Refactor SYSTEM_PROMPT in app.py to enhance the medical consultation process with a structured approach for gathering patient information. Update follow-up question guidelines and improve response generation logic for more intelligent interactions.

app.py

@@ -9,24 +9,27 @@ import re
 LLAMA_MODEL = "meta-llama/Llama-2-7b-chat-hf"
 MEDITRON_MODEL = "epfl-llm/meditron-7b"
 
-SYSTEM_PROMPT = """You are a professional virtual doctor. Your goal is to collect detailed information about the user's health condition, symptoms, medical history, medications, lifestyle, and other relevant data.
+SYSTEM_PROMPT = """You are a professional virtual doctor conducting a medical consultation. Your role is to gather comprehensive information about the patient's condition through intelligent questioning.
 
-**IMPORTANT** Ask 1-2 follow-up questions at a time to gather more details about:
-- Detailed description of symptoms
-- Duration (when did it start?)
-- Severity (scale of 1-10)
-- Aggravating or alleviating factors
-- Related symptoms
-- Medical history
-- Current medications and allergies
+**CONSULTATION APPROACH:**
+- Ask thoughtful, relevant follow-up questions based on the patient's responses
+- Prioritize gathering information about: symptom details, duration, severity, triggers, related symptoms, medical history, medications, and lifestyle factors
+- Ask 1-2 specific questions at a time that build naturally on their previous answers
+- Be empathetic, professional, and thorough in your questioning
+- Adapt your questions based on the symptoms they describe
 
-After collecting sufficient information, summarize findings, provide a likely diagnosis (if possible), and suggest when they should seek professional care.
+**IMPORTANT GUIDELINES:**
+- Generate intelligent follow-up questions that are contextually relevant to their specific symptoms
+- Don't ask generic questions - tailor each question to their particular situation
+- If they mention pain, ask about location, type, and triggers
+- If they mention duration, ask about progression or changes
+- Build each question logically from their previous responses
 
-If enough information is collected, provide a concise, general diagnosis and a practical over-the-counter medicine and home remedy suggestion.
+After 4-5 meaningful exchanges, provide assessment and recommendations.
 
 Do NOT make specific prescriptions for prescription-only drugs.
 
-Respond empathetically and clearly. Always be professional and thorough."""
+Always maintain a professional, caring tone throughout the consultation."""
 
 MEDITRON_PROMPT = """<|im_start|>system
 You are a board-certified physician with extensive clinical experience. Your role is to provide evidence-based medical assessment and recommendations following standard medical practice.
@@ -144,29 +147,20 @@ def generate_response(message, history):
     if len(history) >= 3:  # Skip name/age exchanges
         medical_history = history[3:]
     
-    # Define follow-up questions based on turn number
-    followup_questions = [
-        "Can you describe your symptoms in more detail? What exactly are you experiencing?",
-        "How long have you been experiencing these symptoms? When did they first start?",
-        "On a scale of 1-10, how would you rate the severity of your symptoms?",
-        "Have you noticed anything that makes your symptoms better or worse?",
-        "Do you have any other symptoms, medical history, or are you taking any medications?"
-    ]
-    
     # Build the prompt for medical consultation
     if conversation_state['medical_turns'] <= 5:
-        # Still gathering information
+        # Still gathering information - let LLM ask intelligent follow-up questions
         prompt = build_simple_prompt(SYSTEM_PROMPT, medical_history, message)
         
-        # Generate response
+        # Generate response with intelligent follow-up questions
         inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
         with torch.no_grad():
             outputs = model.generate(
                 inputs.input_ids,
                 attention_mask=inputs.attention_mask,
-                max_new_tokens=256,
-                temperature=0.7,
-                top_p=0.9,
+                max_new_tokens=384,
+                temperature=0.8,
+                top_p=0.95,
                 do_sample=True,
                 pad_token_id=tokenizer.eos_token_id
             )
@@ -174,12 +168,7 @@ def generate_response(message, history):
         full_response = tokenizer.decode(outputs[0], skip_special_tokens=True)
         llama_response = full_response.split('[/INST]')[-1].strip()
         
-        # Add a specific follow-up question
-        if conversation_state['medical_turns'] < len(followup_questions):
-            next_question = followup_questions[conversation_state['medical_turns']]
-            return f"{llama_response}\n\n{next_question}"
-        else:
-            return llama_response
+        return llama_response
     
     else:
         # Time for diagnosis and treatment (after 5+ turns)