add template

app.py

@@ -194,7 +194,7 @@ def review_pitchdeck(text: str) -> dict:
     
     deck_text = text[:12000]
     
-    prompt = f"""You are a senior venture capitalist with 15+ years of experience evaluating thousands of pitch decks. You know the patterns that lead to funding vs. ghosting. Based on extensive research analyzing hundreds of decks, these are the critical failure points:
+    system_message = """You are a senior venture capitalist with 15+ years of experience evaluating thousands of pitch decks. You know the patterns that lead to funding vs. ghosting. Based on extensive research analyzing hundreds of decks, these are the critical failure points:
 
 1. Beautiful decks missing commercial backbone (GTM, financials, market sizing, clear ask)
 2. Giant market claims without credibility - claiming $50B TAM instead of sharp, addressable market
@@ -214,7 +214,9 @@ THE 5 CRITICAL QUESTIONS every deck must answer clearly:
 4. Why you?
 5. What do you need?
 
-Deck Content:
+Be brutally honest. Commercial clarity keeps doors open - GTM and financials get you funded. Emotion opens the door, but logic closes the deal."""
+
+    user_message = f"""Deck Content:
 {deck_text}
 
 TASK:
@@ -274,13 +276,17 @@ Produce ONLY valid JSON with these exact fields:
   }},
   "slide_count": number,
   "slide_count_assessment": "optimal (12-14)" | "too_many" | "too_few"
-}}
-
-Be brutally honest. Commercial clarity keeps doors open - GTM and financials get you funded. Emotion opens the door, but logic closes the deal.
-"""
+}}"""
 
     try:
+        messages = [
+            {"role": "system", "content": system_message},
+            {"role": "user", "content": user_message}
+        ]
+        
+        prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
         inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=8192).to(model.device)
+        
         outputs = model.generate(
             **inputs,
             max_new_tokens=2000,
@@ -291,6 +297,9 @@ Be brutally honest. Commercial clarity keeps doors open - GTM and financials get
         )
         raw_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
         
+        if "<|assistant|>" in raw_output:
+            raw_output = raw_output.split("<|assistant|>")[-1]
+        
         start = raw_output.find('{')
         end = raw_output.rfind('}') + 1
         
@@ -320,10 +329,7 @@ def generate_improvement_pointers(review: dict) -> dict:
             "improvement_pointers": []
         }
     
-    improvement_prompt = f"""You are a pitch deck consultant with expertise from reviewing hundreds of founder decks. Based on this VC review, provide actionable, specific improvement pointers grounded in real-world failure patterns.
-
-VC Review:
-{json.dumps(review, indent=2)}
+    system_message = """You are a pitch deck consultant with expertise from reviewing hundreds of founder decks. Provide actionable, specific improvement pointers grounded in real-world failure patterns.
 
 Focus on fixing the TOP 10 COMMON FAILURES:
 1. Add commercial backbone: GTM plan, financials, market sizing, clear ask
@@ -343,7 +349,10 @@ Generate 5-10 prioritized improvement pointers addressing the specific failures
 - What to fix first, second, third
 - Slide-by-slide improvements where critical issues were identified
 - How to address deal breakers and common failure patterns
-- Quick wins vs. strategic changes
+- Quick wins vs. strategic changes"""
+
+    user_message = f"""VC Review:
+{json.dumps(review, indent=2)}
 
 Return ONLY valid JSON:
 {{
@@ -360,11 +369,17 @@ Return ONLY valid JSON:
   "quick_wins": ["quick fix 1", "quick fix 2"],
   "strategic_changes": ["strategic change 1", "strategic change 2"],
   "critical_fixes": ["must-fix issue 1", "must-fix issue 2"]
-}}
-"""
+}}"""
 
     try:
-        inputs = tokenizer(improvement_prompt, return_tensors="pt", truncation=True, max_length=8192).to(model.device)
+        messages = [
+            {"role": "system", "content": system_message},
+            {"role": "user", "content": user_message}
+        ]
+        
+        prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=8192).to(model.device)
+        
         outputs = model.generate(
             **inputs,
             max_new_tokens=1500,
@@ -374,6 +389,9 @@ Return ONLY valid JSON:
         )
         raw_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
         
+        if "<|assistant|>" in raw_output:
+            raw_output = raw_output.split("<|assistant|>")[-1]
+        
         start = raw_output.find('{')
         end = raw_output.rfind('}') + 1