3 to 4.0.11 by @dependabot in #84
• Bump actions/checkout from 3 to 4 by @dependabot in #86
• Bump actions/setup-python from 4 to 5 by @dependabot in #87
• Bump pytest from 7.4.0 to 7.4.4 by @dependabot in #89
• Bump mypy from 1.4.1 to 1.8.0 by @dependabot in #92
• Bump openai from 1.1.1 to 1.9.0 by @dependabot in #88
• Bump litellm from 1.0.0 to 1.18.9 by @dependabot in #90
• Bump pytest from 7.4.4 to 8.0.0 by @dependabot in #93
• Bump pytest-asyncio from 0.21.1 to 0.23.4 by @dependabot in #94
• Bump openai from 1.10.0 to 1.12.0 by @dependabot in #105
• Bump pytest-asyncio from 0.23.3 to 0.23.5 by @dependabot in #103
• Bump litellm from 1.20.6 to 1.23.14 by @dependabot in #108

Full Changelog: v0.14.0...v0.14.1

v0.14.0

What's Changed

• Add stop param by @jackmpcollins and @mnicstruwig in #80

Full Changelog: v0.13.0...v0.14.0