Update app.py

app.py

@@ -297,15 +297,23 @@ def generate_llm_answer(
         logger.error("  → LLM client not initialized")
         return None
     
-    # Build focused context
+    # Build focused context with relevance filtering
     query_lower = query.lower()
     query_words = set(query_lower.split())
     
-    # Score documents by relevance
+    # ANTI-HALLUCINATION: Filter for fashion-relevant documents only
+    fashion_terms = {'wear', 'outfit', 'style', 'fashion', 'clothing', 'color', 'dress', 'fabric'}
+    
     scored_docs = []
     for doc in retrieved_docs[:20]:
         content = doc.page_content.lower()
         doc_words = set(content.split())
+        
+        # Check if document contains fashion terms
+        has_fashion = any(term in content for term in fashion_terms)
+        if not has_fashion:
+            continue  # Skip non-fashion documents
+        
         overlap = len(query_words.intersection(doc_words))
         
         # Boost for verified/curated
@@ -318,6 +326,11 @@ def generate_llm_answer(
         
         scored_docs.append((doc, overlap))
     
+    # If no fashion-relevant docs found, return None
+    if not scored_docs:
+        logger.warning("  ⚠️ No fashion-relevant documents found")
+        return None
+    
     # Sort and take top 8
     scored_docs.sort(key=lambda x: x[1], reverse=True)
     top_docs = [doc[0] for doc in scored_docs[:8]]
@@ -358,14 +371,21 @@ def generate_llm_answer(
     # Create COMPACT T5 prompt to stay under 512 tokens (critical!)
     model_type = CONFIG.get("model_type", "t5")
     
-    # T5 format - simple and effective to minimize tokens
-    # Keep prompt minimal to leave room for generation
-    user_prompt = f"""Fashion Question: {query}
+    # T5 format - with explicit constraints to prevent hallucination
+    user_prompt = f"""You are a fashion expert. Answer ONLY about fashion, clothing, and style.
+
+Question: {query}
 
-Relevant Fashion Tips:
+Fashion Knowledge:
 {context_text[:600]}
 
-Provide detailed fashion advice (200-400 words):"""
+Rules:
+- Answer ONLY using the fashion knowledge provided
+- Focus on clothing, outfits, colors, fabrics, and styling
+- DO NOT mention: politics, history, wars, empires, architecture
+- If unsure, say "I don't have enough information"
+
+Fashion Answer:"""
 
     try:
         logger.info(f"  → Calling {CONFIG['llm_model']} (temp={temperature}, tokens={max_new_tokens})...")
@@ -452,7 +472,7 @@ def generate_answer_langchain(
     
     # Step 3: If all attempts fail, return error
     if not llm_answer:
-        logger.error(f"  ✗ All 4 LLM attempts failed")
+        logger.error(f"  ✗ All 2 LLM attempts failed")
         return "I apologize, but I'm having trouble generating a response. Please try rephrasing your question or ask something else."
     
     return llm_answer
@@ -470,10 +490,45 @@ def fashion_chatbot(message: str, history: List[List[str]]):
             yield "Please ask a fashion-related question!"
             return
         
-        # Show searching indicator
+        # ANTI-HALLUCINATION: Validate if question is fashion-related
+        query_lower = message.strip().lower()
+        fashion_keywords = [
+            'wear', 'outfit', 'dress', 'style', 'fashion', 'clothing', 'clothes',
+            'color', 'match', 'look', 'shirt', 'pants', 'shoes', 'accessory',
+            'wardrobe', 'fit', 'fabric', 'pattern', 'casual', 'formal', 'seasonal',
+            'wedding', 'meeting', 'interview', 'date', 'party', 'jeans', 'suit',
+            'skirt', 'jacket', 'coat', 'sweater', 'blouse', 'tie', 'scarf', 'boots',
+            'hat', 'bag', 'purse', 'jewelry', 'necklace', 'bracelet', 'watch'
+        ]
+        
+        # Reject obviously non-fashion questions FIRST (higher priority)
+        non_fashion_indicators = [
+            'crisis', 'collapse', 'empire', 'war', 'politics', 'economy',
+            'architecture', 'building', 'nebula', 'space', 'republic',
+            'soviet', 'ottoman', 'history', 'government', 'president', 'designed',
+            'architect', 'eastern', 'western', 'communist', 'russia', 'political',
+            'military', 'sapphire crisis', 'who designed', 'what caused'
+        ]
+        
+        has_non_fashion = any(indicator in query_lower for indicator in non_fashion_indicators)
+        
+        # STRICT CHECK: If non-fashion detected, reject immediately
+        if has_non_fashion:
+            logger.info(f"❌ Non-fashion query rejected: {message.strip()}")
+            yield "I'm a fashion advisor and can only answer questions about clothing, style, and fashion. Please ask me about outfits, styling, colors, or wardrobe advice!"
+            return
+        
+        # Check if query contains fashion keywords
+        is_fashion_query = any(keyword in query_lower for keyword in fashion_keywords)
+        
+        if not is_fashion_query:
+            yield "I'm a fashion advisor and can only answer questions about clothing, style, and fashion. Please ask me about outfits, styling, colors, or wardrobe advice!"
+            return
+        
+        # Show searching indicator (only for valid fashion queries)
         yield "🔍 Searching fashion knowledge..."
         
-        # Retrieve documents
+        # Retrieve documents (only after validation passes)
         retrieved_docs, confidence = retrieve_knowledge_langchain(
             message.strip(),
             vectorstore,
@@ -484,6 +539,11 @@ def fashion_chatbot(message: str, history: List[List[str]]):
             yield "I couldn't find relevant information to answer your question."
             return
         
+        # ANTI-HALLUCINATION: Check retrieval quality
+        if confidence < 0.35:
+            yield "I don't have enough reliable information about this specific topic. Could you rephrase or ask about common fashion topics like outfit recommendations, color matching, or styling advice?"
+            return
+        
         # Show generating indicator
         yield f"💭 Generating answer ({len(retrieved_docs)} sources found)..."
         
@@ -494,12 +554,36 @@ def fashion_chatbot(message: str, history: List[List[str]]):
             llm_answer = generate_llm_answer(message.strip(), retrieved_docs, llm_client, attempt)
             
             if llm_answer:
-                break
+                # ANTI-HALLUCINATION: Validate answer relevance
+                answer_lower = llm_answer.lower()
+                
+                # Check for hallucination indicators
+                hallucination_markers = [
+                    'empire', 'ottoman', 'soviet', 'russia', 'collapse', 'crisis',
+                    'republic', 'communist', 'nebula', 'architecture', 'political',
+                    'government', 'war', 'military', 'economic'
+                ]
+                
+                has_hallucination = any(marker in answer_lower for marker in hallucination_markers)
+                
+                # Check if answer contains fashion terms
+                fashion_terms = [
+                    'wear', 'outfit', 'style', 'clothing', 'fabric', 'color',
+                    'match', 'fit', 'look', 'fashion', 'dress', 'suit'
+                ]
+                has_fashion_content = any(term in answer_lower for term in fashion_terms)
+                
+                if has_hallucination or not has_fashion_content:
+                    logger.warning(f"  ⚠️ Hallucination detected in attempt {attempt}, retrying...")
+                    llm_answer = None
+                    continue
+                else:
+                    break
         
         # If LLM fails, show error
         if not llm_answer:
-            logger.error(f"  ✗ All LLM attempts failed")
-            yield "I apologize, but I'm having trouble generating a response. Please try rephrasing your question."
+            logger.error(f"  ✗ All LLM attempts failed or produced hallucinations")
+            yield "I apologize, but I'm having trouble generating a reliable fashion answer. Please ask about specific fashion topics like outfit recommendations, color coordination, or styling tips."
             return
         
         # Stream the answer word by word for natural flow