[FIX] Preserve configured LLM in non-batch fallback extractors

lexical-graph/src/graphrag_toolkit/lexical_graph/indexing/extract/batch_llm_proposition_extractor_sync.py

@@ -67,6 +67,7 @@ def _run_non_batch_extractor(self, nodes):
         all_nodes = [node for node in nodes]
 
         extractor = LLMPropositionExtractor(
+            llm=self.llm,
             prompt_template=self.prompt_template, 
             source_metadata_field=self.source_metadata_field
         )

lexical-graph/src/graphrag_toolkit/lexical_graph/indexing/extract/batch_topic_extractor_sync.py

@@ -80,6 +80,7 @@ def _run_non_batch_extractor(self, nodes):
         all_nodes = [node for node in nodes]
 
         extractor = TopicExtractor( 
+            llm=self.llm,
             prompt_template=self.prompt_template, 
             source_metadata_field=self.source_metadata_field,
             entity_classification_provider=self.entity_classification_provider,

lexical-graph/tests/unit/indexing/extract/test_batch_llm_proposition_extractor_sync.py

@@ -70,3 +70,44 @@ def test_update_node_with_missing_node_id(self):
         result = extractor._update_node(node, node_metadata_map)
 
         assert result.metadata[PROPOSITIONS_KEY] == []
+
+
+class TestBatchLLMPropositionExtractorSyncRunNonBatchExtractor:
+    """Tests for _run_non_batch_extractor method.
+
+    Regression tests: when a node set falls below Bedrock's minimum batch
+    size, BatchLLMPropositionExtractorSync falls back to the non-batch
+    LLMPropositionExtractor. That fallback must reuse the configured LLM
+    (self.llm) rather than silently defaulting to
+    GraphRAGConfig.extraction_llm, which is a us.* inference profile and
+    is invalid in any other Bedrock region.
+    """
+
+    def _make_extractor(self, llm, prompt_template="prompt", source_metadata_field=None):
+        """Create a BatchLLMPropositionExtractorSync instance with fields
+        populated via model_construct, bypassing validation/__init__ so no
+        real BatchConfig or AWS setup is needed for this unit test."""
+        return BatchLLMPropositionExtractorSync.model_construct(
+            llm=llm,
+            prompt_template=prompt_template,
+            source_metadata_field=source_metadata_field,
+        )
+
+    @patch("graphrag_toolkit.lexical_graph.indexing.extract.batch_llm_proposition_extractor_sync.LLMPropositionExtractor")
+    def test_run_non_batch_extractor_passes_configured_llm(self, mock_extractor_cls):
+        """Verify the non-batch fallback is constructed with the extractor's
+        own configured llm, not left to default to GraphRAGConfig.extraction_llm."""
+        configured_llm = Mock(name="configured-llm")
+        extractor = self._make_extractor(llm=configured_llm)
+
+        mock_instance = mock_extractor_cls.return_value
+        mock_instance.extract.return_value = [{PROPOSITIONS_KEY: []}]
+
+        nodes = [TextNode(text="test", id_="node-1")]
+        extractor._run_non_batch_extractor(nodes)
+
+        mock_extractor_cls.assert_called_once_with(
+            llm=configured_llm,
+            prompt_template="prompt",
+            source_metadata_field=None,
+        )

lexical-graph/tests/unit/indexing/extract/test_batch_topic_extractor_sync.py

@@ -68,3 +68,54 @@ def test_update_node_with_missing_node_id(self):
         result = extractor._update_node(node, node_metadata_map)
 
         assert result.metadata[TOPICS_KEY] == {'topics': []}
+
+
+class TestBatchTopicExtractorSyncRunNonBatchExtractor:
+    """Tests for _run_non_batch_extractor method.
+
+    Regression tests: when a node set falls below Bedrock's minimum batch
+    size, BatchTopicExtractorSync falls back to the non-batch
+    TopicExtractor. That fallback must reuse the configured LLM (self.llm)
+    rather than silently defaulting to GraphRAGConfig.extraction_llm, which
+    is a us.* inference profile and is invalid in any other Bedrock region.
+    """
+
+    def _make_extractor(self, llm, prompt_template="prompt", source_metadata_field=None,
+                         entity_classification_provider=None, topic_provider=None):
+        """Create a BatchTopicExtractorSync instance with fields populated
+        via model_construct, bypassing validation/__init__ so no real
+        BatchConfig or AWS setup is needed for this unit test."""
+        return BatchTopicExtractorSync.model_construct(
+            llm=llm,
+            prompt_template=prompt_template,
+            source_metadata_field=source_metadata_field,
+            entity_classification_provider=entity_classification_provider or Mock(),
+            topic_provider=topic_provider or Mock(),
+        )
+
+    @patch("graphrag_toolkit.lexical_graph.indexing.extract.batch_topic_extractor_sync.TopicExtractor")
+    def test_run_non_batch_extractor_passes_configured_llm(self, mock_extractor_cls):
+        """Verify the non-batch fallback is constructed with the extractor's
+        own configured llm, not left to default to GraphRAGConfig.extraction_llm."""
+        configured_llm = Mock(name="configured-llm")
+        entity_classification_provider = Mock(name="entity-classification-provider")
+        topic_provider = Mock(name="topic-provider")
+        extractor = self._make_extractor(
+            llm=configured_llm,
+            entity_classification_provider=entity_classification_provider,
+            topic_provider=topic_provider,
+        )
+
+        mock_instance = mock_extractor_cls.return_value
+        mock_instance.extract.return_value = [{TOPICS_KEY: {'topics': []}}]
+
+        nodes = [TextNode(text="test", id_="node-1")]
+        extractor._run_non_batch_extractor(nodes)
+
+        mock_extractor_cls.assert_called_once_with(
+            llm=configured_llm,
+            prompt_template="prompt",
+            source_metadata_field=None,
+            entity_classification_provider=entity_classification_provider,
+            topic_provider=topic_provider,
+        )