Simple, unified interface to multiple Generative AI providers.
aisuite makes it easy for developers to interact with multiple Gen-AI services through a standardized interface. Using an interface similar to OpenAI's, aisuite supports chat completions and audio transcription, making it easy to work with the most popular AI providers and compare results. It is a thin wrapper around python client libraries, and allows creators to seamlessly swap out and test different providers without changing their code.
All of the top providers are supported. Sample list of supported providers include - Anthropic, AWS, Azure, Cerebras, Cohere, Google, Groq, HuggingFace, Ollama, Mistral, OpenAI, Sambanova, Watsonx and others.
To maximize stability, aisuite uses either the HTTP endpoint or the SDK for making calls to the provider.
You can install just the base aisuite package, or install a provider's package along with aisuite.
This installs just the base package without installing any provider's SDK.
pip install aisuiteThis installs aisuite along with anthropic's library.
pip install 'aisuite[anthropic]'This installs all the provider-specific libraries
pip install 'aisuite[all]'To get started, you will need API Keys for the providers you intend to use. You'll need to install the provider-specific library either separately or when installing aisuite.
The API Keys can be set as environment variables, or can be passed as config to the aisuite Client constructor.
You can use tools like python-dotenv or direnv to set the environment variables manually. Please take a look at the examples folder to see usage.
Here is a short example of using aisuite to generate chat completion responses from gpt-4o and claude-3-5-sonnet.
Set the API keys.
export OPENAI_API_KEY="your-openai-api-key"
export ANTHROPIC_API_KEY="your-anthropic-api-key"Use the python client.
import aisuite as ai
client = ai.Client()
models = ["openai:gpt-4o", "anthropic:claude-3-5-sonnet-20240620"]
messages = [
{"role": "system", "content": "Respond in Pirate English."},
{"role": "user", "content": "Tell me a joke."},
]
for model in models:
response = client.chat.completions.create(
model=model,
messages=messages,
temperature=0.75
)
print(response.choices[0].message.content)Note that the model name in the create() call uses the format - <provider>:<model-name>.
aisuite will call the appropriate provider with the right parameters based on the provider value.
For a list of provider values, you can look at the directory - aisuite/providers/. The list of supported providers are of the format - <provider>_provider.py in that directory. We welcome providers adding support to this library by adding an implementation file in this directory. Please see section below for how to contribute.
For more examples, check out the examples directory where you will find several notebooks that you can run to experiment with the interface.
We have made easy for a provider or volunteer to add support for a new platform.
We follow a convention-based approach for loading providers, which relies on strict naming conventions for both the module name and the class name. The format is based on the model identifier in the form provider:model.
- The provider's module file must be named in the format
<provider>_provider.py. - The class inside this module must follow the format: the provider name with the first letter capitalized, followed by the suffix
Provider.
-
Hugging Face: The provider class should be defined as:
class HuggingfaceProvider(BaseProvider)
in providers/huggingface_provider.py.
-
OpenAI: The provider class should be defined as:
class OpenaiProvider(BaseProvider)
in providers/openai_provider.py
This convention simplifies the addition of new providers and ensures consistency across provider implementations.
aisuite provides a simple abstraction for tool/function calling that works across supported providers. This is in addition to the regular abstraction of passing JSON spec of the tool to the tools parameter. The tool calling abstraction makes it easy to use tools with different LLMs without changing your code.
There are two ways to use tools with aisuite:
This is the default behavior when max_turns is not specified.
You can pass tools in the OpenAI tool format:
def will_it_rain(location: str, time_of_day: str):
"""Check if it will rain in a location at a given time today.
Args:
location (str): Name of the city
time_of_day (str): Time of the day in HH:MM format.
"""
return "YES"
tools = [{
"type": "function",
"function": {
"name": "will_it_rain",
"description": "Check if it will rain in a location at a given time today",
"parameters": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "Name of the city"
},
"time_of_day": {
"type": "string",
"description": "Time of the day in HH:MM format."
}
},
"required": ["location", "time_of_day"]
}
}
}]
response = client.chat.completions.create(
model="openai:gpt-4o",
messages=messages,
tools=tools
)When max_turns is specified, you can pass a list of callable Python functions as the tools parameter. aisuite will automatically handle the tool calling flow:
def will_it_rain(location: str, time_of_day: str):
"""Check if it will rain in a location at a given time today.
Args:
location (str): Name of the city
time_of_day (str): Time of the day in HH:MM format.
"""
return "YES"
client = ai.Client()
messages = [{
"role": "user",
"content": "I live in San Francisco. Can you check for weather "
"and plan an outdoor picnic for me at 2pm?"
}]
# Automatic tool execution with max_turns
response = client.chat.completions.create(
model="openai:gpt-4o",
messages=messages,
tools=[will_it_rain],
max_turns=2 # Maximum number of back-and-forth tool calls
)
print(response.choices[0].message.content)When max_turns is specified, aisuite will:
- Send your message to the LLM
- Execute any tool calls the LLM requests
- Send the tool results back to the LLM
- Repeat until the conversation is complete or max_turns is reached
In addition to response.choices[0].message, there is an additional field response.choices[0].intermediate_messages: which contains the list of all messages including tool interactions used. This can be used to continue the conversation with the model.
For more detailed examples of tool calling, check out the examples/tool_calling_abstraction.ipynb notebook.
Note: Audio transcription support is currently under development. The API and features described below are subject to change.
aisuite provides audio transcription (speech-to-text) with the same unified interface pattern used for chat completions. Transcribe audio files across multiple providers with consistent code.
import aisuite as ai
client = ai.Client()
# Transcribe an audio file
result = client.audio.transcriptions.create(
model="openai:whisper-1",
file="meeting.mp3"
)
print(result.text)
# Switch providers without changing your code
result = client.audio.transcriptions.create(
model="deepgram:nova-2",
file="meeting.mp3"
)
print(result.text)Use OpenAI-style parameters that work across all providers:
result = client.audio.transcriptions.create(
model="openai:whisper-1",
file="interview.mp3",
language="en", # Specify audio language
prompt="Technical discussion about AI", # Context hints
temperature=0.2 # Sampling temperature (where supported)
)These parameters are automatically mapped to each provider's native format.
Each provider offers unique capabilities you can access directly:
OpenAI Whisper:
result = client.audio.transcriptions.create(
model="openai:whisper-1",
file="speech.mp3",
response_format="verbose_json", # Get detailed metadata
timestamp_granularities=["word"] # Word-level timestamps
)Deepgram:
result = client.audio.transcriptions.create(
model="deepgram:nova-2",
file="meeting.mp3",
punctuate=True, # Auto-add punctuation
diarize=True, # Identify speakers
sentiment=True, # Sentiment analysis
summarize=True # Auto-summarization
)Google Speech-to-Text:
result = client.audio.transcriptions.create(
model="google:default",
file="call.mp3",
enable_automatic_punctuation=True,
enable_speaker_diarization=True,
diarization_speaker_count=2
)Hugging Face:
result = client.audio.transcriptions.create(
model="huggingface:openai/whisper-large-v3",
file="presentation.mp3",
return_timestamps="word" # Word-level timestamps
)For real-time or large audio files, use streaming:
async def transcribe_stream():
stream = client.audio.transcriptions.create_stream_output(
model="deepgram:nova-2",
file="long_recording.mp3"
)
async for chunk in stream:
print(chunk.text, end="", flush=True)
if chunk.is_final:
print() # New line for final results
# Run the async function
import asyncio
asyncio.run(transcribe_stream())- OpenAI:
whisper-1 - Deepgram:
nova-2,nova,enhanced,base - Google:
default,latest_long,latest_short - Hugging Face:
openai/whisper-large-v3,openai/whisper-tiny,facebook/wav2vec2-base-960h,facebook/wav2vec2-large-xlsr-53
Install transcription providers:
# Install with specific provider
pip install 'aisuite[openai]' # For OpenAI Whisper
pip install 'aisuite[deepgram]' # For Deepgram
pip install 'aisuite[google]' # For Google Speech-to-Text
pip install 'aisuite[huggingface]' # For Hugging Face models
# Install all providers
pip install 'aisuite[all]'Set API keys:
export OPENAI_API_KEY="your-openai-api-key"
export DEEPGRAM_API_KEY="your-deepgram-api-key"
export GOOGLE_APPLICATION_CREDENTIALS="path/to/credentials.json"
export HF_TOKEN="your-huggingface-token"For more examples and advanced usage, check out examples/asr_example.ipynb.
aisuite is released under the MIT License. You are free to use, modify, and distribute the code for both commercial and non-commercial purposes.
If you would like to contribute, please read our Contributing Guide and join our Discord server!