fix: LLM - CodeGenerationModel now supports safety attributes

PiperOrigin-RevId: 560967317
googleapis · Aug 29, 2023 · c2c8a5e · c2c8a5e
1 parent 2a08535
commit c2c8a5e
Show file tree

Hide file tree

Showing 2 changed files with 36 additions and 12 deletions.
diff --git a/tests/unit/aiplatform/test_language_models.py b/tests/unit/aiplatform/test_language_models.py
@@ -257,9 +257,9 @@
 
 _TEST_CODE_GENERATION_PREDICTION = {
     "safetyAttributes": {
-        "categories": [],
-        "blocked": False,
-        "scores": [],
+        "blocked": True,
+        "categories": ["Finance"],
+        "scores": [0.1],
     },
     "content": """
 ```python
@@ -2188,6 +2188,17 @@ def test_code_generation(self):
                 temperature=0.2,
             )
             assert response.text == _TEST_CODE_GENERATION_PREDICTION["content"]
+            expected_safety_attributes_raw = _TEST_CODE_GENERATION_PREDICTION[
+                "safetyAttributes"
+            ]
+            expected_safety_attributes = dict(
+                zip(
+                    expected_safety_attributes_raw["categories"],
+                    expected_safety_attributes_raw["scores"],
+                )
+            )
+            assert response.safety_attributes == expected_safety_attributes
+            assert response.is_blocked == expected_safety_attributes_raw["blocked"]
 
         # Validating the parameters
         predict_temperature = 0.1

diff --git a/vertexai/language_models/_language_models.py b/vertexai/language_models/_language_models.py
@@ -789,6 +789,26 @@ def predict_streaming(
             )
 
 
+def _parse_text_generation_model_response(
+    prediction_response: aiplatform.models.Prediction,
+    prediction_idx: int = 0,
+) -> TextGenerationResponse:
+    """Converts the raw text_generation model response to `TextGenerationResponse`."""
+    prediction = prediction_response.predictions[prediction_idx]
+    safety_attributes_dict = prediction.get("safetyAttributes", {})
+    return TextGenerationResponse(
+        text=prediction["content"],
+        _prediction_response=prediction_response,
+        is_blocked=safety_attributes_dict.get("blocked", False),
+        safety_attributes=dict(
+            zip(
+                safety_attributes_dict.get("categories") or [],
+                safety_attributes_dict.get("scores") or [],
+            )
+        ),
+    )
+
+
 class _ModelWithBatchPredict(_LanguageModel):
     """Model that supports batch prediction."""
 
@@ -1754,11 +1774,7 @@ def predict(
             instances=[prediction_request.instance],
             parameters=prediction_request.parameters,
         )
-
-        return TextGenerationResponse(
-            text=prediction_response.predictions[0]["content"],
-            _prediction_response=prediction_response,
-        )
+        return _parse_text_generation_model_response(prediction_response)
 
     def predict_streaming(
         self,
@@ -1800,10 +1816,7 @@ def predict_streaming(
                 predictions=[prediction_dict],
                 deployed_model_id="",
             )
-            yield TextGenerationResponse(
-                text=prediction_dict["content"],
-                _prediction_response=prediction_obj,
-            )
+            yield _parse_text_generation_model_response(prediction_obj)
 
 
 class _PreviewCodeGenerationModel(CodeGenerationModel, _TunableModelMixin):