liminal-sdk-python 2024.4.3

The Liminal SDK for Python

Liminal Python SDK

The Liminal SDK for Python provides a clean, straightforward, asyncio-based interface for interacting with the Liminal API.

• Installation
• Python Versions
• Quickstart
• Initial Authentication
  • Microsoft Entra ID
    • Device Code Flow
• Ongoing Authentication
  • Manually Interacting with the Refresh Token
  • Creating a Liminal Client from a Stored Refresh Token
• Endpoints
  • Getting Model Instances
  • Managing Threads
  • Submitting Prompts
• Connection Pooling
• Running Examples
• Contributing

Installation

pip install liminal-sdk-python

Python Versions

liminal is currently supported on:

• Python 3.11
• Python 3.12

Quickstart

Presuming the use of Microsoft Entra ID as your auth provider, instantiating a Liminal API object is easy:

import asyncio

from liminal import Client
from liminal.auth.microsoft.device_code_flow import DeviceCodeFlowProvider


async def main() -> None:
    """Create the aiohttp session and run the example."""
    # Create an auth provider to authenticate the user:
    auth_provider = DeviceCodeFlowProvider("<TENANT_ID>", "<CLIENT_ID>")

    # Create the liminal SDK instance:
    liminal = Client(auth_provider, "<LIMINAL_API_SERVER_URL>")


asyncio.run(main())

You can see several examples of how to use this API object via the examples folder in this repo.

Initial Authentication

Liminal supports the concept of authenticating via various auth providers. Currently, the following auth providers are supported:

• Microsoft Entra ID

Microsoft Entra ID

Device Code Flow

This authentication process with Microsoft Entra ID involves an OAuth 2.0 Device Authorization Grant. This flow requires you to start your app, retrieve a device code from the logs produced by this SDK, and provide that code to Microsoft via a web browser. Once you complete the login process, the SDK will be authenticated for use with your Liminal instance.

To authenticate with this flow, you will need an Entra ID client and tenant ID:

• Log into your Azure portal.
• Navigate to Microsoft Entra ID.
• Click on App registrations.
• Either create a new app registration or select an existing one.
• In the Overview of the registration, look for the Application (client) ID and Directory (tenant) ID values.

With a client ID and tenant ID, you can create a Liminal client object and authenticate it:

import asyncio

from liminal import Client
from liminal.auth.microsoft.device_code_flow import DeviceCodeFlowProvider


async def main() -> None:
    """Create the aiohttp session and run the example."""
    # Create an auth provider to authenticate the user:
    auth_provider = DeviceCodeFlowProvider("<TENANT_ID>", "<CLIENT_ID>")

    # Create the liminal SDK instance and authenticate it:
    liminal = Client(auth_provider, "<LIMINAL_API_SERVER_URL>")
    await liminal.authenticate_from_auth_provider()


asyncio.run(main())

In your application logs, you will see a message that looks like this:

INFO:liminal:To sign in, use a web browser to open the page
https://microsoft.com/devicelogin and enter the code XXXXXXXXX to authenticate.

Leaving your application running, open a browser at that URL and input the code as instructed. Once you successfully complete authentication via Entra ID, your Liminal client object will automatically authenticate with your Liminal API server.

Ongoing Authentication

After the initial authentication with your auth provider, the Liminal client object will internally store access and refresh tokens to ensure the ongoing ability to communicate with your Liminal API server. The client object will automatically handle using the stored refresh token to request new access tokens as appropriate.

Manually Interacting with the Refresh Token

The Liminal client object provides a add_refresh_token_callback method to manually interact with refresh tokens as they are generated. This is useful in situations where you want to retrieve that token and do something else with (store it in a database, for example). Every time a new refresh token is generated by the client object, the registered callbacks will be called.

Refresh token callbacks are methods that take a single str parameter (denoting the refresh token). These callbacks are not expected to return any value. Lastly, when calling add_refresh_token_callback, the returned method allows you to cancel the callback at any time.

import asyncio

from liminal import Client
from liminal.auth.microsoft.device_code_flow import DeviceCodeFlowProvider


async def main() -> None:
    """Create the aiohttp session and run the example."""
    # Create an auth provider to authenticate the user:
    auth_provider = DeviceCodeFlowProvider("<TENANT_ID>", "<CLIENT_ID>")

    # Create the liminal SDK instance and authenticate it:
    liminal = Client(microsoft_auth_provider, "<LIMINAL_API_SERVER_URL>")
    await liminal.authenticate_from_auth_provider()

    def do_something_with_refresh_token(refresh_token: str) -> None:
        """Do something with the refresh token."""
        pass

    # Register the refresh token callback:
    remove_callback = liminal.add_refresh_token_callback(
        do_something_with_refresh_token
    )

    # Later, if you want to remove the callback:
    remove_callback()


asyncio.run(main())

Creating a Liminal Client from a Stored Refresh Token

Assuming you have a stored refresh token (collected from the callback process shown above), it is simple to create a new Limina client using that token:

import asyncio

from liminal import Client


async def main() -> None:
    """Create the aiohttp session and run the example."""
    # Retrieve your stored refresh_token:
    refresh_token = "12345"

    # Create the client:
    liminal = await client.authenticate_from_refresh_token(refresh_token=refresh_token)


asyncio.run(main())

Endpoints

Getting Model Instances

Every LLM instance connected in the Liminal admin dashboard is referred to as a "model instance." The SDK provides several methods to interact with model instances:

import asyncio

from liminal import Client
from liminal.auth.microsoft.device_code_flow import DeviceCodeFlowProvider


async def main() -> None:
    # Assuming you have an authenticated `liminal` object:

    # Get all available model instances:
    model_instances = await liminal.llm.get_available_model_instances()
    # >>> [ModelInstance(...), ModelInstance(...)]

    # Get a specific model instance (if it exists):
    model_instance = await liminal.llm.get_model_instance("My Model")
    # >>> ModelInstance(...)


asyncio.run(main())

Managing Threads

Threads are conversations with an LLM instance:

import asyncio

from liminal import Client
from liminal.auth.microsoft.device_code_flow import DeviceCodeFlowProvider


async def main() -> None:
    # Assuming you have an authenticated `liminal` object:

    # Get all available threads:
    threads = await liminal.thread.get_available()
    # >>> [Thread(...), Thread(...)]

    # Get a specific thread by ID:
    thread = await liminal.thread.get_by_id(123)
    # >>> Thread(...)

    # Some operations require a model instance:
    model_instance = await liminal.llm.get_model_instance("My Model")

    # Create a new thread:
    thread = await liminal.thread.create(model_instance.id, "New Thread")
    # >>> Thread(...)


asyncio.run(main())

Submitting Prompts

Submitting prompts is easy:

import asyncio

from liminal import Client
from liminal.auth.microsoft.device_code_flow import DeviceCodeFlowProvider


async def main() -> None:
    # Assuming you have an authenticated `liminal` object:

    # Prompt operations require a model instance:
    model_instance = await liminal.llm.get_model_instance(model_instance_name)

    # Prompt operations optionally take an existing thread:
    thread = await liminal.thread.get_by_id(123)
    # >>> Thread(...)

    # Analayze a prompt for sensitive info:
    findings = await liminal.prompt.analyze(
        model_instance.id, "Here is a sensitive prompt"
    )
    # >>> AnalyzeResponse(...)

    # Cleanse input text by applying the policies defined in the Liminal admin
    # dashboard. You can optionally provide existing analysis finidings; if not
    # provided, analyze is # called automatically):
    cleansed = await liminal.prompt.cleanse(
        model_instance.id,
        "Here is a sensitive prompt",
        findings=findings,
        thread_id=thread.id,
    )
    # >>> CleanseResponse(...)

    # Submit a prompt to an LLM, cleansing it in the process (once again, providing optional
    # findings):
    response = await liminal.prompt.submit(
        model_instance.id,
        "Here is a sensitive prompt",
        findings=findings,
        thread_id=thread.id,
    )
    # >>> SubmitResponse(...)

    # Rehydrate a response with sensitive data:
    hydrated = await liminal.prompt.hydrate(
        model_instance.id, "Here is a response to rehdyrate", thread_id=thread.id
    )
    # >>> HydrateResponse(...)


asyncio.run(main())

Connection Pooling

By default, the library creates a new connection to the Liminal API server with each coroutine. If you are calling a large number of coroutines (or merely want to squeeze out every second of runtime savings possible), an httpx AsyncClient can be used for connection pooling:

import asyncio

from liminal import Client
from liminal.auth.microsoft.device_code_flow import DeviceCodeFlowProvider


async def main() -> None:
    # Create an auth provider to authenticate the user:
    microsoft_auth_provider = MicrosoftAuthProvider("<TENANT_ID>", "<CLIENT_ID>")

    # Create the liminal SDK instance with a shared HTTPX AsyncClient:
    async with httpx.AsyncClient() as client:
        liminal = Client(
            microsoft_auth_provider, "<LIMINAL_API_SERVER_URL>", httpx_client=client
        )

        # Get to work!
        # ...


asyncio.run(main())

Check out the examples, the tests, and the source files themselves for method signatures and more examples.

Running Examples

You can see examples of how to use this SDK via the examples folder in this repo. Each example follows a similar "call" format by asking for inputs via environment variables; for example:

LIMINAL_API_SERVER_URL=https://api.DOMAIN.liminal.ai \
CLIENT_ID=xxxxxxxxxxxxxxxx \
TENANT_ID=xxxxxxxxxxxxxxxx \
MODEL_INSTANCE_NAME=model-instance-name \
python3 examples/quickstart_with_microsoft.py

Contributing

Thanks to all of our contributors so far!

1. Check for open features/bugs or initiate a discussion on one.
2. Fork the repository.
3. (optional, but highly recommended) Create a virtual environment: python3 -m venv .venv
4. (optional, but highly recommended) Enter the virtual environment: source ./.venv/bin/activate
5. Install the dev environment: ./scripts/setup.sh
6. Code your new feature or bug fix on a new branch.
7. Write tests that cover your new functionality.
8. Run tests and ensure 100% code coverage: poetry run pytest --cov liminal tests
9. Update README.md with any new documentation.
10. Submit a pull request!