Execution & Healing

The Execution Engine is responsible for generating and applying remediations to fix identified data quality issues.

Execution Modes

ADQA operates in three distinct modes to balance autonomy with safety:

1. Advisory Mode

• Behavior: Profiles and detects issues, but only proposes actions.
• Output: An ActionPlan containing a list of recommended fixes.
• Use Case: Initial analysis or low-trust environments.

2. Human-in-the-loop (HIL)

• Behavior: Proposes an ActionPlan and waits for external approval.
• Output: Executes only the actions that have been explicitly approved.
• Use Case: Critical datasets where every change must be verified.

3. Automatic Mode

• Behavior: Fully autonomous. Detects and applies remediations immediately.
• Output: A new, healed DataFrame.
• Use Case: CI/CD pipelines, data pre-processing for ML.

Common Remediations

• DROP: Removes constant columns or redundant features.
• DEDUPE: Removes duplicate rows.
• IMPUTE: Fills missing values using mean, median, or constant values.
• CLIP: Winsorizes outliers to a specific percentile.
• MASK: Anonymizes PII data.
• TRANSFORM: Applies mathematical transformations (e.g., Log scale for skewed data).

Example: Automatic Healing

from adqa import ADQA, ADQAConfig, ExecutionMode

config = ADQAConfig(execution_mode=ExecutionMode.AUTOMATIC)
agent = ADQA.from_path("dirty_data.csv", config=config)

# This will detect issues and apply fixes in one call
result = agent.analyze()

# result.dataframe now contains the healed data
healed_df = result.dataframe