Implementation:Kserve Kserve AlibiExplainer

Knowledge Sources	Kserve_Kserve
Domains	Explainability, Model Serving
Last Updated	2026-02-13 00:00 GMT

Overview

Concrete tool for providing a unified KServe model interface that routes explanation requests to the appropriate Alibi algorithm provided by the KServe sample code.

Description

This module defines two classes: ExplainerMethod (an Enum with values anchor_tabular, anchor_images, and anchor_text) and AlibiExplainer (which extends kserve.Model). The AlibiExplainer constructor accepts a model name, predictor host, explainer method, configuration mapping, and an optional pre-built explainer object. Based on the selected ExplainerMethod, it instantiates the corresponding wrapper class (AnchorTabular, AnchorImages, or AnchorText) with an internal _predict_fn that calls the predictor service. The explain() method delegates to the selected wrapper, converts the Alibi Explanation result to JSON, and returns it as a dictionary. The _predict_fn uses nest_asyncio to bridge synchronous Alibi calls with the async KServe prediction pipeline.

Usage

Use this class as the main explainer model server when deploying an Alibi-based explanation service alongside a KServe predictor, supporting tabular, image, or text anchor explanations.

Code Reference

Source Location

• Repository: Kserve_Kserve
• File: docs/samples/explanation/alibi/alibiexplainer/alibiexplainer/explainer.py
• Lines: 1-90

Signature

class ExplainerMethod(Enum):
    anchor_tabular = "AnchorTabular"
    anchor_images = "AnchorImages"
    anchor_text = "AnchorText"


class AlibiExplainer(kserve.Model):
    def __init__(
        self,
        name: str,
        predictor_host: str,
        method: ExplainerMethod,
        config: Mapping,
        explainer: object = None,
    ):
        ...

    def _predict_fn(self, arr: Union[np.ndarray, List]) -> np.ndarray:
        ...

    def explain(self, payload: Dict, headers: Dict[str, str] = None) -> Any:
        ...

Import

from alibiexplainer.explainer import AlibiExplainer, ExplainerMethod

I/O Contract

Inputs

Constructor

Name	Type	Required	Description
name	str	Yes	Model name for KServe routing
predictor_host	str	Yes	Hostname of the predictor service to send prediction requests to
method	ExplainerMethod	Yes	The Alibi explanation algorithm to use (anchor_tabular, anchor_images, or anchor_text)
config	Mapping	Yes	Configuration parameters passed to the selected explanation wrapper
explainer	object	No	Optional pre-built Alibi explainer object (default: None)

explain()

Name	Type	Required	Description
payload	Dict	Yes	Dictionary with "instances" key containing the data to explain
headers	Dict[str, str]	No	Optional HTTP headers

Outputs

explain()

Name	Type	Description
explanation	Dict	JSON-serializable dictionary containing the Alibi explanation result

Usage Examples

Basic Usage

from alibiexplainer.explainer import AlibiExplainer, ExplainerMethod

# Create an Alibi explainer for tabular data
explainer = AlibiExplainer(
    name="income-explainer",
    predictor_host="income-predictor.default.svc.cluster.local",
    method=ExplainerMethod.anchor_tabular,
    config={"threshold": 0.95},
    explainer=pretrained_anchor_tabular,
)

# Handle an explain request
payload = {"instances": [[25, 50000, 0.8]]}
result = explainer.explain(payload)
print(result)  # {"meta": {...}, "data": {"anchor": ["income > 40000"], ...}}

Related Pages

• Environment:Kserve_Kserve_Python_Runtime