Install evidently and load reference (training) and current (production) feature data as pandas DataFrames with matching column names
Create a Report with DataDriftPreset() and optionally DataQualityPreset() to cover missing values and distribution shifts in one pass
Run report.run(reference_data=ref_df, current_data=cur_df) and call report.save_html('drift_report.html') to generate a standalone visual report
Access numerical drift results with report.as_dict() to extract per-feature drift scores and the overall dataset drift flag for alerting
Schedule the report generation daily or per batch using a workflow orchestrator (Airflow, Prefect) and publish results to a monitoring dashboard
Known gotchas
Evidently auto-selects a statistical test per column type (chi-squared for categorical, KS test for numerical) — override with column_mapping and custom test selection when domain knowledge demands a specific test
The DataDriftPreset uses a share_of_drifted_columns threshold of 0.5 by default to flag dataset-level drift; tune this threshold based on acceptable drift levels for your use case
Comparing DataFrames with mismatched column sets raises a schema mismatch error — ensure training and production feature pipelines produce identical column names and dtypes
Give your agent this knowledge — and 200+ more routes
One MCP install gives any agent live access to the full route map, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp