Glider

any_guardrail.guardrails.glider.glider

Glider

Bases: ThreeStageGuardrail[ChatMessages, str, None, str, int | None]

A prompt based guardrail from Patronus AI that utilizes pass criteria and a rubric to judge text.

For more information, see the model card:GLIDER. It outputs its reasoning, highlights for what determined the score, and an integer score.

Parameters:

Name	Type	Description	Default
model_id	str | None	HuggingFace path to model.	None
pass_criteria	str	A question or description of what you are validating.	required
rubric	str	A scoring rubric, describing to the model how to score the provided data.	required
provider	StandardProvider | None	Reserved for future extensibility. Currently unused.	None

Raise

ValueError: Can only use model path to GLIDER from HuggingFace.

Source code in src/any_guardrail/guardrails/glider/glider.py

class Glider(ThreeStageGuardrail[ChatMessages, str, None, str, int | None]):
    """A prompt based guardrail from Patronus AI that utilizes pass criteria and a rubric to judge text.

    For more information, see the model card:[GLIDER](https://huggingface.co/PatronusAI/glider). It outputs its reasoning,
    highlights for what determined the score, and an integer score.

    Args:
        model_id: HuggingFace path to model.
        pass_criteria: A question or description of what you are validating.
        rubric: A scoring rubric, describing to the model how to score the provided data.
        provider: Reserved for future extensibility. Currently unused.

    Raise:
        ValueError: Can only use model path to GLIDER from HuggingFace.

    """

    SUPPORTED_MODELS: ClassVar = ["PatronusAI/glider"]

    def __init__(
        self,
        pass_criteria: str,
        rubric: str,
        model_id: str | None = None,
        provider: StandardProvider | None = None,  # Reserved for future extensibility
    ) -> None:
        """Initialize the GLIDER guardrail."""
        self.model_id = default(model_id, self.SUPPORTED_MODELS)
        self.pass_criteria = pass_criteria
        self.rubric = rubric
        self.system_prompt = SYSTEM_PROMPT_GLIDER
        self.provider = provider  # Reserved for future extensibility
        self.model = pipeline("text-generation", self.model_id, max_new_tokens=2048, return_full_text=False)

    def validate(self, input_text: str, output_text: str | None = None) -> GuardrailOutput[None, str, int | None]:  # type: ignore[override]
        """Use the provided pass criteria and rubric to judge the input and output text provided.

        Args:
            input_text: The initial text to evaluate.
            output_text: Optional subsequent text to evaluate alongside input.

        Returns:
            GuardrailOutput with explanation in the format provided by the system prompt.

        """
        model_inputs = self._pre_processing(input_text, output_text)
        model_outputs = self._inference(model_inputs)
        return self._post_processing(model_outputs)

    def _pre_processing(
        self, input_text: str, output_text: str | None = None
    ) -> GuardrailPreprocessOutput[ChatMessages]:
        if output_text is None:
            data = INPUT_DATA_FORMAT.format(input_text=input_text)
        else:
            data = INPUT_OUTPUT_DATA_FORMAT.format(input_text=input_text, output_text=output_text)
        prompt = self.system_prompt.format(data=data, pass_criteria=self.pass_criteria, rubric=self.rubric)
        return GuardrailPreprocessOutput(data=[{"role": "user", "content": prompt}])

    def _inference(self, message: GuardrailPreprocessOutput[ChatMessages]) -> GuardrailInferenceOutput[str]:
        """Run text-generation pipeline on chat messages."""
        generated_text: str = self.model(message.data)[0]["generated_text"]  # type: ignore[assignment]
        return GuardrailInferenceOutput(data=generated_text)

    def _post_processing(self, model_outputs: GuardrailInferenceOutput[str]) -> GuardrailOutput[None, str, int | None]:
        score = re.findall(r"<score>\n(\d+)\n</score>", model_outputs.data)
        if len(score) != 0 and score[0].isdigit():
            final_score = int(score[0])
        else:
            final_score = None

        return GuardrailOutput(explanation=model_outputs.data, score=final_score)

__init__(pass_criteria, rubric, model_id=None, provider=None)

Initialize the GLIDER guardrail.

Source code in src/any_guardrail/guardrails/glider/glider.py

def __init__(
    self,
    pass_criteria: str,
    rubric: str,
    model_id: str | None = None,
    provider: StandardProvider | None = None,  # Reserved for future extensibility
) -> None:
    """Initialize the GLIDER guardrail."""
    self.model_id = default(model_id, self.SUPPORTED_MODELS)
    self.pass_criteria = pass_criteria
    self.rubric = rubric
    self.system_prompt = SYSTEM_PROMPT_GLIDER
    self.provider = provider  # Reserved for future extensibility
    self.model = pipeline("text-generation", self.model_id, max_new_tokens=2048, return_full_text=False)

validate(input_text, output_text=None)

Use the provided pass criteria and rubric to judge the input and output text provided.

Parameters:

Name	Type	Description	Default
input_text	str	The initial text to evaluate.	required
output_text	str | None	Optional subsequent text to evaluate alongside input.	None

Returns:

Type	Description
GuardrailOutput[None, str, int | None]	GuardrailOutput with explanation in the format provided by the system prompt.

Source code in src/any_guardrail/guardrails/glider/glider.py

def validate(self, input_text: str, output_text: str | None = None) -> GuardrailOutput[None, str, int | None]:  # type: ignore[override]
    """Use the provided pass criteria and rubric to judge the input and output text provided.

    Args:
        input_text: The initial text to evaluate.
        output_text: Optional subsequent text to evaluate alongside input.

    Returns:
        GuardrailOutput with explanation in the format provided by the system prompt.

    """
    model_inputs = self._pre_processing(input_text, output_text)
    model_outputs = self._inference(model_inputs)
    return self._post_processing(model_outputs)