feat(langchain): stream metadata in RunnableRails output

Enhance streaming in RunnableRails to include generation metadata in
streamed chunks. Skips END_OF_STREAM markers and updates chunk formatting
to support metadata for AIMessageChunk outputs. This improves compatibility
with consumers expecting metadata in streaming responses.

Author: Pouyanpi

nemoguardrails/integrations/langchain/runnable_rails.py

@@ -221,6 +221,7 @@ def _create_passthrough_messages(self, _input) -> List[Dict[str, Any]]:
                 "role": "context",
                 "content": {
                     "passthrough_input": _input,
+                    # We also set all the input variables as top level context variables
                     **(_input if isinstance(_input, dict) else {}),
                 },
             },
@@ -841,55 +842,87 @@ async def astream(
             streaming_enabled = True
 
         try:
-            async for chunk in self.rails.stream_async(messages=input_messages):
+            from nemoguardrails.streaming import END_OF_STREAM
+
+            async for chunk in self.rails.stream_async(
+                messages=input_messages, include_generation_metadata=True
+            ):
+                # Skip END_OF_STREAM markers
+                chunk_text = (
+                    chunk["text"]
+                    if isinstance(chunk, dict) and "text" in chunk
+                    else chunk
+                )
+                if chunk_text is END_OF_STREAM:
+                    continue
+
                 # Format the chunk based on the input type for streaming
                 formatted_chunk = self._format_streaming_chunk(input, chunk)
                 yield formatted_chunk
         finally:
             if streaming_enabled and hasattr(self.rails.llm, "streaming"):
                 self.rails.llm.streaming = original_streaming
 
-    def _format_streaming_chunk(self, input: Any, chunk: str) -> Any:
+    def _format_streaming_chunk(self, input: Any, chunk) -> Any:
         """Format a streaming chunk based on the input type.
 
         Args:
             input: The original input
-            chunk: The current text chunk
+            chunk: The current chunk (string or dict with text/generation_info)
 
         Returns:
             The formatted streaming chunk (using AIMessageChunk for LangChain compatibility)
         """
+        # Extract text and metadata from chunk if it's a dict with generation metadata
+        text_content = chunk
+        metadata = {}
+
+        if isinstance(chunk, dict) and "text" in chunk:
+            text_content = chunk["text"]
+            generation_info = chunk.get("generation_info", {})
+
+            # Use generation_info as metadata for streaming chunks
+            if generation_info:
+                metadata = generation_info.copy()
         if isinstance(input, ChatPromptValue):
-            return AIMessageChunk(content=chunk)
+            return AIMessageChunk(content=text_content, **metadata)
         elif isinstance(input, StringPromptValue):
-            return chunk
+            return text_content  # String outputs don't support metadata
         elif isinstance(input, (HumanMessage, AIMessage, BaseMessage)):
-            return AIMessageChunk(content=chunk)
+            return AIMessageChunk(content=text_content, **metadata)
         elif isinstance(input, list) and all(
             isinstance(msg, BaseMessage) for msg in input
         ):
-            return AIMessageChunk(content=chunk)
+            return AIMessageChunk(content=text_content, **metadata)
         elif isinstance(input, dict):
             output_key = self.passthrough_bot_output_key
             if self.passthrough_user_input_key in input or "input" in input:
                 user_input = input.get(
                     self.passthrough_user_input_key, input.get("input")
                 )
                 if isinstance(user_input, str):
-                    return {output_key: chunk}
+                    return {
+                        output_key: text_content
+                    }  # Dict outputs don't support metadata
                 elif isinstance(user_input, list):
                     if all(
                         isinstance(msg, dict) and "role" in msg for msg in user_input
                     ):
-                        return {output_key: {"role": "assistant", "content": chunk}}
+                        return {
+                            output_key: {"role": "assistant", "content": text_content}
+                        }
                     elif all(isinstance(msg, BaseMessage) for msg in user_input):
-                        return {output_key: AIMessageChunk(content=chunk)}
-                    return {output_key: chunk}
+                        return {
+                            output_key: AIMessageChunk(content=text_content, **metadata)
+                        }
+                    return {output_key: text_content}
                 elif isinstance(user_input, BaseMessage):
-                    return {output_key: AIMessageChunk(content=chunk)}
-            return {output_key: chunk}
+                    return {
+                        output_key: AIMessageChunk(content=text_content, **metadata)
+                    }
+            return {output_key: text_content}
         elif isinstance(input, str):
-            return AIMessageChunk(content=chunk)
+            return AIMessageChunk(content=text_content, **metadata)
         else:
             raise ValueError(f"Unexpected input type: {type(input)}")
 

tests/runnable_rails/test_metadata.py

@@ -312,3 +312,74 @@ def test_partial_metadata(self, mock_rails_config):
         assert result.additional_kwargs == {"custom_field": "value"}
         # Missing fields should be handled gracefully (None or default values)
         assert result.response_metadata is None or result.response_metadata == {}
+
+    def test_streaming_metadata_preservation(self, mock_rails_config):
+        """Test that streaming preserves metadata in chunks."""
+        from unittest.mock import AsyncMock
+
+        mock_rails = AsyncMock()
+        mock_generation_response = Mock()
+        mock_generation_response.response = "Streaming response"
+        mock_generation_response.output_data = {}
+        mock_generation_response.tool_calls = None
+        mock_generation_response.llm_metadata = {
+            "additional_kwargs": {"finish_reason": "stop"},
+            "response_metadata": {"model_name": "test-model"},
+        }
+
+        async def mock_stream(*args, **kwargs):
+            chunks = [
+                {
+                    "text": "Hello ",
+                    "generation_info": {"model": "test-model", "finish_reason": "stop"},
+                },
+                {
+                    "text": "world!",
+                    "generation_info": {"model": "test-model", "finish_reason": "stop"},
+                },
+            ]
+            for chunk in chunks:
+                yield chunk
+
+        mock_rails.stream_async = mock_stream
+
+        runnable_rails = RunnableRails(config=mock_rails_config, passthrough=True)
+        runnable_rails.rails = mock_rails
+
+        chunks = list(runnable_rails.stream("Test input"))
+
+        assert len(chunks) == 2
+        for chunk in chunks:
+            assert hasattr(chunk, "content")
+            assert hasattr(chunk, "additional_kwargs") or hasattr(chunk, "model")
+            assert hasattr(chunk, "response_metadata") or hasattr(
+                chunk, "finish_reason"
+            )
+
+    async def test_async_streaming_metadata_preservation(self, mock_rails_config):
+        """Test that async streaming preserves metadata in chunks."""
+        from unittest.mock import AsyncMock
+
+        mock_rails = AsyncMock()
+
+        async def mock_stream(*args, **kwargs):
+            chunks = [
+                {"text": "Async ", "generation_info": {"model": "test-model"}},
+                {"text": "stream!", "generation_info": {"model": "test-model"}},
+            ]
+            for chunk in chunks:
+                yield chunk
+
+        mock_rails.stream_async = mock_stream
+
+        runnable_rails = RunnableRails(config=mock_rails_config, passthrough=True)
+        runnable_rails.rails = mock_rails
+
+        chunks = []
+        async for chunk in runnable_rails.astream("Test input"):
+            chunks.append(chunk)
+
+        assert len(chunks) == 2
+        for chunk in chunks:
+            assert hasattr(chunk, "content")
+            assert hasattr(chunk, "additional_kwargs") or hasattr(chunk, "model")

tests/runnable_rails/test_streaming.py

@@ -480,3 +480,55 @@ def test_streaming_with_different_input_types():
         ), f"Failed for {input_type}: {full_content}"
 
     assert llm.streaming == False
+
+
+def test_streaming_metadata_preservation():
+    """Test that streaming chunks preserve metadata structure."""
+    llm = FakeLLM(responses=["Test response"])
+    config = RailsConfig.from_content(config={"models": []})
+    model_with_rails = RunnableRails(config, llm=llm)
+
+    chunks = []
+    for chunk in model_with_rails.stream("Test input"):
+        chunks.append(chunk)
+
+    assert len(chunks) > 0
+
+    for chunk in chunks:
+        assert hasattr(chunk, "content")
+        assert hasattr(chunk, "additional_kwargs")
+        assert hasattr(chunk, "response_metadata")
+        assert isinstance(chunk.additional_kwargs, dict)
+        assert isinstance(chunk.response_metadata, dict)
+
+
+async def test_async_streaming_metadata_preservation():
+    """Test that async streaming chunks preserve metadata structure."""
+    llm = FakeLLM(responses=["Test async response"])
+    config = RailsConfig.from_content(config={"models": []})
+    model_with_rails = RunnableRails(config, llm=llm)
+
+    chunks = []
+    async for chunk in model_with_rails.astream("Test input"):
+        chunks.append(chunk)
+
+    assert len(chunks) > 0
+
+    for chunk in chunks:
+        assert hasattr(chunk, "content")
+        assert hasattr(chunk, "additional_kwargs")
+        assert hasattr(chunk, "response_metadata")
+        assert isinstance(chunk.additional_kwargs, dict)
+        assert isinstance(chunk.response_metadata, dict)
+
+
+def test_streaming_chunk_types():
+    """Test that streaming returns proper AIMessageChunk types."""
+    llm = FakeLLM(responses=["Hello world"])
+    config = RailsConfig.from_content(config={"models": []})
+    model_with_rails = RunnableRails(config, llm=llm)
+
+    chunks = list(model_with_rails.stream("Hi"))
+
+    for chunk in chunks:
+        assert chunk.__class__.__name__ == "AIMessageChunk"