feat(runnable-rails): implement AIMessage metadata parity in RunnableRails

Ensure AIMessage responses from RunnableRails contain the same metadata
fields (response_metadata, usage_metadata, additional_kwargs, id) as
direct LLM calls, enabling consistent LangChain integration behavior.

Author: Pouyanpi

nemoguardrails/actions/llm/utils.py

@@ -26,6 +26,7 @@
 from nemoguardrails.colang.v2_x.runtime.flows import InternalEvent, InternalEvents
 from nemoguardrails.context import (
     llm_call_info_var,
+    llm_response_metadata_var,
     reasoning_trace_var,
     tool_calls_var,
 )
@@ -85,6 +86,7 @@ async def llm_call(
         response = await _invoke_with_message_list(llm, prompt, all_callbacks, stop)
 
     _store_tool_calls(response)
+    _store_response_metadata(response)
     return _extract_content(response)
 
 
@@ -173,6 +175,20 @@ def _store_tool_calls(response) -> None:
     tool_calls_var.set(tool_calls)
 
 
+def _store_response_metadata(response) -> None:
+    """Store response metadata excluding content for metadata preservation."""
+    if hasattr(response, "model_fields"):
+        metadata = {}
+        for field_name in response.model_fields:
+            if (
+                field_name != "content"
+            ):  # Exclude content since it may be modified by rails
+                metadata[field_name] = getattr(response, field_name)
+        llm_response_metadata_var.set(metadata)
+    else:
+        llm_response_metadata_var.set(None)
+
+
 def _extract_content(response) -> str:
     """Extract text content from response."""
     if hasattr(response, "content"):
@@ -655,3 +671,15 @@ def get_and_clear_tool_calls_contextvar() -> Optional[list]:
         tool_calls_var.set(None)
         return tool_calls
     return None
+
+
+def get_and_clear_response_metadata_contextvar() -> Optional[dict]:
+    """Get the current response metadata and clear it from the context.
+
+    Returns:
+        Optional[dict]: The response metadata if it exists, None otherwise.
+    """
+    if metadata := llm_response_metadata_var.get():
+        llm_response_metadata_var.set(None)
+        return metadata
+    return None

nemoguardrails/context.py

@@ -42,3 +42,8 @@
 tool_calls_var: contextvars.ContextVar[Optional[list]] = contextvars.ContextVar(
     "tool_calls", default=None
 )
+
+# The response metadata from the current LLM response.
+llm_response_metadata_var: contextvars.ContextVar[
+    Optional[dict]
+] = contextvars.ContextVar("llm_response_metadata", default=None)

nemoguardrails/integrations/langchain/runnable_rails.py

@@ -393,11 +393,21 @@ def _format_passthrough_output(self, result: Any, context: Dict[str, Any]) -> An
         return passthrough_output
 
     def _format_chat_prompt_output(
-        self, result: Any, tool_calls: Optional[list] = None
+        self,
+        result: Any,
+        tool_calls: Optional[list] = None,
+        metadata: Optional[dict] = None,
     ) -> AIMessage:
         """Format output for ChatPromptValue input."""
         content = self._extract_content_from_result(result)
-        if tool_calls:
+
+        if metadata and isinstance(metadata, dict):
+            metadata_copy = metadata.copy()
+            metadata_copy.pop("content", None)
+            if tool_calls:
+                metadata_copy["tool_calls"] = tool_calls
+            return AIMessage(content=content, **metadata_copy)
+        elif tool_calls:
             return AIMessage(content=content, tool_calls=tool_calls)
         return AIMessage(content=content)
 
@@ -406,11 +416,21 @@ def _format_string_prompt_output(self, result: Any) -> str:
         return self._extract_content_from_result(result)
 
     def _format_message_output(
-        self, result: Any, tool_calls: Optional[list] = None
+        self,
+        result: Any,
+        tool_calls: Optional[list] = None,
+        metadata: Optional[dict] = None,
     ) -> AIMessage:
         """Format output for BaseMessage input types."""
         content = self._extract_content_from_result(result)
-        if tool_calls:
+
+        if metadata and isinstance(metadata, dict):
+            metadata_copy = metadata.copy()
+            metadata_copy.pop("content", None)
+            if tool_calls:
+                metadata_copy["tool_calls"] = tool_calls
+            return AIMessage(content=content, **metadata_copy)
+        elif tool_calls:
             return AIMessage(content=content, tool_calls=tool_calls)
         return AIMessage(content=content)
 
@@ -434,25 +454,50 @@ def _format_dict_output_for_dict_message_list(
         }
 
     def _format_dict_output_for_base_message_list(
-        self, result: Any, output_key: str, tool_calls: Optional[list] = None
+        self,
+        result: Any,
+        output_key: str,
+        tool_calls: Optional[list] = None,
+        metadata: Optional[dict] = None,
     ) -> Dict[str, Any]:
         """Format dict output when user input was a list of BaseMessage objects."""
         content = self._extract_content_from_result(result)
-        if tool_calls:
+
+        if metadata and isinstance(metadata, dict):
+            metadata_copy = metadata.copy()
+            metadata_copy.pop("content", None)
+            if tool_calls:
+                metadata_copy["tool_calls"] = tool_calls
+            return {output_key: AIMessage(content=content, **metadata_copy)}
+        elif tool_calls:
             return {output_key: AIMessage(content=content, tool_calls=tool_calls)}
         return {output_key: AIMessage(content=content)}
 
     def _format_dict_output_for_base_message(
-        self, result: Any, output_key: str, tool_calls: Optional[list] = None
+        self,
+        result: Any,
+        output_key: str,
+        tool_calls: Optional[list] = None,
+        metadata: Optional[dict] = None,
     ) -> Dict[str, Any]:
         """Format dict output when user input was a BaseMessage."""
         content = self._extract_content_from_result(result)
-        if tool_calls:
+
+        if metadata:
+            metadata_copy = metadata.copy()
+            if tool_calls:
+                metadata_copy["tool_calls"] = tool_calls
+            return {output_key: AIMessage(content=content, **metadata_copy)}
+        elif tool_calls:
             return {output_key: AIMessage(content=content, tool_calls=tool_calls)}
         return {output_key: AIMessage(content=content)}
 
     def _format_dict_output(
-        self, input_dict: dict, result: Any, tool_calls: Optional[list] = None
+        self,
+        input_dict: dict,
+        result: Any,
+        tool_calls: Optional[list] = None,
+        metadata: Optional[dict] = None,
     ) -> Dict[str, Any]:
         """Format output for dictionary input."""
         output_key = self.passthrough_bot_output_key
@@ -471,13 +516,13 @@ def _format_dict_output(
                     )
                 elif all(isinstance(msg, BaseMessage) for msg in user_input):
                     return self._format_dict_output_for_base_message_list(
-                        result, output_key, tool_calls
+                        result, output_key, tool_calls, metadata
                     )
                 else:
                     return {output_key: result}
             elif isinstance(user_input, BaseMessage):
                 return self._format_dict_output_for_base_message(
-                    result, output_key, tool_calls
+                    result, output_key, tool_calls, metadata
                 )
 
         # Generic fallback for dictionaries
@@ -490,6 +535,7 @@ def _format_output(
         result: Any,
         context: Dict[str, Any],
         tool_calls: Optional[list] = None,
+        metadata: Optional[dict] = None,
     ) -> Any:
         """Format the output based on the input type and rails result.
 
@@ -512,17 +558,17 @@ def _format_output(
             return self._format_passthrough_output(result, context)
 
         if isinstance(input, ChatPromptValue):
-            return self._format_chat_prompt_output(result, tool_calls)
+            return self._format_chat_prompt_output(result, tool_calls, metadata)
         elif isinstance(input, StringPromptValue):
             return self._format_string_prompt_output(result)
         elif isinstance(input, (HumanMessage, AIMessage, BaseMessage)):
-            return self._format_message_output(result, tool_calls)
+            return self._format_message_output(result, tool_calls, metadata)
         elif isinstance(input, list) and all(
             isinstance(msg, BaseMessage) for msg in input
         ):
-            return self._format_message_output(result, tool_calls)
+            return self._format_message_output(result, tool_calls, metadata)
         elif isinstance(input, dict):
-            return self._format_dict_output(input, result, tool_calls)
+            return self._format_dict_output(input, result, tool_calls, metadata)
         elif isinstance(input, str):
             return self._format_string_prompt_output(result)
         else:
@@ -669,7 +715,9 @@ def _full_rails_invoke(
             result = result[0]
 
         # Format and return the output based in input type
-        return self._format_output(input, result, context, res.tool_calls)
+        return self._format_output(
+            input, result, context, res.tool_calls, res.llm_metadata
+        )
 
     async def ainvoke(
         self,
@@ -731,7 +779,9 @@ async def _full_rails_ainvoke(
         result = res.response
 
         # Format and return the output based on input type
-        return self._format_output(input, result, context, res.tool_calls)
+        return self._format_output(
+            input, result, context, res.tool_calls, res.llm_metadata
+        )
 
     def stream(
         self,

nemoguardrails/rails/llm/llmrails.py

@@ -33,6 +33,7 @@
 from nemoguardrails.actions.llm.generation import LLMGenerationActions
 from nemoguardrails.actions.llm.utils import (
     get_and_clear_reasoning_trace_contextvar,
+    get_and_clear_response_metadata_contextvar,
     get_and_clear_tool_calls_contextvar,
     get_colang_history,
 )
@@ -1086,6 +1087,7 @@ async def generate_async(
                 options.log.internal_events = True
 
         tool_calls = get_and_clear_tool_calls_contextvar()
+        llm_metadata = get_and_clear_response_metadata_contextvar()
 
         # If we have generation options, we prepare a GenerationResponse instance.
         if options:
@@ -1106,6 +1108,9 @@ async def generate_async(
             if tool_calls:
                 res.tool_calls = tool_calls
 
+            if llm_metadata:
+                res.llm_metadata = llm_metadata
+
             if self.config.colang_version == "1.0":
                 # If output variables are specified, we extract their values
                 if options.output_vars:

nemoguardrails/rails/llm/options.py

@@ -412,6 +412,10 @@ class GenerationResponse(BaseModel):
         default=None,
         description="Tool calls extracted from the LLM response, if any.",
     )
+    llm_metadata: Optional[dict] = Field(
+        default=None,
+        description="Metadata from the LLM response (additional_kwargs, response_metadata, usage_metadata, etc.)",
+    )
 
 
 if __name__ == "__main__":