Quick start

Install the client library

The library is compatible with Python 3.9 to 3.12.

Shell
pip install giskard-hub

Get your API key

Head over to your Giskard Hub instance and click on the user icon in the top right corner. You will find your personal API key, click on the button to copy it.

Note

If you don’t see your API key in the UI, it means your administrator has not enabled API keys. Please contact them to get one.

Configure your environment

You can set the following environment variables to avoid passing them as arguments to the client:

Shell
export GSK_API_KEY=your_api_key
export GSK_HUB_URL=https://your-giskard-hub-instance.com/_api

Create a project and run an evaluation

You can now use the client to interact with the Hub. You will be able to control the Hub programmatically, independently of the UI. Let’s start by initializing a client instance:

from giskard_hub import HubClient

hub = HubClient()

If you didn’t set up the environment variables, you can provide the API key and Hub URL as arguments:

hub = HubClient(
    api_key="YOUR_GSK_API_KEY",
    hub_url="THE_GSK_HUB_URL",
)

You can now use the hub client to control the LLM Hub! Let’s start by creating a fresh project.

Create a project

project = hub.projects.create(
    name="My first project",
    description="This is a test project to get started with the Giskard Hub client library",
)

That’s it! You have created a project.

Tip

If you have an already existing project, you can easily retrieve it. Either use hub.projects.list() to get a list of all projects, or use hub.projects.retrieve("YOUR_PROJECT_ID") to get a specific project.

Import a dataset

Let’s now create a dataset and add a conversation example.

# Let's create a dataset
dataset = hub.datasets.create(
    project_id=project.id,
    name="My first dataset",
    description="This is a test dataset",
)

We can now add a conversation example to the dataset. This will be used for the model evaluation.

# Add a conversation example
hub.conversations.create(
    dataset_id=dataset.id,
    messages=[
        dict(role="user", content="What is the capital of France?"),
        dict(role="assistant", content="Paris"),
        dict(role="user", content="What is the capital of Germany?"),
    ],
    demo_output=dict(role="assistant", content="I don't know that!"),
    checks=[
        dict(identifier="correctness", params={"reference": "Berlin"}),
        dict(identifier="conformity", params={"rules": ["The agent should always provide short and concise answers."]}),
    ]
)

These are the attributes you can set for a conversation (the only required attribute is messages):

  • messages: A list of messages in the conversation. Each message is a dictionary with the following keys:

    • role: The role of the message, either “user” or “assistant”.

    • content: The content of the message.

  • demo_output: A demonstration of a (possibly wrong) output from the model. This is just for demonstration purposes.

  • checks: A list of checks that the conversation should pass. This is used for evaluation. Each check is a dictionary with the following keys:
    • identifier: The identifier of the check. If it’s a built-in check, you will also need to provide the params dictionary. The built-in checks are:

      • correctness: The output of the model should match the reference.

      • conformity: The conversation should follow a set of rules.

      • groundedness: The output of the model should be grounded in the conversation.

      • string_match: The output of the model should contain a specific string (keyword or sentence).

    • params: A dictionary of parameters for the check. The parameters depend on the check type:

      • For the correctness check, the parameter is reference (type: str), which is the expected output.

      • For the conformity check, the parameter is rules (type: list[str]), which is a list of rules that the conversation should follow.

      • For the groundedness check, the parameter is context (type: str), which is the context in which the model should ground its output.

      • For the string_match check, the parameter is keyword (type: str), which is the string that the model’s output should contain.

You can add as many conversations as you want to the dataset.

Configure a model/agent

Note

In this section we will run evaluation against models configured in the Hub. If you want to evaluate a local model that is not yet exposed with an API, check the Evaluate local models.

Before running our first evaluation, we’ll need to set up a model. You’ll need an API endpoint ready to serve the model. Then, you can configure the model API in the Hub:

model = hub.models.create(
    project_id=project.id,
    name="My Bot",
    description="A chatbot for demo purposes",
    url="https://my-model-endpoint.example.com/bot_v1",
    supported_languages=["en", "fr"],
    # if your model endpoint needs special headers:
    headers={"X-API-Key": "MY_TOKEN"},
)

We can test that everything is working well by running a chat with the model:

response = model.chat(
    messages=[
        dict(role="user", content="What is the capital of France?"),
        dict(role="assistant", content="Paris"),
        dict(role="user", content="What is the capital of Germany?"),
    ],
)

print(response)

If all is working well, this will return something like

ModelOutput(
    message=ChatMessage(
        role='assistant',
        content='The capital of Germany is Berlin.'
    ),
    metadata={}
)

Run a remote evaluation

We can now launch a remote evaluation of our model!

eval_run = hub.evaluate(
    model=model,
    dataset=dataset,
    name="test-run",  # optional
)

The evaluation will run asynchronously on the Hub. For this reason, the giskard_hub.dat.EvaluationRun object returned by the evaluate method may miss some attributes (e.g. eval_run.metrics may be empty) until the evaluation is complete.

To wait until the evaluation run has finished running, you can use:

eval_run.wait_for_completion()

Once ready, you can print the evaluation metrics:

eval_run.print_metrics()
"Metrics"

Tip

You can directly pass IDs to the evaluate function, e.g. model=model_id and dataset=dataset_id, without having to retrieve the objects first.