diff --git a/nemoguardrails/library/hf_classifier/flows.co b/nemoguardrails/library/hf_classifier/flows.co
index 34cc4dcd78..3d1f2de8b4 100644
--- a/nemoguardrails/library/hf_classifier/flows.co
+++ b/nemoguardrails/library/hf_classifier/flows.co
@@ -22,4 +22,6 @@ flow hf classifier check retrieval $classifier
   $response = await HfClassifierCheckRetrievalAction(classifier=$classifier)
 
   if $response.is_transform
+    # FIX: Added global declaration to propagate the change to calling context
+    global $relevant_chunks
     $relevant_chunks = $response.transform_text["relevant_chunks"]
diff --git a/nemoguardrails/library/hf_classifier/flows.v1.co b/nemoguardrails/library/hf_classifier/flows.v1.co
index 015ca8530b..91052a11e7 100644
--- a/nemoguardrails/library/hf_classifier/flows.v1.co
+++ b/nemoguardrails/library/hf_classifier/flows.v1.co
@@ -25,4 +25,6 @@ define subflow hf classifier check retrieval
   $response = execute hf_classifier_check_retrieval(classifier=$classifier)
 
   if $response.is_transform
-    $relevant_chunks = $response.transform_text["relevant_chunks"]
+    # In Colang 1.0, subflows share global scope.
+    # The caller must initialize $relevant_chunks before calling this subflow.
+    $relevant_chunks = $response.transform_text["relevant_chunks"]
\ No newline at end of file
diff --git a/tests/test_hf_classifier_retrieval_global.py b/tests/test_hf_classifier_retrieval_global.py
new file mode 100644
index 0000000000..3ce742bf5f
--- /dev/null
+++ b/tests/test_hf_classifier_retrieval_global.py
@@ -0,0 +1,195 @@
+import unittest
+from unittest.mock import AsyncMock
+
+from nemoguardrails import LLMRails, RailsConfig
+
+
+class MockHfClassifierResponse:
+    def __init__(self, is_transform=False, transform_text=None, is_blocked=False):
+        self.is_transform = is_transform
+        self.transform_text = transform_text or {}
+        self.is_blocked = is_blocked
+
+
+class TestHfClassifierRetrievalGlobal(unittest.TestCase):
+
+    def test_v1_retrieval_transform_propagates_global(self):
+        mock_action = AsyncMock()
+        mock_action.return_value = MockHfClassifierResponse(
+            is_transform=True,
+            transform_text={
+                "relevant_chunks": ["new_chunk_1", "new_chunk_2"],
+                "response": "Transformed response"
+            }
+        )
+
+        config = RailsConfig.from_content(
+            colang_content="""
+            define user express greeting
+                "Hello"
+
+            define flow test_retrieval
+                $user_input = "test query"
+                $classifier = "test_classifier"
+                $relevant_chunks = execute hf classifier check retrieval
+                $test_output = $relevant_chunks
+            """,
+            yaml_content="""
+            models:
+              - type: main
+                engine: openai
+                model: gpt-3.5-turbo
+            """
+        )
+
+        app = LLMRails(config)
+        app.register_action("hf_classifier_check_retrieval", mock_action)
+        app.generate(messages=[{"role": "user", "content": "Hello"}])
+
+        self.assertEqual(app.context.get("relevant_chunks"), ["new_chunk_1", "new_chunk_2"])
+        self.assertEqual(app.context.get("test_output"), ["new_chunk_1", "new_chunk_2"])
+
+    def test_original_retrieval_transform_propagates_global(self):
+        mock_action = AsyncMock()
+        mock_action.return_value = MockHfClassifierResponse(
+            is_transform=True,
+            transform_text={
+                "relevant_chunks": ["original_chunk_1", "original_chunk_2"],
+                "response": "Original transformed response"
+            }
+        )
+
+        config = RailsConfig.from_content(
+            colang_content="""
+            define user express greeting
+                "Hello"
+
+            define flow test_retrieval
+                $user_input = "test query"
+                $classifier = "test_classifier"
+                execute hf classifier check retrieval $classifier
+                $test_output = $relevant_chunks
+            """,
+            yaml_content="""
+            models:
+              - type: main
+                engine: openai
+                model: gpt-3.5-turbo
+            """
+        )
+
+        app = LLMRails(config)
+        app.register_action("HfClassifierCheckRetrievalAction", mock_action)
+        app.generate(messages=[{"role": "user", "content": "Hello"}])
+
+        self.assertEqual(app.context.get("relevant_chunks"), ["original_chunk_1", "original_chunk_2"])
+        self.assertEqual(app.context.get("test_output"), ["original_chunk_1", "original_chunk_2"])
+
+    def test_no_transform_doesnt_modify_chunks(self):
+        mock_action = AsyncMock()
+        mock_action.return_value = MockHfClassifierResponse(
+            is_transform=False,
+            is_blocked=False
+        )
+
+        config = RailsConfig.from_content(
+            colang_content="""
+            define user express greeting
+                "Hello"
+
+            define flow test_no_transform
+                $user_input = "test query"
+                $original_chunks = ["original_chunk_1", "original_chunk_2"]
+                $relevant_chunks = $original_chunks
+                $classifier = "test_classifier"
+                execute hf classifier check retrieval
+                $test_output = $relevant_chunks
+            """,
+            yaml_content="""
+            models:
+              - type: main
+                engine: openai
+                model: gpt-3.5-turbo
+            """
+        )
+
+        app = LLMRails(config)
+        app.register_action("hf_classifier_check_retrieval", mock_action)
+        app.generate(messages=[{"role": "user", "content": "Hello"}])
+
+        self.assertEqual(app.context.get("relevant_chunks"), ["original_chunk_1", "original_chunk_2"])
+        self.assertEqual(app.context.get("test_output"), ["original_chunk_1", "original_chunk_2"])
+
+    def test_v1_retrieval_with_empty_transform_text(self):
+        mock_action = AsyncMock()
+        mock_action.return_value = MockHfClassifierResponse(
+            is_transform=True,
+            transform_text={"relevant_chunks": []}
+        )
+
+        config = RailsConfig.from_content(
+            colang_content="""
+            define user express greeting
+                "Hello"
+
+            define flow test_empty_transform
+                $user_input = "test query"
+                $original_chunks = ["original"]
+                $relevant_chunks = $original_chunks
+                $classifier = "test_classifier"
+                $relevant_chunks = execute hf classifier check retrieval
+                $test_output = $relevant_chunks
+            """,
+            yaml_content="""
+            models:
+              - type: main
+                engine: openai
+                model: gpt-3.5-turbo
+            """
+        )
+
+        app = LLMRails(config)
+        app.register_action("hf_classifier_check_retrieval", mock_action)
+        app.generate(messages=[{"role": "user", "content": "Hello"}])
+
+        self.assertEqual(app.context.get("relevant_chunks"), [])
+        self.assertEqual(app.context.get("test_output"), [])
+
+    def test_v1_retrieval_blocked_flow(self):
+        mock_action = AsyncMock()
+        mock_action.return_value = MockHfClassifierResponse(
+            is_transform=False,
+            is_blocked=True
+        )
+
+        config = RailsConfig.from_content(
+            colang_content="""
+            define user express greeting
+                "Hello"
+
+            define flow test_blocked
+                $user_input = "test query"
+                $original_chunks = ["original"]
+                $relevant_chunks = $original_chunks
+                $classifier = "test_classifier"
+                $config.enable_rails_exceptions = True
+                execute hf classifier check retrieval
+                $test_output = $relevant_chunks
+            """,
+            yaml_content="""
+            models:
+              - type: main
+                engine: openai
+                model: gpt-3.5-turbo
+            """
+        )
+
+        app = LLMRails(config)
+        app.register_action("hf_classifier_check_retrieval", mock_action)
+        app.generate(messages=[{"role": "user", "content": "Hello"}])
+
+        self.assertEqual(app.context.get("relevant_chunks"), ["original"])
+
+
+if __name__ == "__main__":
+    unittest.main()
\ No newline at end of file