Implementation:Evidentlyai Evidently Embeddings Drift Metric

Knowledge Sources	Evidentlyai_Evidently
Domains	Metrics, Data Drift, Embeddings
Last Updated	2026-02-14 12:00 GMT

Overview

Implements the EmbeddingsDrift metric for detecting distribution drift in embedding columns using configurable drift detection methods.

Description

The Embeddings Drift Metric module provides a modern metric interface for detecting drift in embedding representations between current and reference datasets. It wraps the legacy EmbeddingsDriftMetric through the LegacyMetricCalculation bridge pattern.

Key classes:

• EmbeddingsDrift -- A SingleValueMetric that specifies which embeddings to analyze (embeddings_name) and an optional drift_method for the detection algorithm. The drift score is returned as a single numeric value.

• EmbeddingsDriftCalculation -- The calculation class that combines SingleValueCalculation and LegacyMetricCalculation. It overrides _gen_input_data to ensure embedding column definitions are properly propagated to the legacy input data. The legacy_metric method creates the corresponding EmbeddingsDriftMetric instance. The calculate_value method extracts the drift score from the legacy result and attaches the legacy rendering widgets.

The module re-exports the following drift method classes for convenience:

• ModelDriftMethod -- Uses a trained classifier to detect drift
• DistanceDriftMethod -- Uses distance metrics between distributions
• RatioDriftMethod -- Uses share of drifted embedding components
• MMDDriftMethod -- Uses Maximum Mean Discrepancy

Usage

Use EmbeddingsDrift in an Evidently report to detect whether the distribution of named embedding columns has shifted between reference and current datasets. Specify the embeddings_name that corresponds to an entry in the data definition's embeddings mapping. Optionally provide a drift_method to control the detection algorithm.

Code Reference

Source Location

• Repository: Evidentlyai_Evidently
• File: src/evidently/metrics/embeddings.py

Signature

class EmbeddingsDrift(SingleValueMetric):
    embeddings_name: str
    drift_method: Optional[DriftMethod] = None

class EmbeddingsDriftCalculation(
    SingleValueCalculation[EmbeddingsDrift],
    LegacyMetricCalculation[SingleValue, EmbeddingsDrift, EmbeddingsDriftMetricResults, EmbeddingsDriftMetric],
):
    def _gen_input_data(self, context: "Context", task_name: Optional[str]) -> InputData:
        ...
    def legacy_metric(self) -> EmbeddingsDriftMetric:
        ...
    def calculate_value(self, context, legacy_result, render) -> SingleValue:
        ...
    def display_name(self) -> str:
        ...

Import

from evidently.metrics.embeddings import (
    EmbeddingsDrift,
    EmbeddingsDriftCalculation,
    ModelDriftMethod,
    DistanceDriftMethod,
    RatioDriftMethod,
    MMDDriftMethod,
)

I/O Contract

Inputs

Name	Type	Required	Description
embeddings_name	str	Yes	Name of the embedding group as defined in the data definition's embeddings mapping
drift_method	Optional[DriftMethod]	No	Drift detection method to use (ModelDriftMethod, DistanceDriftMethod, RatioDriftMethod, or MMDDriftMethod); defaults to None (auto-selected)

Outputs

Name	Type	Description
result	SingleValue	A single numeric drift score value with attached visualization widgets
display_name	str	Formatted as "Drift in embeddings: {embeddings_name}"

Usage Examples

from evidently.metrics.embeddings import EmbeddingsDrift, MMDDriftMethod
from evidently.core.report import Report

# Basic embeddings drift detection
report = Report(metrics=[
    EmbeddingsDrift(embeddings_name="text_embeddings"),
])

# With a specific drift method
report = Report(metrics=[
    EmbeddingsDrift(
        embeddings_name="image_embeddings",
        drift_method=MMDDriftMethod(),
    ),
])

report.run(reference_data=ref_df, current_data=curr_df)

Related Pages

• Environment:Evidentlyai_Evidently_Python_Core_Environment