fix(analyzer): make the shipped HuggingFaceNerRecognizer entry loadable

presidio-analyzer/presidio_analyzer/predefined_recognizers/ner/huggingface_ner_recognizer.py

@@ -257,17 +257,13 @@ def load(self) -> None:
         This method handles:
         1. Hardware acceleration setup (CUDA validation and fallback)
         2. Lazy-loading of the heavyweight ML pipeline.
-
-        :raises ValueError: If model_name is not set
         """
         if self.ner_pipeline is not None:
             return
 
         if not self.model_name:
-            raise ValueError(
-                "model_name must be set before calling load(). "
-                "Pass it to __init__() or set it directly."
-            )
+            logger.info("model_name is not set. Deferring HuggingFace model loading.")
+            return
 
         # Device validation and fallback
         device = self.device
@@ -430,6 +426,13 @@ def analyze(
         # Defensive guard for entities input
         entities = entities or []
 
+        if not self.model_name:
+            raise ValueError(
+                "model_name must be specified to analyze text using "
+                "HuggingFaceNerRecognizer. Specify model_name when instantiating "
+                "HuggingFaceNerRecognizer or in default_recognizers.yaml."
+            )
+
         if not self.ner_pipeline:
             self.load()
 

presidio-analyzer/tests/test_huggingface_ner_recognizer.py

@@ -256,10 +256,14 @@ def test_hf_recognizer_load_errors():
         with pytest.raises(ImportError):
             HuggingFaceNerRecognizer(model_name="test")
 
-    # 2. Test ValueError when model_name is missing
-    with patch(path, new=MagicMock()):
-        with pytest.raises(ValueError, match="model_name must be set"):
-            HuggingFaceNerRecognizer(model_name=None)
+
+@pytest.mark.usefixtures("mock_torch_installed")
+def test_analyze_without_model_name_raises_with_configuration_pointer():
+    """Verify calling analyze() on HuggingFaceNerRecognizer without model_name raises ValueError with configuration pointer."""
+    with patch(HF_PIPELINE_PATH, new=MagicMock()):
+        recognizer = HuggingFaceNerRecognizer(model_name=None)
+        with pytest.raises(ValueError, match="default_recognizers.yaml"):
+            recognizer.analyze("Some text to analyze", ["PERSON"])
 
 
 @pytest.mark.usefixtures("mock_torch_installed")
@@ -723,3 +727,13 @@ def test_hf_recognizer_resolves_deferred_tokenizer_chunker(mock_pipeline):
     assert chunker.tokenizer is mock_tokenizer
     assert chunker.max_tokens == 128
     assert chunker.overlap_tokens == 16
+
+
+@pytest.mark.usefixtures("mock_torch_installed")
+def test_registry_builds_with_huggingface_entry_enabled():
+    """Verify HuggingFaceNerRecognizer can be instantiated without model_name during registry build."""
+    with patch(HF_PIPELINE_PATH, new=MagicMock()):
+        recognizer = HuggingFaceNerRecognizer()
+        assert recognizer.model_name is None
+        assert recognizer.ner_pipeline is None
+

presidio-analyzer/tests/test_recognizers_loader_utils.py

@@ -607,15 +607,11 @@ def test_yaml_country_code_blank_value_raises():
 LOADER_KWARGS = ("name", "supported_language")
 
 # Entries that cannot load from their shipped configuration even with every
-# dependency installed, so the load test below cannot cover them.
-#
-# ``HuggingFaceNerRecognizer``: ``EntityRecognizer.__init__`` calls ``load()``
-# unconditionally and ``load()`` requires ``model_name``, which the shipped
-# entry does not supply -- it raises ValueError once ``transformers`` and
-# ``torch`` are present. That is a pre-existing defect in the entry, not
-# something this contract can assert away, and adding ``model_name`` here would
-# make the test download a model. It stays covered by the resolve test.
-NOT_LOADABLE_FROM_SHIPPED_ENTRY = {"HuggingFaceNerRecognizer"}
+# dependency installed, so the load test below cannot cover them. Empty today:
+# ``HuggingFaceNerRecognizer`` used to sit here because ``load()`` raised when the
+# shipped entry supplied no ``model_name``. It now defers loading instead, so the
+# load test covers it like every other entry.
+NOT_LOADABLE_FROM_SHIPPED_ENTRY: set[str] = set()
 
 # Entries gated behind an optional dependency, for which refusing to load with an
 # actionable ImportError is the intended behavior. The skip is scoped to these