Reduce verbosity: role constraints + token limits

backend/core/conversation_manager.py

@@ -32,6 +32,24 @@ sys.path.insert(0, str(Path(__file__).parent.parent))
 from backend.core.llm_client import create_llm_client
 from backend.core.persona_system import PersonaSystem
 
+SURVEYOR_MAX_TOKENS = 140
+PATIENT_MAX_TOKENS = 240
+
+SURVEYOR_STYLE_GUIDANCE = (
+    "Constraints:\n"
+    "- Be concise (1–2 short sentences).\n"
+    "- Ask at most ONE question.\n"
+    "- Focus on listening/clarifying and encouraging the patient to elaborate.\n"
+    "- Avoid long explanations or multi-paragraph monologues.\n"
+)
+
+PATIENT_STYLE_GUIDANCE = (
+    "Constraints:\n"
+    "- Respond in 2–4 sentences.\n"
+    "- Answer the question directly, then add 1–2 relevant details.\n"
+    "- Avoid long tangents or multi-paragraph monologues.\n"
+)
+
 
 class ConversationState(Enum):
     """States for conversation lifecycle."""
@@ -155,19 +173,29 @@ class ConversationManager:
 
             # Build prompt based on message type
             if msg_type == "greeting":
-                user_prompt = "Please introduce yourself and ask your first survey question."
+                user_prompt = (
+                    "Please introduce yourself briefly and ask your first survey question.\n\n"
+                    f"{SURVEYOR_STYLE_GUIDANCE}"
+                )
             elif msg_type == "closing":
-                user_prompt = "Thank the patient and conclude the survey professionally."
+                user_prompt = (
+                    "Thank the patient and conclude the survey professionally.\n\n"
+                    f"{SURVEYOR_STYLE_GUIDANCE}"
+                )
             else:  # follow_up
                 if self.history:
                     last_patient_msg = next((msg for msg in reversed(self.history)
                                            if msg["role"] == "patient"), None)
                     if last_patient_msg:
-                        user_prompt = f"The patient just said: '{last_patient_msg['content']}'. Please ask an appropriate follow-up question."
+                        user_prompt = (
+                            f"The patient just said: '{last_patient_msg['content']}'. "
+                            "Respond with a brief acknowledgment and ask an appropriate follow-up question.\n\n"
+                            f"{SURVEYOR_STYLE_GUIDANCE}"
+                        )
                     else:
-                        user_prompt = "Please ask the next survey question."
+                        user_prompt = f"Please ask the next survey question.\n\n{SURVEYOR_STYLE_GUIDANCE}"
                 else:
-                    user_prompt = "Please ask the next survey question."
+                    user_prompt = f"Please ask the next survey question.\n\n{SURVEYOR_STYLE_GUIDANCE}"
 
             # Convert history to conversation format
             conversation_history = [
@@ -186,7 +214,9 @@ class ConversationManager:
             # Generate response
             response = await self.client.generate(
                 prompt=prompt_with_history,
-                system_prompt=system_prompt
+                system_prompt=system_prompt,
+                max_tokens=SURVEYOR_MAX_TOKENS,
+                temperature=0.4
             )
 
             # Create message dict
@@ -234,9 +264,13 @@ class ConversationManager:
                                     if msg["role"] == "surveyor"), None)
 
             if not last_surveyor_msg:
-                user_prompt = "Please respond to the survey interviewer's greeting."
+                user_prompt = f"Please respond to the survey interviewer's greeting.\n\n{PATIENT_STYLE_GUIDANCE}"
             else:
-                user_prompt = f"The interviewer just said: '{last_surveyor_msg['content']}'. Please respond naturally as your persona would."
+                user_prompt = (
+                    f"The interviewer just said: '{last_surveyor_msg['content']}'. "
+                    "Please respond naturally as your persona would.\n\n"
+                    f"{PATIENT_STYLE_GUIDANCE}"
+                )
 
             # Convert history to conversation format
             conversation_history = [
@@ -255,7 +289,9 @@ class ConversationManager:
             # Generate response
             response = await self.client.generate(
                 prompt=prompt_with_history,
-                system_prompt=system_prompt
+                system_prompt=system_prompt,
+                max_tokens=PATIENT_MAX_TOKENS,
+                temperature=0.7
             )
 
             # Create message dict

data/patient_personas.yaml

@@ -52,6 +52,9 @@ personas:
       especially medical terms. You tend to give detailed answers and share related health 
       experiences. You're doing well with your conditions but worry about your health. 
       Be polite, cooperative, and occasionally ask for clarification.
+      
+      Keep your responses to 2–4 sentences. Answer the question directly, then add a
+      couple of relevant details. Avoid long tangents or multi-paragraph monologues.
 
   - id: "anxious_parent_001"
     name: "Jennifer Chen"
@@ -102,6 +105,9 @@ personas:
       protective and sometimes defensive about your parenting choices. You want reassurance 
       that you're doing the right things for your child's health. Be questioning but not 
       hostile, and provide detailed information about your child's condition when asked.
+      
+      Keep your responses to 2–4 sentences. Answer the question directly, then add a
+      couple of relevant details. Avoid long tangents or multi-paragraph monologues.
 
   - id: "busy_professional_001" 
     name: "David Rodriguez"
@@ -152,4 +158,6 @@ personas:
       You have work stress, irregular sleep, and occasional back pain from sitting. You're 
       slightly impatient but not rude. If the interviewer acknowledges your time constraints, 
       become more cooperative. Give minimal answers unless specifically asked to elaborate. 
-      Occasionally mention checking the time or having another meeting soon.
+      Occasionally mention checking the time or having another meeting soon.
+      
+      Keep your responses to 1–3 sentences unless the interviewer asks you to elaborate.

data/surveyor_personas.yaml

@@ -48,6 +48,10 @@ personas:
       interest in responses, ask clarifying follow-ups, and ensure respondents feel 
       heard and respected. Begin surveys with a warm introduction and clear explanation 
       of the survey's purpose and confidentiality measures.
+      
+      Keep your turns short: 1–2 sentences, with at most one question. Your role is to
+      listen, reassure, and ask concise follow-up questions that help the patient share
+      relevant details (avoid long monologues).
 
   - id: "friendly_researcher_001"
     name: "Alex Thompson"
@@ -95,4 +99,8 @@ personas:
       time. If someone seems nervous or confused, take extra time to explain and 
       reassure. Your goal is to create a relaxed atmosphere where people feel safe 
       sharing health information. Always express gratitude for their responses and 
-      maintain a supportive, non-judgmental tone throughout.
+      maintain a supportive, non-judgmental tone throughout.
+      
+      Keep your turns short: 1–2 sentences, with at most one question. Your role is to
+      listen, reassure, and ask concise follow-up questions that help the patient share
+      relevant details (avoid long monologues).