Merge branch 'feat/toxicity-hub-validators' into feat/toxicity-huggingface-model

Author: rkritika1508

backend/app/core/validators/README.md

@@ -407,7 +407,7 @@ Notes / limitations:
   ```bash
   GUARDRAILS_HUB_API_KEY=<your-key> ENABLE_REMOTE_INFERENCING=true bash scripts/install_guardrails_from_hub.sh
   ```
-- `on_fail=fix` behaves like `on_fail=exception` — LlamaGuard has no programmatic fix, so validation stops immediately on failure to prevent downstream validators from receiving `None` as input.
+- `on_fail=fix` returns `""` on failure — LlamaGuard has no programmatic fix, so `safe_text` will be `""` and the response `metadata.reason` will identify this validator as the cause.
 - LlamaGuard policy classification may produce false positives in news, clinical, or legal contexts.
 
 ### 9) NSFW Text Validator (`nsfw_text`)

backend/app/core/validators/config/base_validator_config.py

@@ -19,8 +19,14 @@ def _on_fix(self, value: str, fail_result: FailResult):
         fix_value = fail_result.fix_value if fail_result else None
         if not fix_value:
             self.validator_metadata = {
+<<<<<<< feat/toxicity-huggingface-model
                 "reason": f"Empty string has been returned since the validation failed for: {self.type}"
             }
+=======
+                "reason": f"Empty string has been returned since the validation failed for: {self.type}"  # type: ignore[attr-defined]
+            }
+            return ""
+>>>>>>> feat/toxicity-hub-validators
         return fix_value
 
     def resolve_on_fail(self):

backend/app/core/validators/config/llamaguard_7b_safety_validator_config.py

@@ -1,6 +1,5 @@
 from typing import List, Literal, Optional
 
-from guardrails import OnFailAction
 from guardrails.hub import LlamaGuard7B
 
 from app.core.enum import GuardrailOnFail
@@ -34,12 +33,7 @@ def _resolve_policies(self) -> Optional[List[str]]:
         return resolved
 
     def build(self):
-        on_fail = self.resolve_on_fail()
-        # LlamaGuard7B has no programmatic fix. If on_fail=fix is requested,
-        # fall back to exception so downstream validators don't receive None as input.
-        if self.on_fail == GuardrailOnFail.Fix:
-            on_fail = OnFailAction.EXCEPTION
         return LlamaGuard7B(
             policies=self._resolve_policies(),
-            on_fail=on_fail,  # type: ignore[arg-type]
+            on_fail=self.resolve_on_fail(),  # type: ignore[arg-type]
         )

backend/app/tests/test_toxicity_hub_validators.py

@@ -107,16 +107,29 @@ def test_build_returns_validator_instance(self):
 
         assert result == mock_validator.return_value
 
-    def test_on_fail_fix_remaps_to_exception(self):
-        # LlamaGuard has no programmatic fix; on_fail=fix is silently remapped to
-        # exception to prevent downstream validators from receiving None as input.
+    def test_on_fail_fix_resolves_to_callable(self):
         config = LlamaGuard7BSafetyValidatorConfig(type="llamaguard_7b", on_fail="fix")
 
         with patch(_LLAMAGUARD_PATCH) as mock_validator:
             config.build()
 
         _, kwargs = mock_validator.call_args
-        assert kwargs["on_fail"] == OnFailAction.EXCEPTION
+        assert callable(kwargs["on_fail"])
+
+    def test_on_fix_sets_validator_metadata_when_fix_value_empty(self):
+        from unittest.mock import MagicMock
+        from guardrails.validators import FailResult
+
+        config = LlamaGuard7BSafetyValidatorConfig(type="llamaguard_7b", on_fail="fix")
+        fail_result = MagicMock(spec=FailResult)
+        fail_result.fix_value = None
+
+        result = config._on_fix("some unsafe input", fail_result)
+
+        assert result == ""
+        assert config.validator_metadata == {
+            "reason": "Empty string has been returned since the validation failed for: llamaguard_7b"
+        }
 
     def test_on_fail_exception_resolves_to_exception_action(self):
         config = LlamaGuard7BSafetyValidatorConfig(