twilio: get greeting message from agents

apps/twilio_handler/app.py

@@ -204,33 +204,38 @@ def handle_call(call_sid, workflow_id, project_id=None):
         # Create TwiML response
         response = VoiceResponse()
 
-        # Check if this is a new call
+# Check if this is a new call (no turns yet)
         if call_state.get('turn_count', 0) == 0:
-            # Initial greeting for new calls
-            greeting = "Hello! I'm your RowBoat assistant. How can I help you today?"
-            logger.info(f"New call, preparing greeting: {greeting}")
+            logger.info("First turn: generating AI greeting using an empty user input...")
 
+            # Generate greeting by calling process_conversation_turn with empty user input
             try:
-                # Use streaming audio endpoint instead of generating files
-                # Include a unique ID to prevent caching
-                unique_id = str(uuid.uuid4())
-                # Use a relative URL - Twilio will use the same host as the webhook
-                audio_url = f"/stream-audio/{call_sid}/greeting/{unique_id}"
-                logger.info(f"Streaming greeting from relative URL: {audio_url}")
-
-                # Play the greeting via streaming
-                response.play(audio_url)
+                ai_greeting, updated_messages, updated_state = process_conversation_turn(
+                    user_input="",  # empty to signal "give me your greeting"
+                    workflow_id=call_state['workflow_id'],
+                    system_prompt=call_state['system_prompt'],
+                    previous_messages=[],
+                    previous_state=None,
+                    project_id=call_state.get('project_id')
+                )
             except Exception as e:
-                logger.error(f"Error with audio streaming for greeting: {str(e)}")
-                import traceback
-                logger.error(traceback.format_exc())
-                # Fallback to Twilio TTS
-                response.say(greeting, voice='alice')
+                logger.error(f"Error generating AI greeting: {str(e)}")
+                ai_greeting = "Hello, I encountered an issue creating a greeting. How can I help you?"
 
-            # Update call state
+                # Fallback: no changes to updated_messages/updated_state
+                updated_messages = []
+                updated_state = None
+
+            # Update call_state with AI greeting
+            call_state['messages'] = updated_messages
+            call_state['state'] = updated_state
+            call_state['conversation_history'].append({
+                'user': "",  # empty user
+                'assistant': ai_greeting
+            })
             call_state['turn_count'] = 1
 
-            # Save to MongoDB (primary source of truth)
+            # Save changes to MongoDB
             try:
                 save_call_state(call_sid, call_state)
                 logger.info(f"Saved greeting state to MongoDB for {call_sid}")
@@ -238,24 +243,24 @@ def handle_call(call_sid, workflow_id, project_id=None):
                 logger.error(f"Error saving greeting state to MongoDB: {str(e)}")
                 raise RuntimeError(f"Failed to save greeting state to MongoDB: {str(e)}")
 
-            # Update local memory cache
             active_calls[call_sid] = call_state
 
-            # Instead of using both Gather and Record which compete for input,
-            # just use Gather for speech recognition, and rely on its SpeechResult
-            # This is more reliable than trying to use Record and Deepgram
+            # Play the greeting via streaming audio
+            unique_id = str(uuid.uuid4())
+            audio_url = f"/stream-audio/{call_sid}/greeting/{unique_id}"
+            logger.info(f"Will stream greeting from {audio_url}")
+            response.play(audio_url)
+
+            # Gather user input next
             gather = Gather(
                 input='speech',
                 action=f'/process_speech?call_sid={call_sid}',
                 speech_timeout='auto',
                 language='en-US',
-                enhanced=True,  # Enable enhanced speech recognition
-                speechModel='phone_call'  # Optimize for phone calls
+                enhanced=True,
+                speechModel='phone_call'
             )
             response.append(gather)
-
-            # If no input detected, redirect to twiml endpoint
-            # Call state will be retrieved from MongoDB
             response.redirect('/twiml')
 
         logger.info(f"Returning response: {str(response)}")
@@ -492,10 +497,7 @@ def stream_audio(call_sid, text_type, unique_id):
         # Determine what text to synthesize
         text_to_speak = ""
 
-        if text_type == "greeting":
-            # Use default greeting
-            text_to_speak = "Hello! I'm your RowBoat assistant. How can I help you today?"
-        elif text_type == "response":
+        if text_type == "greeting" or text_type == "response":
             # Get the text from call state (try MongoDB first, then memory)
             call_state = None
 

apps/twilio_handler/twilio_api.py

@@ -55,8 +55,9 @@ def process_conversation_turn(
         if not messages or not any(msg.role == 'system' for msg in messages):
             messages.append(SystemMessage(role='system', content=system_prompt))
 
-        # Add the user's new message
-        messages.append(UserMessage(role='user', content=user_input))
+        # Add the user's new
+        if user_input:
+            messages.append(UserMessage(role='user', content=user_input))
 
         # Process the conversation using the stateless API
         logger.info(f"Sending to RowBoat API with {len(messages)} messages")
@@ -84,20 +85,6 @@ def process_conversation_turn(
         # Update messages list with the new responses
         final_messages = messages + response_messages
 
-        # dump_data = {
-        #     'messages': [msg.model_dump() for msg in messages],
-        #     'response_messages': [msg.model_dump() for msg in response_messages],
-        #     'state': new_state
-        # }
-        # # Write messages to a debug file for inspection
-        # fname = f'debug_dump_{time.time()}.json'
-        # try:
-        #     with open(fname, 'w') as f:
-        #         json.dump(dump_data, f, indent=2)
-        #     logger.info(f"Wrote debug info to {fname}")
-        # except Exception as e:
-        #     logger.error(f"Failed to write message debug file: {str(e)}")
-
 
         logger.info(f"Got response from RowBoat API: {assistant_response[:100]}...")
         return assistant_response, final_messages, new_state

docker-compose.yml

@@ -184,7 +184,7 @@ services:
       context: ./apps/twilio_handler
       dockerfile: Dockerfile
     ports:
-      - "3009:3009"
+      - "4010:4010"
     environment:
       - ELEVENLABS_API_KEY=${ELEVENLABS_API_KEY}
       - ROWBOAT_API_HOST=http://rowboat:3000