Simplify prompt and validation - robust creative output

Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>

backend/src/rag_system.py

@@ -23,59 +23,14 @@ logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
 
 
-# Prompt for Sarvam-1 - Creative writing inspired by Hindi literature
-CREATIVE_PROMPT = """You are a multilingual creative writer deeply inspired by the traditions of Hindi literature and Indian nature poetry.
+# Prompt for Sarvam-1 - Simple, direct creative writing
+CREATIVE_PROMPT = """{context_str}
 
-You will receive:
-1. Retrieved context passages from a Hindi literature knowledge base
-2. A user prompt requesting a poem or story
+---
+Write a poem or short story about: {query}
+Language: {language}
 
-Your task is to create an ORIGINAL creative piece inspired by the themes present in the context.
-
-IMPORTANT RULES:
-
-1. The retrieved context is ONLY for thematic inspiration.
-2. DO NOT summarize, explain, or quote the context.
-3. DO NOT mention authors, poets, books, or literary works from the context.
-4. DO NOT repeat sentences or phrases from the context.
-5. The output must be completely original.
-
-Creative Guidelines:
-
-• Extract inspiration from imagery, emotions, philosophy, and themes found in the context.
-• Use vivid descriptions of nature, emotions, and atmosphere.
-• Maintain a style reminiscent of classical Hindi literary prose or poetry.
-• Use metaphors, sensory details, and cultural depth.
-• Write naturally in the same language as the user's query.
-
-Behavior Rules:
-
-If the context contains biographies, explanations, or references to poets:
-→ IGNORE those parts and focus only on nature-related themes.
-
-If the context is weak or irrelevant:
-→ Still generate a creative piece using general themes of nature and Hindi literary style.
-
-Output Requirements:
-
-• Return ONLY the poem or story.
-• Do not explain your reasoning.
-• Do not mention the context.
-• Do not include headings like "Based on the context".
-
-------------------------------------
-
-Context:
-{context_str}
-
-------------------------------------
-
-User Prompt:
-{query}
-
-------------------------------------
-
-Generate an original creative response in {language}."""
+Write only the creative piece, nothing else:"""
 
 UNSUPPORTED_LANG_PROMPT = """I apologize, but I only support the following languages:
 
@@ -197,113 +152,60 @@ class SimpleRAGSystem:
         return self._synthesize_answer(query, context_docs)
 
     def _is_valid_creative_output(self, text: str) -> bool:
-        """Check if output is valid creative writing vs meta-commentary/prompt leakage"""
-        if not text or len(text) < 20:
+        """Check if output is valid creative writing vs prompt/instruction leakage"""
+        if not text or len(text) < 30:
             return False
         
         text_lower = text.lower()
         
-        # Reject if it contains prompt/instruction keywords
-        invalid_patterns = [
-            # Apologies/inability
-            "i cannot", "i can't", "unable to", "sorry", "apologize",
-            "i'm not", "i am not",
-            # Prompt structure leakage
-            "you are a multilingual", "retrieved context", "user prompt",
-            "your task is to", "important rules", "creative guidelines",
-            "behavior rules", "output requirements", "do not mention",
-            "do not explain", "do not include", "based on the context",
-            # Prompt content leakage
-            "extract inspiration", "vivid descriptions", "sensory details",
-            "classical hindi", "literary prose", "cultural depth",
-            "write naturally", "same language as", "user's query",
-            # Context references
-            "context contains", "biographies", "explanations", "references to poets",
-            "context is weak", "irrelevant", "general themes",
-            # Separator leakage
+        # Hard reject - these indicate broken output
+        hard_reject = [
+            "you are a", "you will receive", "your task is",
+            "important rules", "creative guidelines", "output requirements",
+            "please provide", "at least one", "feel free", "add more",
+            "based on these", "you can use", "when generating",
+            "topic description", "use as input", "generating content",
+            "write only", "write a poem", "write a story", "language:",
+            "context:", "user prompt:", "generate an original",
+            "i cannot", "i can't", "unable to", "sorry",
             "------------------------------------",
-            # Meta commentary
-            "generate an original", "creative response", "creative piece",
-            "thematic inspiration", "summarize", "explain", "or quote",
         ]
         
-        if any(pattern in text_lower for pattern in invalid_patterns):
+        if any(pattern in text_lower for pattern in hard_reject):
             return False
-            
-        # Accept if it looks like creative writing (has multiple lines)
-        lines = [l.strip() for l in text.split('\n') if l.strip()]
-        if len(lines) >= 3:
-            return True
-            
-        # Single line but long enough and no invalid patterns
-        if len(text) > 50:
-            return True
-            
-        return False
+        
+        # Must have multiple lines to be creative writing
+        lines = [l.strip() for l in text.split('\n') if l.strip() and len(l.strip()) > 10]
+        return len(lines) >= 2
 
     def _clean_output(self, text: str) -> str:
-        """Clean up LLM output - remove informal/garbled text and artifacts"""
+        """Clean up LLM output - remove artifacts and garbage"""
         if not text:
             return text
         
-        # Remove common artifacts
+        # Remove artifacts
         text = re.sub(r'</?s>', '', text)
         text = re.sub(r'\[/?INST\]', '', text)
         text = re.sub(r'\[/?SYS\]', '', text)
         text = re.sub(r'\[\d+\]', '', text)
+        text = re.sub(r'-{10,}', '', text)
         
-        # Remove prompt structure leakage
-        text = re.sub(r'You are a multilingual.*', '', text, flags=re.IGNORECASE)
-        text = re.sub(r'You will receive:.*', '', text, flags=re.IGNORECASE)
-        text = re.sub(r'Your task is to.*', '', text, flags=re.IGNORECASE)
-        text = re.sub(r'IMPORTANT RULES:.*', '', text, flags=re.IGNORECASE)
-        text = re.sub(r'Creative Guidelines:.*', '', text, flags=re.IGNORECASE)
-        text = re.sub(r'Behavior Rules:.*', '', text, flags=re.IGNORECASE)
-        text = re.sub(r'Output Requirements:.*', '', text, flags=re.IGNORECASE)
-        text = re.sub(r'Context:.*', '', text, flags=re.IGNORECASE)
-        text = re.sub(r'User Prompt:.*', '', text, flags=re.IGNORECASE)
-        text = re.sub(r'Generate an original.*', '', text, flags=re.IGNORECASE)
-        text = re.sub(r'-{10,}', '', text)  # Remove separator lines
-        
-        # Remove meta-commentary lines
+        # Remove lines that are clearly not creative content
         lines = text.split('\n')
-        cleaned_lines = []
-        
+        cleaned = []
         for line in lines:
             line = line.strip()
-            line_lower = line.lower()
-            
-            # Skip meta-commentary patterns
-            skip_patterns = [
-                # Apologies/inability
-                "i'm not good", "i cannot", "i can't", "i am unable",
-                "sorry", "apologize",
-                # Prompt leakage
-                "you are a", "you will receive", "your task is",
-                "important rules", "creative guidelines", "behavior rules",
-                "output requirements", "do not mention", "do not explain",
-                "based on the context", "thematic inspiration",
-                # Instructions
-                "extract inspiration", "use vivid", "maintain a style",
-                "write naturally", "generate a creative",
-                # Meta
-                "fetched documents", "retrieved context",
-            ]
-            if any(marker in line_lower for marker in skip_patterns):
+            # Skip very short lines, emoji lines, instruction-like lines
+            if len(line) < 5:
                 continue
-            if line.count('!') > 2 or line.count('?') > 2 or line.count('~') > 1:
+            if ': )' in line or ':)' in line or line.count('!') > 2:
                 continue
-            if len(line) < 15 and any(c in line for c in ['<3', '><', 'xp', 'xd']):
+            if line.lower().startswith(('please', 'thank', 'sorry', 'note:', 'context:', 'topic:')):
                 continue
-            if line:
-                cleaned_lines.append(line)
+            cleaned.append(line)
         
-        result = '\n'.join(cleaned_lines)
-        
-        if len(result) < len(text) * 0.3:
-            return text
-            
-        return result.strip()
+        result = '\n'.join(cleaned).strip()
+        return result if result else text
 
     def _synthesize_answer(self, query: str, context_docs: List[Dict]) -> str:
         """Synthesize an answer from retrieved documents when LLM is unavailable"""