Improve conversation quality: use full history, increase tokens to 150, better sampling

app.py

@@ -98,56 +98,25 @@ def chat_completions():
         
         messages = data.get("messages", [])
         
-        # Extract user messages (ignore system messages from Vapi)
-        user_messages = [msg for msg in messages if msg.get("role") == "user"]
-        
-        if not user_messages:
-            prompt = data.get("prompt", "")
-        else:
-            # Get the last user message
-            prompt = user_messages[-1].get("content", "")
-        
-        if not prompt:
-            return jsonify({"error": "No prompt provided"}), 400
-        
-        print(f"User prompt: {prompt}")
-        
-        # Sales-only keyword filter
-        sales_keywords = ['sale', 'price', 'objection', 'close', 'deal', 'customer', 'prospect', 
-                         'lead', 'pitch', 'negotiate', 'client', 'business', 'product', 'service',
-                         'wisemate', 'revenue', 'quota', 'conversion', 'crm', 'follow up', 'meeting',
-                         'appointment', 'schedule', 'calendar', 'demo', 'call', 'consultation']
-        
-        # Check if question is sales-related
-        prompt_lower = prompt.lower()
-        is_sales_related = any(keyword in prompt_lower for keyword in sales_keywords)
-        
-        # If not sales-related, return redirect message
-        if not is_sales_related and len(prompt.split()) > 3:  # Only filter if it's a real question
-            redirect_message = "I'm Daniel from Wisemate, and I specialize in sales and business topics. I can help you with sales techniques, handling objections, closing deals, and Wisemate's services. How can I assist you with your sales needs?"
+        # Build conversation history in Gemma format
+        # Only use user and assistant messages, skip system (Vapi's system prompt is just context)
+        conversation_parts = []
+        for msg in messages:
+            role = msg.get("role")
+            content = msg.get("content", "")
             
-            return jsonify({
-                "id": "chatcmpl-redirect",
-                "object": "chat.completion",
-                "created": int(__import__('time').time()),
-                "model": MODEL_NAME,
-                "choices": [{
-                    "index": 0,
-                    "message": {
-                        "role": "assistant",
-                        "content": redirect_message
-                    },
-                    "finish_reason": "stop"
-                }],
-                "usage": {
-                    "prompt_tokens": 0,
-                    "completion_tokens": 0,
-                    "total_tokens": 0
-                }
-            }), 200
-        
-        # Gemma formatında prompt
-        formatted_prompt = f"<start_of_turn>user\n{prompt}<end_of_turn>\n<start_of_turn>model\n"
+            if role == "user":
+                conversation_parts.append(f"<start_of_turn>user\n{content}<end_of_turn>")
+            elif role == "assistant":
+                conversation_parts.append(f"<start_of_turn>model\n{content}<end_of_turn>")
+            # Skip system messages - Vapi's system prompt is too long for context
+        
+        # Add the model turn start
+        conversation_parts.append("<start_of_turn>model\n")
+        formatted_prompt = "\n".join(conversation_parts)
+        
+        print(f"Formatted prompt length: {len(formatted_prompt)} chars")
+        print(f"Last user message: {[m['content'] for m in messages if m.get('role')=='user'][-1] if messages else 'none'}")
 
         # Model yanıtı üret
         inputs = tokenizer(formatted_prompt, return_tensors="pt")
@@ -159,9 +128,11 @@ def chat_completions():
         print(f"Generating response...")
         outputs = model.generate(
             **inputs,
-            max_new_tokens=50,  # Drastically reduced for real-time voice (was 100)
-            temperature=0.3,  # Lower for faster, more focused responses
-            do_sample=False,  # Greedy decoding for speed
+            max_new_tokens=150,  # Increased for complete appointment booking responses
+            temperature=0.7,  # Higher for more natural conversation
+            do_sample=True,  # Enable sampling for variety
+            top_p=0.9,
+            repetition_penalty=1.1,
             pad_token_id=tokenizer.pad_token_id,
             eos_token_id=tokenizer.eos_token_id
         )