fix: llm generation in case of user idle

Send for LLM generation in case of user idle rather than speaking a hardcoded sentence

api/services/pipecat/transport_setup.py

@@ -127,6 +127,7 @@ async def create_twilio_transport(
         ),
     )
 
+
 async def create_cloudonix_transport(
     websocket_client: WebSocket,
     stream_sid: str,
@@ -207,7 +208,7 @@ async def create_cloudonix_transport(
             if ENABLE_RNNOISE
             else None,
         ),
-    )    
+    )
 
 
 async def create_vonage_transport(

api/services/telephony/providers/cloudonix_provider.py

@@ -270,7 +270,7 @@ class CloudonixProvider(TelephonyProvider):
         Cloudonix embeds CXML directly in the API call during initiate_call(),
         so webhook endpoints are never called and signature verification is not needed.
         This method only exists to satisfy the abstract base class requirement.
-        
+
         Always returns True since no actual webhook verification is performed.
         """
         logger.warning(

api/services/workflow/pipecat_engine.py

@@ -13,7 +13,6 @@ from pipecat.frames.frames import (
     CancelFrame,
     EndFrame,
     FunctionCallResultProperties,
-    FunctionCallsFromLLMInfoFrame,
     LLMContextFrame,
     TTSSpeakFrame,
 )

api/services/workflow/pipecat_engine_callbacks.py

@@ -16,9 +16,7 @@ from typing import TYPE_CHECKING, Awaitable, Callable
 from loguru import logger
 
 from pipecat.frames.frames import (
-    LLMFullResponseEndFrame,
-    LLMFullResponseStartFrame,
-    TTSSpeakFrame,
+    LLMMessagesAppendFrame,
 )
 from pipecat.processors.filters.stt_mute_filter import STTMuteFilter
 from pipecat.utils.enums import EndTaskReason
@@ -68,21 +66,18 @@ def create_user_idle_callback(engine: "PipecatEngine"):
         logger.debug(f"Handling user_idle, attempt: {retry_count}")
 
         if retry_count == 1:
-            # Simulate an LLM generation, so that we can have the LLM context
-            # updated with the new message
-            await engine.task.queue_frames(
-                [
-                    LLMFullResponseStartFrame(),
-                    TTSSpeakFrame("Just checking in to see if you're still there."),
-                    LLMFullResponseEndFrame(),
-                ]
-            )
+            message = {
+                "role": "system",
+                "content": "The user has been quiet. Politely and briefly ask if they're still there in the language that the user has been speaking so far.",
+            }
+            await user_idle.push_frame(LLMMessagesAppendFrame([message], run_llm=True))
             return True
 
-        # Second attempt: terminate the call due to inactivity.
-        await user_idle.push_frame(
-            TTSSpeakFrame("It seems like you're busy right now. Have a nice day!")
-        )
+        message = {
+            "role": "system",
+            "content": "The user has been quiet. We will be disconnecting the call now. Wish them a good day.",
+        }
+        await user_idle.push_frame(LLMMessagesAppendFrame([message], run_llm=True))
         await engine.send_end_task_frame(
             EndTaskReason.USER_IDLE_MAX_DURATION_EXCEEDED.value
         )

api/services/workflow/pipecat_engine_utils.py

@@ -55,13 +55,12 @@ def update_llm_context(
     tools_schema = ToolsSchema(standard_tools=functions)
     previous_interactions = context.messages
 
-    # Filter out old system messages but keep user/assistant/function content.
-    messages: List[Dict[str, Any]] = [system_message]
-    messages.extend(
-        interaction
-        for interaction in previous_interactions
-        if interaction["role"] != "system"
-    )
+    # Replace the first message if it's a system message, otherwise prepend.
+    # Keep any system messages that appear in the middle of the conversation.
+    if previous_interactions and previous_interactions[0]["role"] == "system":
+        messages = [system_message] + previous_interactions[1:]
+    else:
+        messages = [system_message] + previous_interactions
 
     context.set_messages(messages)