Merge pull request #1042 from Accute9/spacy-cache-split-chunking

perf(split): avoid repeated spaCy model loading in split/chunking paths

CHANGELOG.md

@@ -73,6 +73,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
+- **`split`/chunking paths bypassed the centralized spaCy model cache, reloading the model on every call** (#1042, closes #998) by @Accute9, reviewed by @Sameer6305
+  - `semantica/split/methods.py`'s `split_by_sentences()` and `semantica/split/semantic_chunker.py`'s `SemanticChunker.__init__` each called `spacy.load()` directly instead of reusing the process-level cache added in #889/`semantic_extract/methods.py`'s `load_spacy_model()` — every call/construction re-paid the ~120ms model-load cost independently of `NERExtractor`, which already used the cache
+  - Both now route through `load_spacy_model()`, sharing one cached `Language` instance per model name across `split_by_sentences()`, `SemanticChunker`, and `NERExtractor`; a missing model still falls back to regex/paragraph chunking without poisoning the cache for a later successful load
+  - **Fixed during review** (@Sameer6305): `NERExtractor.__init__()` still had a direct `spacy.load()` call site with the same cache-bypass issue, outside the two files named in #998 but sharing the same root cause; routed through the cache alongside stale test patch targets and a strengthened cache-configuration assertion
+  - **Fixed during review** (@KaifAhmad1): `SemanticChunker.__init__` only caught `OSError` around `load_spacy_model()`, while the sibling fix to `NERExtractor` in this same PR added a broader `except Exception` for a model that is installed but fails at runtime (e.g. a config incompatible with the installed spaCy version). A broken-but-present model crashed `SemanticChunker()` outright instead of degrading to fallback chunking like every other path in this PR. Added the matching `except Exception` branch, leaving `self.nlp` as `None`; new `test_semantic_chunker_falls_back_when_spacy_runtime_is_broken` mirrors the existing `NERExtractor` regression test for the same scenario
+  - New `tests/split/test_spacy_model_cache.py`: cache reuse across repeated calls/instances, shared cache between `split_by_sentences()`/`SemanticChunker`/`NERExtractor`, distinct model names loading separately, missing-model fallback without poisoning the cache, and the broken-runtime fallback added above
+  - `pytest tests/split/test_spacy_model_cache.py tests/split/test_splitter.py tests/split/test_chunkers.py`: all passing (3 pre-existing, unrelated `tests/test_ner_configurations.py` failures confirmed present on `main` before this PR)
+
 - **`export_yaml` raised a raw `AttributeError` on list input, silently wrote empty exports for unrecognized dict keys, and graph payloads were reconciled differently by every exporter** (#958, closes #956, #952, #953) by @pravit-amp, reviewed by @Sameer6305
   - Graph payloads circulate under two vocabularies, `entities`/`relationships` and `nodes`/`edges`, and each exporter reconciled them locally with a different idiom — `LPGExporter` in particular dropped every entity whenever `nodes` was present but empty, the exact shape `JSONExporter` emits. A new `normalize_graph_payload()` in `utils/helpers.py` centralizes that decision once, adopted by `LPGExporter`, `ArangoAQLExporter`, `Neo4jCSVExporter`, and both YAML exporters; `ContextGraph.to_dict()` now round-trips through YAML correctly as a result
   - `export_yaml(records, path)` on a bare list previously failed with `AttributeError` from inside the exporter; it and the other YAML methods now reject non-mapping input with an actionable `ProcessingError` naming the expected keys, since these formats distinguish entities/relationships/triplets and guessing which one a list represents would mislabel the records

semantica/semantic_extract/ner_extractor.py

@@ -144,7 +144,12 @@ class NERExtractor:
         self._ml_runtime_usable = True
         if "ml" in self.method and SPACY_AVAILABLE:
             try:
-                self.nlp = spacy.load(self.model_name)
+                # Deferred import: keeps semantic_extract.methods out of the
+                # module-level import graph and routes loading through the
+                # process-level cache so repeated NERExtractor constructions
+                # never pay the ~120 ms spacy.load() cost more than once.
+                from .methods import load_spacy_model
+                self.nlp = load_spacy_model(self.model_name)
             except OSError:
                 self.logger.warning(
                     f"spaCy model {self.model_name} not found. ML method will fallback."

semantica/split/methods.py

@@ -97,7 +97,7 @@ from .semantic_chunker import Chunk
 logger = get_logger("split_methods")
 
 # Try to import optional dependencies
-spacy, SPACY_AVAILABLE = safe_import("spacy")
+_, SPACY_AVAILABLE = safe_import("spacy")
 
 nltk, NLTK_AVAILABLE = safe_import("nltk")
 tiktoken, TIKTOKEN_AVAILABLE = safe_import("tiktoken")
@@ -336,7 +336,8 @@ def split_by_sentences(
     # Try spaCy first
     if SPACY_AVAILABLE and kwargs.get("use_spacy", True):
         try:
-            nlp = spacy.load("en_core_web_sm")
+            from ..semantic_extract.methods import load_spacy_model
+            nlp = load_spacy_model("en_core_web_sm")
             doc = nlp(text)
             sentences = [sent.text for sent in doc.sents]
         except Exception:

semantica/split/semantic_chunker.py

@@ -36,7 +36,8 @@ from ..utils.helpers import safe_import
 from ..utils.logging import get_logger
 from ..utils.progress_tracker import get_progress_tracker
 
-spacy, SPACY_AVAILABLE = safe_import("spacy")
+
+_, SPACY_AVAILABLE = safe_import("spacy")
 
 
 @dataclass
@@ -79,11 +80,19 @@ class SemanticChunker:
         if SPACY_AVAILABLE:
             model_name = config.get("model", "en_core_web_sm")
             try:
-                self.nlp = spacy.load(model_name)
+                from ..semantic_extract.methods import load_spacy_model
+                self.nlp = load_spacy_model(model_name)
             except OSError:
                 self.logger.warning(
                     f"spaCy model {model_name} not found. Using fallback chunking."
                 )
+            except Exception:
+                self.logger.warning(
+                    "spaCy model %s failed to initialize and will be disabled "
+                    "for this chunker instance. Using fallback chunking.",
+                    model_name,
+                    exc_info=True,
+                )
 
     def chunk(self, text: str, **options) -> List[Chunk]:
         """

tests/split/test_spacy_model_cache.py

@@ -0,0 +1,321 @@
+from types import SimpleNamespace
+from unittest.mock import MagicMock
+
+import pytest
+
+from semantica.semantic_extract import methods as se_methods
+from semantica.split import methods as split_methods
+from semantica.split import semantic_chunker
+from semantica.semantic_extract import ner_extractor as ner_extractor_module
+from semantica.semantic_extract.ner_extractor import NERExtractor
+
+
+@pytest.fixture(autouse=True)
+def clear_cache():
+    se_methods.clear_spacy_model_cache()
+    yield
+    se_methods.clear_spacy_model_cache()
+
+
+@pytest.fixture(autouse=True)
+def force_spacy_available(monkeypatch):
+    # split.methods, split.semantic_chunker, and ner_extractor each compute
+    # their own SPACY_AVAILABLE flag from the real environment at import time;
+    # force all true so these tests exercise the spaCy branch regardless of
+    # whether spaCy is actually installed where they run.
+    monkeypatch.setattr(split_methods, "SPACY_AVAILABLE", True)
+    monkeypatch.setattr(semantic_chunker, "SPACY_AVAILABLE", True)
+    monkeypatch.setattr(ner_extractor_module, "SPACY_AVAILABLE", True)
+
+
+def _fake_spacy(load):
+    return SimpleNamespace(
+        load=load,
+        util=SimpleNamespace(is_package=lambda _name: True),
+    )
+
+
+def _nlp_mock(sentences=("Hello world.",)):
+    """A stand-in spaCy Language object: callable, returns a doc with .sents."""
+    nlp = MagicMock()
+    nlp.return_value = SimpleNamespace(
+        sents=[SimpleNamespace(text=s) for s in sentences]
+    )
+    return nlp
+
+
+class TestSpacyModelCache:
+    """split.methods and split.semantic_chunker must share the cached model
+    defined in semantic_extract.methods instead of each calling spacy.load()
+    independently.
+    """
+
+    def test_split_by_sentences_reuses_cached_model(self, monkeypatch):
+        calls = []
+
+        def fake_load(name, **kwargs):
+            calls.append((name, kwargs))
+            return _nlp_mock()
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(fake_load))
+
+        split_methods.split_by_sentences("Hello world. Bye world.")
+        split_methods.split_by_sentences("Another sentence here.")
+        split_methods.split_by_sentences("A third call.")
+
+        assert len(calls) == 1, "spacy.load should run once, not once per call"
+        assert calls[0][0] == "en_core_web_sm"
+
+    def test_semantic_chunker_reuses_cached_model_across_instances(self, monkeypatch):
+        calls = []
+
+        def fake_load(name, **kwargs):
+            calls.append((name, kwargs))
+            return _nlp_mock()
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(fake_load))
+
+        chunker1 = semantic_chunker.SemanticChunker()
+        chunker2 = semantic_chunker.SemanticChunker()
+
+        assert len(calls) == 1, "each new SemanticChunker should not reload the model"
+        assert chunker1.nlp is chunker2.nlp
+
+    def test_split_methods_and_semantic_chunker_share_the_cache(self, monkeypatch):
+        calls = []
+
+        def fake_load(name, **kwargs):
+            calls.append((name, kwargs))
+            return _nlp_mock()
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(fake_load))
+
+        split_methods.split_by_sentences("Test sentence for split.methods.")
+        semantic_chunker.SemanticChunker()
+
+        assert len(calls) == 1, (
+            "split.methods and split.semantic_chunker must share one cached "
+            "model instead of each loading their own"
+        )
+
+    def test_distinct_model_names_load_separately(self, monkeypatch):
+        calls = []
+
+        def fake_load(name, **kwargs):
+            calls.append((name, kwargs))
+            return _nlp_mock()
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(fake_load))
+
+        sm_chunker = semantic_chunker.SemanticChunker(model="en_core_web_sm")
+        lg_chunker = semantic_chunker.SemanticChunker(model="en_core_web_lg")
+        sm_chunker_again = semantic_chunker.SemanticChunker(model="en_core_web_sm")
+
+        assert [name for name, _ in calls] == ["en_core_web_sm", "en_core_web_lg"]
+        assert sm_chunker.nlp is sm_chunker_again.nlp
+        assert sm_chunker.nlp is not lg_chunker.nlp
+
+    def test_no_disable_kwarg_requested(self, monkeypatch):
+        """split.methods and split.semantic_chunker both want the full
+        pipeline (they need .sents, which requires the parser/senter). If
+        either one later starts requesting a trimmed pipeline (e.g.
+        disable=["ner"]), the name-only cache key in load_spacy_model would
+        silently hand back a cached model built for a different config --
+        this test should catch that the moment it happens.
+        """
+        calls = []
+
+        def fake_load(_name, **kwargs):
+            calls.append(kwargs)
+            return _nlp_mock()
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(fake_load))
+
+        split_methods.split_by_sentences("Hello world.")
+        se_methods.clear_spacy_model_cache()
+        semantic_chunker.SemanticChunker()
+
+        assert len(calls) == 2
+        assert all(kwargs == {} for kwargs in calls), (
+            "neither caller should pass any pipeline-configuration kwargs; "
+            "the name-only cache key in load_spacy_model cannot distinguish "
+            "models loaded with different component configs"
+        )
+
+    def test_missing_model_falls_back_without_poisoning_cache(self, monkeypatch):
+        attempts = []
+
+        def failing_load(name, **_kwargs):
+            attempts.append(name)
+            raise OSError(f"Can't find model '{name}'")
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(failing_load))
+
+        # split_by_sentences should fall back to regex splitting, not raise
+        chunks = split_methods.split_by_sentences("Hello world. Bye world.")
+        assert chunks, "fallback splitting should still produce chunks"
+
+        # SemanticChunker should leave .nlp as None rather than propagate
+        chunker = semantic_chunker.SemanticChunker()
+        assert chunker.nlp is None
+
+        assert len(attempts) == 2, "a failed load must not be cached"
+
+        # Once the model is available, both callers should now get it, and
+        # share a single successful load.
+        def working_load(name, **_kwargs):
+            attempts.append(name)
+            return _nlp_mock()
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(working_load))
+
+        chunker2 = semantic_chunker.SemanticChunker()
+        split_methods.split_by_sentences("One more sentence.")
+
+        assert len(attempts) == 3, (
+            "the model should load once after it becomes available"
+        )
+        assert chunker2.nlp is not None
+
+    def test_semantic_chunker_falls_back_when_spacy_runtime_is_broken(
+        self, monkeypatch
+    ):
+        """A spaCy model that is installed but unusable at runtime (e.g. a
+        config incompatible with the installed spaCy version) must degrade
+        SemanticChunker to fallback chunking, not crash __init__ -- mirrors
+        TestNERExtractorSpacyModelCache's equivalent broken-runtime test.
+        """
+
+        def broken_load(name, **_kwargs):
+            raise RuntimeError("ConfigSchemaNlp is not fully defined")
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(broken_load))
+
+        chunker = semantic_chunker.SemanticChunker()
+
+        assert chunker.nlp is None
+
+
+class TestNERExtractorSpacyModelCache:
+    """NERExtractor(method="ml") must reuse the centralized cache in
+    semantic_extract.methods, not call spacy.load() on every construction.
+
+    These tests mirror TestSpacyModelCache but focus on the NERExtractor path,
+    confirming that all three callers (split_by_sentences, SemanticChunker, and
+    NERExtractor) draw from the same process-level cache.
+    """
+
+    def test_ner_extractor_reuses_cached_model_across_instances(self, monkeypatch):
+        """Two NERExtractor(method='ml') constructions with the same model name
+        must cause exactly one underlying spacy.load() call."""
+        calls = []
+
+        def fake_load(name, **kwargs):
+            calls.append(name)
+            return _nlp_mock()
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(fake_load))
+
+        e1 = NERExtractor(method="ml")
+        e2 = NERExtractor(method="ml")
+        e3 = NERExtractor(method="ml", model="en_core_web_sm")
+
+        assert len(calls) == 1, (
+            "repeated NERExtractor constructions should not reload the model"
+        )
+        assert e1.nlp is e2.nlp is e3.nlp
+
+    def test_ner_extractor_and_split_callers_share_one_cached_model(self, monkeypatch):
+        """NERExtractor, SemanticChunker, and split_by_sentences must all use
+        the same cached Language object for the same model name."""
+        calls = []
+
+        def fake_load(name, **kwargs):
+            calls.append(name)
+            return _nlp_mock()
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(fake_load))
+
+        split_methods.split_by_sentences("First sentence.")
+        semantic_chunker.SemanticChunker()
+        NERExtractor(method="ml")
+
+        assert len(calls) == 1, (
+            "split_by_sentences, SemanticChunker, and NERExtractor must share "
+            "one cached model instead of each loading their own"
+        )
+
+    def test_ner_extractor_distinct_model_names_load_separately(self, monkeypatch):
+        """Different model names must produce separate cache entries."""
+        calls = []
+
+        def fake_load(name, **kwargs):
+            calls.append(name)
+            return _nlp_mock()
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(fake_load))
+
+        sm = NERExtractor(method="ml", model="en_core_web_sm")
+        lg = NERExtractor(method="ml", model="en_core_web_lg")
+        sm_again = NERExtractor(method="ml", model="en_core_web_sm")
+
+        assert calls == ["en_core_web_sm", "en_core_web_lg"]
+        assert sm.nlp is sm_again.nlp
+        assert sm.nlp is not lg.nlp
+
+    def test_ner_extractor_failed_load_not_cached_and_retried(self, monkeypatch):
+        """A missing model must not poison the cache.  A subsequent construction
+        after the model becomes available must succeed and share the loaded model."""
+        attempts = []
+
+        def failing_load(name, **_kwargs):
+            attempts.append(name)
+            raise OSError(f"Can't find model '{name}'")
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(failing_load))
+
+        # Construction with missing model: nlp must remain None, no crash
+        extractor1 = NERExtractor(method="ml")
+        assert extractor1.nlp is None
+        assert len(attempts) == 1, "one load attempt expected for the missing model"
+
+        # Second construction: must retry (cache must not hold the failure)
+        extractor2 = NERExtractor(method="ml")
+        assert extractor2.nlp is None
+        assert len(attempts) == 2, "a failed load must not be cached"
+
+        # Now install a working model and verify recovery
+        def working_load(name, **_kwargs):
+            attempts.append(name)
+            return _nlp_mock()
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(working_load))
+
+        extractor3 = NERExtractor(method="ml")
+        extractor4 = NERExtractor(method="ml")
+
+        assert extractor3.nlp is not None
+        assert extractor3.nlp is extractor4.nlp
+        assert len(attempts) == 3, (
+            "exactly one successful load expected after the model becomes available"
+        )
+
+    def test_ner_extractor_non_ml_method_does_not_load_model(self, monkeypatch):
+        """NERExtractor with a non-ml method must not touch the spaCy cache."""
+        calls = []
+
+        def fake_load(name, **kwargs):
+            calls.append(name)
+            return _nlp_mock()
+
+        monkeypatch.setattr(se_methods, "spacy", _fake_spacy(fake_load))
+
+        NERExtractor(method="pattern")
+        NERExtractor(method="llm")
+        NERExtractor(method="regex")
+
+        assert calls == [], "non-ml methods must not trigger any spacy.load()"
+
+
+if __name__ == "__main__":
+    pytest.main([__file__])

tests/split/test_splitter.py

@@ -30,18 +30,16 @@ class TestSplitter(unittest.TestCase):
         splitter = TextSplitter(method=["recursive", "token"])
         self.assertEqual(splitter.methods, ["recursive", "token"])
 
-    @patch('semantica.split.semantic_chunker.spacy')
+    @patch('semantica.semantic_extract.methods.spacy')
     def test_semantic_chunker_initialization(self, mock_spacy):
-        # Mock spacy.load to return a mock nlp object
+        # SemanticChunker now loads spaCy through the centralized
+        # load_spacy_model() in semantic_extract.methods, so we patch
+        # methods.spacy rather than the removed semantic_chunker.spacy binding.
         mock_nlp = MagicMock()
         mock_spacy.load.return_value = mock_nlp
-        
-        # We need to ensure SPACY_AVAILABLE is True for this test context if possible, 
-        # but it is imported at module level. 
-        # If spacy is not installed, it sets SPACY_AVAILABLE = False.
-        # We might need to patch the module attribute or just test fallback if spacy missing.
-        
-        chunker = SemanticChunker(chunk_size=100)
+
+        with patch('semantica.split.semantic_chunker.SPACY_AVAILABLE', True):
+            chunker = SemanticChunker(chunk_size=100)
         self.assertEqual(chunker.chunk_size, 100)
 
     def test_chunk_dataclass(self):

tests/test_ner_configurations.py

@@ -101,9 +101,15 @@ class TestNERConfigurations(unittest.TestCase):
             self.assertEqual(entities[0].metadata["extraction_method"], "ml")
             self.assertEqual(entities[0].metadata["model"], "en_core_web_trf")
 
-    @patch('semantica.semantic_extract.ner_extractor.spacy')
+    @patch('semantica.semantic_extract.methods.spacy')
     def test_ner_ml_init_falls_back_when_spacy_runtime_is_broken(self, mock_spacy):
-        """Test NER init does not crash when spaCy is installed but unusable at runtime."""
+        """Test NER init does not crash when spaCy is installed but unusable at runtime.
+
+        The model load now goes through load_spacy_model() in semantic_extract.methods,
+        so we patch methods.spacy (not ner_extractor.spacy) to inject the failure.
+        """
+        from semantica.semantic_extract.methods import clear_spacy_model_cache
+        clear_spacy_model_cache()
         mock_spacy.load.side_effect = RuntimeError("ConfigSchemaNlp is not fully defined")
 
         with patch('semantica.semantic_extract.ner_extractor.SPACY_AVAILABLE', True):
@@ -112,17 +118,23 @@ class TestNERConfigurations(unittest.TestCase):
         self.assertIsNone(extractor.nlp)
         self.assertFalse(extractor._ml_runtime_usable)
 
-    @patch('semantica.semantic_extract.ner_extractor.spacy')
     @patch('semantica.semantic_extract.methods.get_entity_method')
     @patch('semantica.semantic_extract.methods.spacy')
     def test_ner_ml_runtime_failure_disables_repeated_ml_load_attempts(
         self,
         mock_methods_spacy,
         mock_get_method,
-        mock_init_spacy,
     ):
-        """Test degraded ML mode skips repeated spaCy load attempts after init failure."""
-        mock_init_spacy.load.side_effect = RuntimeError("ConfigSchemaNlp is not fully defined")
+        """Test degraded ML mode skips repeated spaCy load attempts after init failure.
+
+        The model load at construction time now goes through load_spacy_model() in
+        semantic_extract.methods, so methods.spacy is the single mock target for the
+        init-time failure.  After the RuntimeError is raised, _ml_runtime_usable is
+        False and no further spacy.load (or extract_entities_ml) calls are made.
+        """
+        from semantica.semantic_extract.methods import clear_spacy_model_cache
+        clear_spacy_model_cache()
+        mock_methods_spacy.load.side_effect = RuntimeError("ConfigSchemaNlp is not fully defined")
         mock_ml_method = MagicMock(return_value=[])
         mock_get_method.side_effect = lambda name: mock_ml_method if name == "ml" else (lambda *_args, **_kwargs: [])
 
@@ -132,8 +144,9 @@ class TestNERConfigurations(unittest.TestCase):
         entities = extractor.extract_entities(self.text)
 
         self.assertFalse(extractor._ml_runtime_usable)
-        self.assertEqual(mock_init_spacy.load.call_count, 1)
-        self.assertEqual(mock_methods_spacy.load.call_count, 0)
+        # methods.spacy.load called once during __init__ (the RuntimeError); not again
+        # during extract_entities because _filter_unusable_methods removes "ml".
+        self.assertEqual(mock_methods_spacy.load.call_count, 1)
         self.assertEqual(mock_ml_method.call_count, 0)
         self.assertIsInstance(entities, list)