feat(guardrails): add LLMJudgeValidator

packages/uipath-platform/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "uipath-platform"
-version = "0.1.64"
+version = "0.1.65"
 description = "HTTP client library for programmatic access to UiPath Platform"
 readme = { file = "README.md", content-type = "text/markdown" }
 requires-python = ">=3.11"

packages/uipath-platform/src/uipath/platform/guardrails/decorators/__init__.py

@@ -24,6 +24,7 @@
     GuardrailValidatorBase,
     HarmfulContentValidator,
     IntellectualPropertyValidator,
+    LLMJudgeValidator,
     PIIValidator,
     PromptInjectionValidator,
     RuleFunction,
@@ -39,6 +40,7 @@
     "CustomGuardrailValidator",
     "HarmfulContentValidator",
     "IntellectualPropertyValidator",
+    "LLMJudgeValidator",
     "PIIValidator",
     "PromptInjectionValidator",
     "UserPromptAttacksValidator",

packages/uipath-platform/src/uipath/platform/guardrails/decorators/validators/__init__.py

@@ -8,6 +8,7 @@
 from .custom import CustomValidator, RuleFunction
 from .harmful_content import HarmfulContentValidator
 from .intellectual_property import IntellectualPropertyValidator
+from .llm_judge import LLMJudgeValidator
 from .pii import PIIValidator
 from .prompt_injection import PromptInjectionValidator
 from .user_prompt_attacks import UserPromptAttacksValidator
@@ -18,6 +19,7 @@
     "CustomGuardrailValidator",
     "HarmfulContentValidator",
     "IntellectualPropertyValidator",
+    "LLMJudgeValidator",
     "PIIValidator",
     "PromptInjectionValidator",
     "UserPromptAttacksValidator",

packages/uipath-platform/src/uipath/platform/guardrails/decorators/validators/llm_judge.py

@@ -0,0 +1,87 @@
+"""LLM-as-judge guardrail validator."""
+
+from uuid import uuid4
+
+from uipath.platform.guardrails.guardrails import (
+    BuiltInValidatorGuardrail,
+    NumberParameterValue,
+    StringParameterValue,
+)
+
+from ._base import BuiltInGuardrailValidator
+
+
+class LLMJudgeValidator(BuiltInGuardrailValidator):
+    """Validate data with an LLM acting as judge against free-form criteria.
+
+    Delegates to the UiPath LLM-as-judge guardrail backend. Supported at all
+    stages — provide judging criteria written from the perspective of the
+    data being evaluated (input at PRE, output at POST).
+
+    Args:
+        criteria: Natural-language description of what the judge should check
+            for. The judge passes when the data satisfies the criteria.
+        model: LLM model identifier to use for judging. Defaults to ``"gpt-4o-mini"``.
+        threshold: Score threshold in [0.0, 1.0] above which the judge
+            considers the data compliant. Defaults to ``0.5``.
+
+    Raises:
+        ValueError: If *criteria* is empty or *threshold* is outside [0.0, 1.0].
+    """
+
+    def __init__(
+        self,
+        criteria: str,
+        model: str = "gpt-4o-mini",
+        threshold: float = 0.5,
+    ) -> None:
+        """Initialize LLMJudgeValidator with criteria, model, and threshold."""
+        if not criteria or not criteria.strip():
+            raise ValueError("criteria must be a non-empty string")
+        if not 0.0 <= threshold <= 1.0:
+            raise ValueError(f"threshold must be between 0.0 and 1.0, got {threshold}")
+        self.criteria = criteria
+        self.model = model
+        self.threshold = threshold
+
+    def get_built_in_guardrail(
+        self,
+        name: str,
+        description: str | None,
+        enabled_for_evals: bool,
+    ) -> BuiltInValidatorGuardrail:
+        """Build an LLM-as-judge :class:`BuiltInValidatorGuardrail`.
+
+        Args:
+            name: Name for the guardrail.
+            description: Optional description.
+            enabled_for_evals: Whether active in evaluation scenarios.
+
+        Returns:
+            Configured :class:`BuiltInValidatorGuardrail` for LLM-as-judge.
+        """
+        return BuiltInValidatorGuardrail(
+            id=str(uuid4()),
+            name=name,
+            description=description or f"LLM-as-judge ({self.model}): {self.criteria}",
+            enabled_for_evals=enabled_for_evals,
+            guardrail_type="builtInValidator",
+            validator_type="llm_judge",
+            validator_parameters=[
+                StringParameterValue(
+                    parameter_type="string",
+                    id="criteria",
+                    value=self.criteria,
+                ),
+                StringParameterValue(
+                    parameter_type="string",
+                    id="model",
+                    value=self.model,
+                ),
+                NumberParameterValue(
+                    parameter_type="number",
+                    id="threshold",
+                    value=self.threshold,
+                ),
+            ],
+        )

packages/uipath-platform/src/uipath/platform/guardrails/guardrails.py

@@ -37,8 +37,21 @@ class NumberParameterValue(BaseModel):
     model_config = ConfigDict(populate_by_name=True, extra="allow")
 
 
+class StringParameterValue(BaseModel):
+    """String parameter value."""
+
+    parameter_type: Literal["string"] = Field(alias="$parameterType")
+    id: str
+    value: str
+
+    model_config = ConfigDict(populate_by_name=True, extra="allow")
+
+
 ValidatorParameter = Annotated[
-    EnumListParameterValue | MapEnumParameterValue | NumberParameterValue,
+    EnumListParameterValue
+    | MapEnumParameterValue
+    | NumberParameterValue
+    | StringParameterValue,
     Field(discriminator="parameter_type"),
 ]
 

packages/uipath-platform/tests/services/test_guardrails_decorators.py

@@ -26,6 +26,7 @@
     GuardrailBlockException,
     GuardrailExclude,
     GuardrailExecutionStage,
+    LLMJudgeValidator,
     LogAction,
     LoggingSeverityLevel,
     PIIDetectionEntity,
@@ -310,6 +311,68 @@ def test_selector_is_none(self):
         assert g.selector is None
 
 
+# ---------------------------------------------------------------------------
+# 5b. LLMJudgeValidator — criteria/model/threshold validation, all stages
+# ---------------------------------------------------------------------------
+
+
+class TestLLMJudgeValidator:
+    def test_empty_criteria_raises(self):
+        with pytest.raises(ValueError, match="criteria"):
+            LLMJudgeValidator(criteria="")
+
+    def test_whitespace_only_criteria_raises(self):
+        with pytest.raises(ValueError, match="criteria"):
+            LLMJudgeValidator(criteria="   ")
+
+    def test_threshold_below_zero_raises(self):
+        with pytest.raises(ValueError, match="threshold"):
+            LLMJudgeValidator(criteria="be concise", threshold=-0.1)
+
+    def test_threshold_above_one_raises(self):
+        with pytest.raises(ValueError, match="threshold"):
+            LLMJudgeValidator(criteria="be concise", threshold=1.5)
+
+    def test_no_scope_restriction(self):
+        v = LLMJudgeValidator(criteria="be polite")
+        v.validate_stage(GuardrailExecutionStage.PRE)
+        v.validate_stage(GuardrailExecutionStage.POST)
+
+    def test_builds_llm_judge_guardrail_with_parameters(self):
+        v = LLMJudgeValidator(
+            criteria="The output must be a valid JSON object.",
+            model="gpt-4o",
+            threshold=0.8,
+        )
+        g = v.get_built_in_guardrail("Judge", None, True)
+        assert g.validator_type == "llm_judge"
+        param_by_id = {p.id: p for p in g.validator_parameters}
+        assert (
+            param_by_id["criteria"].value == "The output must be a valid JSON object."
+        )
+        assert param_by_id["model"].value == "gpt-4o"
+        assert param_by_id["threshold"].value == 0.8
+
+    def test_default_model_and_threshold(self):
+        v = LLMJudgeValidator(criteria="be polite")
+        g = v.get_built_in_guardrail("Judge", None, True)
+        param_by_id = {p.id: p for p in g.validator_parameters}
+        assert param_by_id["model"].value == "gpt-4o-mini"
+        assert param_by_id["threshold"].value == 0.5
+
+    def test_default_description_includes_model_and_criteria(self):
+        v = LLMJudgeValidator(criteria="be polite", model="gpt-4o")
+        g = v.get_built_in_guardrail("Judge", None, True)
+        assert g.description is not None
+        assert "gpt-4o" in g.description
+        assert "be polite" in g.description
+
+    def test_selector_is_none(self):
+        v = LLMJudgeValidator(criteria="be polite")
+        g = v.get_built_in_guardrail("Judge", None, True)
+        assert g.selector is None
+
+
 # ---------------------------------------------------------------------------
 # 6. CustomValidator — rule routing and error handling
 # ---------------------------------------------------------------------------