kani (カニ) is a lightweight and highly hackable framework for chat-based language models with tool usage/function calling.
Compared to other LM frameworks, kani is less opinionated and offers more fine-grained customizability over the parts of the control flow that matter, making it the perfect choice for NLP researchers, hobbyists, and developers alike.
kani comes with support for the following models out of the box, with a model-agnostic framework to add support for many more:
Hosted Models
- OpenAI Models (GPT-3.5-turbo, GPT-4, GPT-4-turbo, GPT-4o)
- Anthropic Models (Claude, Claude Instant)
Open Source Models
kani supports every chat model available on Hugging Face through transformers or llama.cpp!
In particular, we have reference implementations for the following base models, and their fine-tunes:
- LLaMA 3 (all sizes)
- Mistral-7B, Mixtral-8x7B, and Mixtral-8x22B
- Command R and Command R+
- Gemma (all sizes)
- LLaMA 2 (all sizes)
- Vicuna v1.3
Check out the Model Zoo to see how to use each of these models in your application!
Interested in contributing? Check out our guide.
- Lightweight and high-level - kani implements common boilerplate to interface with language models without forcing you to use opinionated prompt frameworks or complex library-specific tooling.
- Model agnostic - kani provides a simple interface to implement: token counting and completion generation. kani lets developers switch which language model runs on the backend without major code refactors.
- Automatic chat memory management - Allow chat sessions to flow without worrying about managing the number of tokens in the history - kani takes care of it.
- Function calling with model feedback and retry - Give models access to functions in just one line of code. kani elegantly provides feedback about hallucinated parameters and errors and allows the model to retry calls.
- You control the prompts - There are no hidden prompt hacks. We will never decide for you how to format your own data, unlike other popular language model libraries.
- Fast to iterate and intuitive to learn - With kani, you only write Python - we handle the rest.
- Asynchronous design from the start - kani can scale to run multiple chat sessions in parallel easily, without having to manage multiple processes or programs.
kani requires Python 3.10 or above. To install model-specific dependencies, kani uses various extras (brackets after
the library name in pip install). To determine which extra(s) to install, see
the model table, or use the [all] extra to install everything.
# for OpenAI models
$ pip install "kani[openai]"
# for Hugging Face models
$ pip install "kani[huggingface]" torch
# or install everything:
$ pip install "kani[all]"For the most up-to-date changes and new models, you can also install the development version from Git's main branch:
$ pip install "kani[all] @ git+https://github.com/zhudotexe/kani.git@main"
kani requires Python 3.10 or above.
First, install the library. In this quickstart, we'll use the OpenAI engine, though kani is model-agnostic.
$ pip install "kani[openai]"Then, let's use kani to create a simple chatbot using ChatGPT as a backend.
# import the library
import asyncio
from kani import Kani, chat_in_terminal
from kani.engines.openai import OpenAIEngine
# Replace this with your OpenAI API key: https://platform.openai.com/account/api-keys
api_key = "sk-..."
# kani uses an Engine to interact with the language model. You can specify other model
# parameters here, like temperature=0.7.
engine = OpenAIEngine(api_key, model="gpt-4o-mini")
# The kani manages the chat state, prompting, and function calling. Here, we only give
# it the engine to call ChatGPT, but you can specify other parameters like
# system_prompt="You are..." here.
ai = Kani(engine)
# kani comes with a utility to interact with a kani through your terminal...
chat_in_terminal(ai)
# or you can use kani programmatically in an async function!
async def main():
resp = await ai.chat_round("What is the airspeed velocity of an unladen swallow?")
print(resp.text)
asyncio.run(main())kani makes the time to set up a working chat model short, while offering the programmer deep customizability over every prompt, function call, and even the underlying language model.
Function calling gives language models the ability to choose when to call a function you provide based off its documentation.
With kani, you can write functions in Python and expose them to the model with just one line of code: the @ai_function
decorator.
# import the library
import asyncio
from typing import Annotated
from kani import AIParam, Kani, ai_function, chat_in_terminal, ChatRole
from kani.engines.openai import OpenAIEngine
# set up the engine as above
api_key = "sk-..."
engine = OpenAIEngine(api_key, model="gpt-4o-mini")
# subclass Kani to add AI functions
class MyKani(Kani):
# Adding the annotation to a method exposes it to the AI
@ai_function()
def get_weather(
self,
# and you can provide extra documentation about specific parameters
location: Annotated[str, AIParam(desc="The city and state, e.g. San Francisco, CA")],
):
"""Get the current weather in a given location."""
# In this example, we mock the return, but you could call a real weather API
return f"Weather in {location}: Sunny, 72 degrees fahrenheit."
ai = MyKani(engine)
# the terminal utility allows you to test function calls...
chat_in_terminal(ai)
# and you can track multiple rounds programmatically.
async def main():
async for msg in ai.full_round("What's the weather in Tokyo?"):
print(msg.role, msg.text)
asyncio.run(main())kani guarantees that function calls are valid by the time they reach your methods while allowing you to focus on writing code. For more information, check out the function calling docs.
kani supports streaming responses from the underlying language model token-by-token, even in the presence of function
calls. Streaming is designed to be a drop-in superset of the chat_round and full_round methods, allowing you to
gradually refactor your code without ever leaving it in a broken state.
async def stream_chat():
stream = ai.chat_round_stream("What does kani mean?")
async for token in stream:
print(token, end="")
print()
msg = await stream.message() # or `await stream`
async def stream_with_function_calling():
async for stream in ai.full_round_stream("What's the weather in Tokyo?"):
async for token in stream:
print(token, end="")
print()
msg = await stream.message()Existing frameworks for language models like LangChain and simpleaichat are opinionated and/or heavyweight - they edit developers' prompts under the hood, are challenging to learn, and are difficult to customize without adding a lot of high-maintenance bloat to your codebase.
We built kani as a more flexible, simple, and robust alternative. A good analogy between frameworks would be to say that kani is to LangChain as Flask (or FastAPI) is to Django.
kani is appropriate for everyone from academic researchers to industry professionals to hobbyists to use without worrying about under-the-hood hacks.
To learn more about how to customize kani with your own prompt wrappers, function calling, and more, read the docs!
Or take a look at the hands-on examples in this repo.
Want to see kani in action? Using 4-bit quantization to shrink the model, we run LLaMA v2 as part of our test suite right on GitHub Actions:
https://github.com/zhudotexe/kani/actions/workflows/pytest.yml?query=branch%3Amain+is%3Asuccess
Simply click on the latest build to see LLaMA's output!
The core development team is made of three PhD students in the Department of Computer and Information Science at the University of Pennsylvania. We're all members of Prof. Chris Callison-Burch's lab, working towards advancing the future of NLP.
- Andrew Zhu started in Fall 2022. His research interests include natural language processing, programming languages, distributed systems, and more. He's also a full-stack software engineer, proficient in all manner of backend, devops, database, and frontend engineering. Andrew strives to make idiomatic, clean, performant, and low-maintenance code — philosophies that are often rare in academia. His research is supported by the NSF Graduate Research Fellowship.
- Liam Dugan started in Fall 2021. His research focuses primarily on large language models and how humans interact with them. In particular, he is interested in human detection of generated text and whether we can apply those insights to automatic detection systems. He is also interested in the practical application of large language models to education.
- Alyssa Hwang started in Fall 2020 and is advised by Chris Callison-Burch and Andrew Head. Her research focuses on AI assistants that effectively communicate complex information, like voice assistants guiding users through instructions or audiobooks allowing users to seamlessly navigate through spoken text. Beyond research, Alyssa chairs the Penn CIS Doctoral Association, founded the CIS PhD Mentorship Program, and was supported by the NSF Graduate Research Fellowship Program.
We use kani actively in our research, and aim to keep it up-to-date with modern NLP practices.
If you use Kani, please cite us as:
@inproceedings{zhu-etal-2023-kani,
title = "Kani: A Lightweight and Highly Hackable Framework for Building Language Model Applications",
author = "Zhu, Andrew and
Dugan, Liam and
Hwang, Alyssa and
Callison-Burch, Chris",
editor = "Tan, Liling and
Milajevs, Dmitrijs and
Chauhan, Geeticka and
Gwinnup, Jeremy and
Rippeth, Elijah",
booktitle = "Proceedings of the 3rd Workshop for Natural Language Processing Open Source Software (NLP-OSS 2023)",
month = dec,
year = "2023",
address = "Singapore",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.nlposs-1.8",
doi = "10.18653/v1/2023.nlposs-1.8",
pages = "65--77",
}
We would like to thank the members of the lab of Chris Callison-Burch for their testing and detailed feedback on the contents of both our paper and the Kani repository. In addition, we’d like to thank Henry Zhu (no relation to the first author) for his early and enthusiastic support of the project.
This research is based upon work supported in part by the Air Force Research Laboratory (contract FA8750-23-C-0507), the IARPA HIATUS Program (contract 2022-22072200005), and the NSF (Award 1928631). Approved for Public Release, Distribution Unlimited. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of IARPA, NSF, or the U.S. Government.
