Tools

Out of the box, Opsml allows you to create tool configurations within you pyproject.toml file to help simplify some workflows. Currently, tool configuration support:

Global Default Variables

• Define global attributes that apply to all cards created in the project.
• Supports space, name, and version.

Example

[tool.opsml.default]
space = "space"
name = "my-card"

Now when you create a card, it will automatically use these values unless overridden.

modelcard = ModelCard( # (1)
    interface=model_interface,
    tags=["foo:bar", "baz:qux"],
    datacard_uid=datacard.uid,
)

model_registry.register_card(modelcard)

1. space and name are automatically set to "space" and "my-card" respectively, as defined in the pyproject.toml file.

Global Registry Variables

• A more granular way to define registry/card-specific variables.
• Allows you to specify space, name, and version for each registry.
• Each line defines a registry type (e.g., model, prompt) and its associated variables.

Example

[tool.opsml.registry]
model = { space = "opsml", name = "my-model" }
prompt = { space = "opsml", name = "my-prompt" }

modelcard = ModelCard( # (1)
    interface=model_interface,
    tags=["foo:bar", "baz:qux"],
    datacard_uid=datacard.uid,
)

model_registry.register_card(modelcard)


card = PromptCard(
    prompt=Prompt( # (2)
        model="gpt-4o",
        provider="openai",
        user_message="Provide a brief summary of the programming language $1.",
        system_message="Be concise, reply with one sentence.",
    ),
)

prompt_registry.register_card(card)

1. The modelcard will automatically use the space "opsml" and name "my-model" as defined in the pyproject.toml file for the model registry. 2. The prompt card will automatically use the space "opsml" and name "my-prompt" as defined in the pyproject.toml file for the prompt registry.

Card Decks

A Card Deck is a collection of cards that can be used to group related cards together. This allows you to:

1. Manage multiple cards as a single entity, making it easier to deploy and maintain.
2. Access a specific card in the deck by its alias. This allows you to use the card's interface to perform operations, such as making predictions with a model. Currently, you can create a card deck through the client API or through a tool configuration in your pyproject.toml file.
3. Lock and update application ML artifacts in a more declarative way.

More information on card decks can be found in the Card Decks documentation.

[[tool.opsml.deck]]
space = "space"
name = "name1"
version = "1"
write_dir = "opsml_app/app1"
cards = [
    {alias = "my_prompt", space="space", name="prompt", version = "1.*", type = "prompt"}, # (1)
    {alias = "my_model", space="space", name="model", version = "1.*", type = "model"} # (2)
]

1. The first card in the deck is a prompt card with the alias "my_prompt", which can be accessed using this alias.
2. The second card in the deck is a model card with the alias "my_model", which can also be accessed using this alias.

CardDeck Arguments

Argument	Description
space	The space to associate the CardDeck with
name	The name to give the CardDeck
version	The version of the CardDeck. This is automatically updated like other card types
write_dir	The optional directory where the CardDeck will be written to when opsml install is run
cards	A list of cards in the deck, each defined as an inline table. More information on accepted arguments is below

Card Arguments

Argument	Required	Description
alias	Yes	The alias to use for the card in the deck. This is how you will access the card in the deck
space	Yes	The space to associate the card with. This is used to identify the card in the registry
name	Yes	The name to give the card
version	No	An optional version for the card. This works similar to other card types. For example "1.*" will match any 1.x version
type	Yes	The type of card. This is used to identify the card in the registry. Currently supported types are "model", "prompt", "data"
drift	No	An optional drift configuration for model cards. This allows you to specify drift profiles, whether to deactivate other models, and the drift profile status. More information below.

Drift Configuration

Argument	Required	Description
active	No	Whether to activate the drift profile when the deck is loaded. Defaults to false
deactivate_others	No	Whether to deactivate previous model version profiles when this model is activated. Defaults to false
drift_type	Yes	A list of drift profile types to load when the deck is loaded

Example Drift Configuration

Note - TOML does not currently support multi-line inline tables. In the future, this may change to allow for more complex configurations/better formatting.

[[tool.opsml.deck]]
type = "deck"
space = "space"
name = "name2"
version = "1"
write_dir =  "opsml_app/app2"
cards = [
    {alias = "my_prompt", space="space", name="prompt", version = "1.*", type = "prompt"},
    {alias = "my_model", space="space", name="model", version = "1.*", type = "model", drift = { active = true, deactivate_others = false, drift_type = ["custom", "psi"] }}
]

Locking a CardDeck

One of the main benefits of defining a CardDeck within your pyproject.toml file is the ability to lock the deck, which will do one of the following:

• If the deck does not exist, it will create the deck and generate an opsml.lock file that contains the decks card configuration and dependencies.
• If the deck already exists, it will update the deck with the latest card configuration and dependencies (based on space, name and version arguments), and update the opsml.lock file accordingly.

Installing a CardDeck

Once you have an opsml.lock file, you can install the deck and all of its artifacts by running the following command:

opsml install

You can now load the CardDeck directly from this path. More information on loading a CardDeck can be found in the Card Decks documentation.