NousResearch/funcchain

NousResearch/funcchain

Description: ⛓️ build cognitive systems, pythonic

Language: Python

License: MIT

Stars: 9

Forks: 4

Open issues: 0

Created: 2024-02-09T19:35:19Z

Pushed: 2024-02-13T03:00:59Z

Default branch: main

Fork: yes

Parent repository: shroominic/funcchain

Archived: no

README:

funcchain

 

pip install funcchain

Introduction

funcchain is the *most pythonic* way of writing cognitive systems. Leveraging pydantic models as output schemas combined with langchain in the backend allows for a seamless integration of llms into your apps. It utilizes OpenAI Functions or LlamaCpp grammars (json-schema-mode) for efficient structured output. In the backend it compiles the funcchain syntax into langchain runnables so you can easily invoke, stream or batch process your pipelines.


Simple Demo

from funcchain import chain
from pydantic import BaseModel

# define your output shape
class Recipe(BaseModel):
ingredients: list[str]
instructions: list[str]
duration: int

# write prompts utilising all native python features
def generate_recipe(topic: str) -> Recipe:
"""
Generate a recipe for a given topic.
"""
return chain() # TodoList | ShoppingList:
"""
The user input is either a shopping List or a todo list.
"""
return chain()

# the model will choose the output type automatically
lst = extract_list(
input("Enter your list: ")
)

# custom handler based on type
match lst:
case ShoppingList(items=items, store=store):
print("Here is your Shopping List: ")
for item in items:
print(f"{item.name}: {item.description}")
print(f"You need to go to: {store}")

case TodoList(todos=todos, urgency=urgency):
print("Here is your Todo List: ")
for item in todos:
print(f"{item.name}: {item.description}")
print(f"Urgency: {urgency}")

Vision Models

from funcchain import Image
from pydantic import BaseModel, Field
from funcchain import chain, settings

# set global llm using model identifiers (see MODELS.md)
settings.llm = "openai/gpt-4-vision-preview"

# everything defined is part of the prompt
class AnalysisResult(BaseModel):
"""The result of an image analysis."""

theme: str = Field(description="The theme of the image")
description: str = Field(description="A description of the image")
objects: list[str] = Field(description="A list of objects found in the image")

# easy use of images as input with structured output
def analyse_image(image: Image) -> AnalysisResult:
"""
Analyse the image and extract its
theme, description and objects.
"""
return chain()

result = analyse_image(Image.open("examples/assets/old_chinese_temple.jpg"))

print("Theme:", result.theme)
print("Description:", result.description)
for obj in result.objects:
print("Found this object:", obj)

Seamless local model support

from pydantic import BaseModel, Field
from funcchain import chain, settings

# auto-download the model from huggingface
settings.llm = "ollama/openchat"

class SentimentAnalysis(BaseModel):
analysis: str
sentiment: bool = Field(description="True for Happy, False for Sad")

def analyze(text: str) -> SentimentAnalysis:
"""
Determines the sentiment of the text.
"""
return chain()

# generates using the local model
poem = analyze("I really like when my dog does a trick!")

# promised structured output (for local models!)
print(poem.analysis)

Features

• 🐍 pythonic
• 🔀 easy swap between openai or local models
• 🔄 dynamic output types (pydantic models, or primitives)
• 👁️ vision llm support
• 🧠 langchain_core as backend
• 📝 jinja templating for prompts
• 🏗️ reliable structured output
• 🔁 auto retry parsing
• 🔧 langsmith support
• 🔄 sync, async, streaming, parallel, fallbacks
• 📦 gguf download from huggingface
• ✅ type hints for all functions and mypy support
• 🗣️ chat router component
• 🧩 composable with langchain LCEL
• 🛠️ easy error handling
• 🚦 enums and literal support
• 📐 custom parsing types

Documentation

Checkout the docs here 👈

Also highly recommend to try and run the examples in the ./examples folder.

Contribution

You want to contribute? Thanks, that's great! For more information checkout the [Contributing Guide](docs/contributing/dev-setup.md). Please run the dev setup to get started:

git clone https://github.com/shroominic/funcchain.git && cd funcchain

./dev_setup.sh