# Copyright 2025 Iguazio
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
from typing import Any, Dict
from transformers import pipeline
class ToxicityGuardrailStep:
def __init__(
self,
context=None,
name=None,
threshold: float = 0.5,
model_name: str = "unitary/toxic-bert",
**kwargs,
):
"""
A serving graph step that filters out toxic requests using a pre-trained
text classification model.
:param context: MLRun context object, injected automatically by the serving graph.
:param name: Name of this step in the serving graph.
:param threshold: Toxicity score threshold; requests whose toxicity score meets or
exceeds this value are blocked with a ValueError. Defaults to 0.5.
:param model_name: HuggingFace model identifier used for text classification.
Defaults to "unitary/toxic-bert".
:param kwargs: Additional keyword arguments forwarded to the serving graph step base.
"""
self.threshold = threshold
self.model_name = model_name
self._classifier = None
def post_init(self, mode="sync", **kwargs):
self._classifier = pipeline("text-classification", model=self.model_name)
def do(self, event: Dict[str, Any]) -> Dict[str, Any]:
question = event.get("question", "")
result = self._classifier(question)[0]
score = (
result["score"]
if result["label"] == "toxic"
else 1 - result["score"]
)
if score >= self.threshold:
raise ValueError(
f"Request blocked: toxicity score {score:.3f} >= {self.threshold}"
)
return event