Update app.py

app.py

@@ -823,47 +823,18 @@ def generate_answer_langchain(
     if not retrieved_docs:
         return "I couldn't find relevant information to answer your question."
     
-    # Natural mode: allow 2 attempts for quality
-    max_attempts = 2
-    
-    llm_answer = None
-    for attempt in range(1, max_attempts + 1):
-        logger.info(f"\n  🤖 LLM Generation Attempt {attempt}/{max_attempts}")
-        llm_answer = generate_llm_answer(query, retrieved_docs, llm_client, attempt)
-        
-        if llm_answer:
-            logger.info(f"  ✅ LLM answer generated successfully")
-            break
-        else:
-            if attempt < max_attempts:
-                logger.warning(f"  → Attempt {attempt}/{max_attempts} failed, retrying...")
-    
-    if not llm_answer:
-        logger.error(f"  ✗ All {max_attempts} LLM attempts failed")
-        
-        # Try scaffold-and-polish as fallback
-        try:
-            logger.info("  → Attempting scaffold-and-polish using PHI model")
-            polished = scaffold_and_polish(query, retrieved_docs, llm_client)
-            if polished:
-                logger.info("  ✅ Scaffold-and-polish produced an answer")
-                return polished
-        except Exception as e:
-            logger.error(f"  ✗ Scaffold-and-polish error: {e}")
-
-        # Final fallback: extractive templated answer (guaranteed deterministic & FAST)
-        try:
-            logger.info("  → Using extractive fallback generator")
-            fallback = generate_extractive_answer(query, retrieved_docs)
-            if fallback:
-                logger.info("  ✅ Extractive fallback produced an answer")
-                return fallback
-        except Exception as e:
-            logger.error(f"  ✗ Extractive fallback error: {e}")
-
-        return "I apologize, but I'm having trouble generating a response. Please try rephrasing your question or ask something else."
+    # Use extractive answer as PRIMARY method - reliable and high-quality
+    # Small LLMs (DistilGPT2) produce nonsensical output on CPU
+    logger.info("  → Using extractive answer generator (primary method)")
+    try:
+        extractive_answer = generate_extractive_answer(query, retrieved_docs)
+        if extractive_answer:
+            logger.info("  ✅ Extractive answer generated successfully")
+            return extractive_answer
+    except Exception as e:
+        logger.error(f"  ✗ Extractive answer error: {e}")
     
-    return llm_answer
+    return "I apologize, but I'm having trouble generating a response. Please try rephrasing your question or ask something else."
 
 # ============================================================================
 # GRADIO INTERFACE
@@ -889,16 +860,12 @@ def fashion_chatbot(message: str, history: List[List[str]]):
         
         yield f"💭 Generating answer ({len(retrieved_docs)} sources found)..."
         
-        llm_answer = None
-        for attempt in range(1, 3):
-            logger.info(f"\n  🤖 LLM Generation Attempt {attempt}/2")
-            llm_answer = generate_llm_answer(message.strip(), retrieved_docs, llm_client, attempt)
-            
-            if llm_answer:
-                break
+        # Use extractive answer - reliable and high-quality
+        logger.info("  → Generating extractive answer")
+        llm_answer = generate_extractive_answer(message.strip(), retrieved_docs)
         
         if not llm_answer:
-            logger.error(f"  ✗ All LLM attempts failed")
+            logger.error(f"  ✗ Extractive answer generation failed")
             yield "I apologize, but I'm having trouble generating a response. Please try rephrasing your question."
             return