Components

What is a Component?

A component is the fundamental building block of DashAI. Every pluggable piece of functionality — models, tasks, metrics, explorers, explainers, converters, data loaders, optimizers, and jobs — is a component.

Component Types

Each component class declares a TYPE class attribute that determines its category:

TYPE	Base class	Purpose	Examples
Model	BaseModel	Train and predict	SVC, RandomForest, DistilBertTransformer
GenerativeModel	BaseGenerativeModel	Generate outputs from prompts/inputs	QwenModel, StableDiffusionV2Model
Task	BaseTask	Define ML task semantics	TextClassification, Regression, Translation
GenerativeTask	BaseGenerativeTask	Define generative task semantics	TextToTextGenerationTask, TextToImageGenerationTask, ControlNetTask
Metric	BaseMetric	Evaluate model performance	Accuracy, F1, RMSE, MAE
Explorer	BaseExplorer	Visualize and analyze data	ScatterPlotExplorer, HistogramPlotExplorer
Explainer	BaseExplainer	Interpret model predictions	KernelShap, PermutationFeatureImportance
Converter	BaseConverter	Transform features	StandardScaler, OneHotEncoder, PCA, SMOTE
DataLoader	BaseDataLoader	Load datasets from files	CSVDataLoader, ExcelDataLoader
Optimizer	BaseOptimizer	Hyperparameter optimization	Optuna-based optimizers
Job	BaseJob	Background task execution	ModelJob, ExplorerJob, PredictJob

Component Metadata

Every component can expose metadata used by the frontend for display and filtering:

• DESCRIPTION — a multilingual description of what the component does.
• DISPLAY_NAME — a human-readable name.
• COLOR — a hex color for UI rendering.
• COMPATIBLE_COMPONENTS — a list of component names this component works with (e.g., a metric that only applies to classification tasks).

---

Component Registry

The Component Registry (back/dependencies/registry/component_registry.py) is a centralized catalog of all available components. It is created during application startup and stored in the DI container.

Registration

When a component class is registered, the registry:

1. Reads the TYPE class attribute to determine the component category.
2. Checks whether the class is a configurable object (has get_schema()).
3. Extracts metadata (DESCRIPTION, DISPLAY_NAME, COLOR, etc.).
4. Stores the component in a hierarchical dictionary keyed by type and name.

Each registered component is stored as a dictionary:

{
    "name": "SVC",
    "type": "Model",
    "class": SVCClass,
    "configurable_object": True,
    "schema": { ... },       # JSON Schema if configurable
    "metadata": { ... },
    "description": MultilingualString(...),
    "display_name": MultilingualString(...),
    "color": "#3498db",
}

Lookup Methods

Method	Description
registry[name]	Direct lookup by component name
get_components_by_types(select, ignore)	Filter components by type (e.g., only Models)
get_child_components(parent_name)	Get all components that inherit from a given parent
get_related_components(component_id)	Get compatible components via COMPATIBLE_COMPONENTS

Initialization

The list of components to register on startup is defined in back/initial_components.py. Additional components can be added at runtime through the plugin system.

---

Configurable Objects

A Configurable Object is any component whose behavior can be customized through user-supplied parameters. The mechanism is built on top of Pydantic and JSON Schema.

How It Works

1. Schema definition — A component defines a SCHEMA class attribute as a Pydantic model. Each field in the model represents a configurable parameter:

   class LogisticRegressionSchema(BaseSchema):
       penalty: schema_field(
           none_type(enum_field(enum=["l1", "l2", "elasticnet"])),
           placeholder="l2",
           description=MultilingualString(
               en="Type of regularization penalty.",
               es="Tipo de penalización de regularización.",
           ),
           alias=MultilingualString(en="Penalty", es="Penalización"),
       )  # type: ignore
       C: schema_field(
           optimizer_float_field(gt=0.0),
           placeholder={"optimize": False, "fixed_value": 1.0,
                        "lower_bound": 0.01, "upper_bound": 100.0},
           description=MultilingualString(
               en="Inverse of regularization strength.",
               es="Inverso de la fuerza de regularización.",
           ),
           alias=MultilingualString(en="C", es="C"),
       )  # type: ignore

   Each field uses schema_field() with a type validator (e.g. optimizer_float_field, enum_field), a placeholder default, a bilingual description, and an alias for the UI label. The frontend uses the generated JSON Schema to render form controls; the optimizer uses type metadata to define search bounds.

2. Schema generation — get_schema() converts the Pydantic model into a JSON Schema dictionary. The frontend uses this schema to dynamically render configuration forms.

3. Validation and transformation — When the user submits a configuration, the backend calls validate_and_transform(params) which:

   • Validates raw parameter data against the Pydantic schema.
   • Recursively instantiates any nested component references (a parameter of type ComponentType is resolved into an actual component instance).

Component Fields

The component_field() utility (back/core/schema_fields/component_field.py) creates parameters that reference other components. For example, a model might accept another model as a parameter:

class BagOfWordsSchema(BaseSchema):
    tabular_classifier: schema_field(
        component_field(component_type="TabularClassificationModel"),
        placeholder=None,
        description=MultilingualString(
            en="Tabular classifier used as the underlying model.",
            es="Clasificador tabular usado como modelo subyacente.",
        ),
        alias=MultilingualString(
            en="Tabular classifier", es="Clasificador tabular"
        ),
    )  # type: ignore

The frontend renders component fields as a searchable dropdown populated from the registry. When the component is instantiated, validate_and_transform() resolves the selected component name into a live instance.