Merge pull request #1886 from pipecat-ai/mb/add-otel-llm-output

Add LLM response tracing to OTel tracing

src/pipecat/utils/tracing/service_attributes.py

@@ -171,6 +171,7 @@ def add_llm_span_attributes(
     model: str,
     stream: bool = True,
     messages: Optional[str] = None,
+    output: Optional[str] = None,
     tools: Optional[str] = None,
     tool_count: Optional[int] = None,
     tool_choice: Optional[str] = None,
@@ -188,6 +189,7 @@ def add_llm_span_attributes(
         model: Model name/identifier
         stream: Whether streaming is enabled
         messages: JSON-serialized messages
+        output: Aggregated output text from the LLM
         tools: JSON-serialized tools configuration
         tool_count: Number of tools available
         tool_choice: Tool selection configuration
@@ -208,6 +210,9 @@ def add_llm_span_attributes(
     if messages:
         span.set_attribute("input", messages)
 
+    if output:
+        span.set_attribute("output", output)
+
     if tools:
         span.set_attribute("tools", tools)
 

src/pipecat/utils/tracing/service_decorators.py

@@ -282,6 +282,7 @@ def traced_llm(func: Optional[Callable] = None, *, name: Optional[str] = None) -
     - Tool configurations
     - Token usage metrics
     - Performance metrics like TTFB
+    - Aggregated output text
 
     Args:
         func: The LLM method to trace.
@@ -313,6 +314,26 @@ def traced_llm(func: Optional[Callable] = None, *, name: Optional[str] = None) -
                     span_name, context=parent_context
                 ) as current_span:
                     try:
+                        # Store original method and output aggregator
+                        original_push_frame = self.push_frame
+                        output_text = ""  # Simple string accumulation
+
+                        async def traced_push_frame(frame, direction=None):
+                            nonlocal output_text
+                            # Capture text from LLMTextFrame during streaming
+                            if (
+                                hasattr(frame, "__class__")
+                                and frame.__class__.__name__ == "LLMTextFrame"
+                                and hasattr(frame, "text")
+                            ):
+                                output_text += frame.text
+
+                            # Call original
+                            if direction is not None:
+                                return await original_push_frame(frame, direction)
+                            else:
+                                return await original_push_frame(frame)
+
                         # For token usage monitoring
                         original_start_llm_usage_metrics = None
                         if hasattr(self, "start_llm_usage_metrics"):
@@ -331,6 +352,9 @@ def traced_llm(func: Optional[Callable] = None, *, name: Optional[str] = None) -
                             self.start_llm_usage_metrics = wrapped_start_llm_usage_metrics
 
                         try:
+                            # Replace push_frame to capture output
+                            self.push_frame = traced_push_frame
+
                             # Detect if we're using Google's service
                             is_google_service = "google" in service_class_name.lower()
 
@@ -411,13 +435,24 @@ def traced_llm(func: Optional[Callable] = None, *, name: Optional[str] = None) -
 
                             # Add all gathered attributes to the span
                             add_llm_span_attributes(span=current_span, **attribute_kwargs)
-                        except Exception as e:
-                            logging.warning(f"Error adding initial LLM attributes: {e}")
 
-                        # Call the original function
+                        except Exception as e:
-                        return await f(self, context, *args, **kwargs)
+                            logging.warning(f"Error setting up LLM tracing: {e}")
+                            # Don't raise - let the function execute anyway
+
+                        # Run function with modified push_frame to capture the output
+                        result = await f(self, context, *args, **kwargs)
+
+                        # Add aggregated output after function completes, if available
+                        if output_text:
+                            current_span.set_attribute("output", output_text)
+
+                        return result
+
                     finally:
-                        # Restore the original methods if we overrode them
+                        # Always restore the original methods
+                        self.push_frame = original_push_frame
+
                         if (
                             "original_start_llm_usage_metrics" in locals()
                             and original_start_llm_usage_metrics