Upload folder using huggingface_hub

src/agent/insights_generator.py

@@ -2,7 +2,10 @@
 Restaurant Insights Generator - EXPANDED VERSION
 Generates role-specific insights for Chef and Manager personas.
 
-UPDATED: Now uses TOP 20 items/aspects instead of TOP 10 for more comprehensive insights.
+UPDATED v3: 
+- New sentiment scale (>= 0.6 positive, 0-0.59 neutral, < 0 negative)
+- Clearer guidance on strengths vs concerns
+- Top 20 items/aspects for comprehensive insights
 """
 
 import json
@@ -14,7 +17,10 @@ class InsightsGenerator:
     """
     Generates actionable insights for different restaurant roles.
     
-    UPDATED: Expanded to use top 20 menu items and aspects for better recommendations.
+    UPDATED: 
+    - New sentiment thresholds (0.6/0 instead of 0.3/-0.3)
+    - Expanded to use top 20 menu items and aspects
+    - Clearer mapping of sentiment to strengths/concerns
     """
     
     def __init__(self, client, model: str = "claude-sonnet-4-20250514"):
@@ -109,6 +115,11 @@ MENU PERFORMANCE (Top items by customer mentions):
 FOOD-RELATED ASPECTS:
 {aspect_summary}
 
+SENTIMENT SCALE:
+- 🟢 POSITIVE (0.6 to 1.0): Customers love this - highlight as a STRENGTH
+- 🟡 NEUTRAL (0.0 to 0.59): Mixed or average feedback - room for improvement
+- 🔴 NEGATIVE (below 0): Customers complained - flag as a CONCERN
+
 YOUR TASK:
 Generate actionable insights specifically for the HEAD CHEF. Focus on:
 - Food quality and taste
@@ -121,30 +132,32 @@ Generate actionable insights specifically for the HEAD CHEF. Focus on:
 
 CRITICAL RULES:
 1. Focus ONLY on food/kitchen topics
-2. Be specific with evidence from reviews
-3. Make recommendations actionable
-4. Reference specific menu items by name
-5. Output ONLY valid JSON, no other text
+2. STRENGTHS should come from items/aspects with sentiment >= 0.6 (🟢 positive)
+3. CONCERNS should come from items/aspects with sentiment < 0 (🔴 negative)
+4. Be specific with evidence from reviews
+5. Make recommendations actionable
+6. Reference specific menu items by name
+7. Output ONLY valid JSON, no other text
 
 OUTPUT FORMAT (JSON):
 {{
   "summary": "2-3 sentence executive summary covering overall kitchen performance",
   "strengths": [
-    "Specific strength 1 with menu item example",
-    "Specific strength 2 with menu item example",
-    "Specific strength 3 with menu item example",
-    "Specific strength 4 with menu item example",
-    "Specific strength 5 with menu item example"
+    "Specific strength 1 - reference a 🟢 positive item with sentiment >= 0.6",
+    "Specific strength 2 - reference a 🟢 positive item with sentiment >= 0.6",
+    "Specific strength 3 - reference a 🟢 positive item with sentiment >= 0.6",
+    "Specific strength 4 - reference a 🟢 positive item with sentiment >= 0.6",
+    "Specific strength 5 - reference a 🟢 positive item with sentiment >= 0.6"
   ],
   "concerns": [
-    "Specific concern 1 with evidence",
-    "Specific concern 2 with evidence",
-    "Specific concern 3 with evidence"
+    "Specific concern 1 - reference a 🔴 negative item with sentiment < 0",
+    "Specific concern 2 - reference a 🔴 negative item with sentiment < 0",
+    "Specific concern 3 - reference a 🔴 negative item with sentiment < 0"
   ],
   "recommendations": [
     {{
       "priority": "high",
-      "action": "Specific action to take",
+      "action": "Specific action to fix a negative sentiment item",
       "reason": "Why this matters based on review data",
       "evidence": "Supporting data from reviews"
     }},
@@ -176,14 +189,15 @@ OUTPUT FORMAT (JSON):
 }}
 
 IMPORTANT: 
-- Provide at least 5 strengths and 5 recommendations
+- Provide at least 5 strengths (from 🟢 items) and 5 recommendations
+- If there are no negative items, focus recommendations on improving neutral items
 - Reference actual menu items from the data above
 - Ensure all JSON is properly formatted with no trailing commas
 
 Generate chef insights:"""
         
         return prompt
-    
+
     def _build_manager_prompt(
         self,
         analysis_data: Dict[str, Any],
@@ -204,6 +218,11 @@ OPERATIONAL ASPECTS (All discovered from reviews):
 MENU OVERVIEW (for context):
 {menu_summary}
 
+SENTIMENT SCALE:
+- 🟢 POSITIVE (0.6 to 1.0): Customers love this - highlight as a STRENGTH
+- 🟡 NEUTRAL (0.0 to 0.59): Mixed or average feedback - room for improvement
+- 🔴 NEGATIVE (below 0): Customers complained - flag as a CONCERN
+
 YOUR TASK:
 Generate actionable insights specifically for the RESTAURANT MANAGER. Focus on:
 - Service quality and speed
@@ -217,30 +236,32 @@ Generate actionable insights specifically for the RESTAURANT MANAGER. Focus on:
 
 CRITICAL RULES:
 1. Focus ONLY on operations/service topics
-2. Be specific with evidence from reviews
-3. Make recommendations actionable
-4. Reference specific aspects by name
-5. Output ONLY valid JSON, no other text
+2. STRENGTHS should come from aspects with sentiment >= 0.6 (🟢 positive)
+3. CONCERNS should come from aspects with sentiment < 0 (🔴 negative)
+4. Be specific with evidence from reviews
+5. Make recommendations actionable
+6. Reference specific aspects by name
+7. Output ONLY valid JSON, no other text
 
 OUTPUT FORMAT (JSON):
 {{
   "summary": "2-3 sentence executive summary covering overall operations",
   "strengths": [
-    "Specific operational strength 1 with evidence",
-    "Specific operational strength 2 with evidence",
-    "Specific operational strength 3 with evidence",
-    "Specific operational strength 4 with evidence",
-    "Specific operational strength 5 with evidence"
+    "Specific operational strength 1 - reference a 🟢 positive aspect with sentiment >= 0.6",
+    "Specific operational strength 2 - reference a 🟢 positive aspect with sentiment >= 0.6",
+    "Specific operational strength 3 - reference a 🟢 positive aspect with sentiment >= 0.6",
+    "Specific operational strength 4 - reference a 🟢 positive aspect with sentiment >= 0.6",
+    "Specific operational strength 5 - reference a 🟢 positive aspect with sentiment >= 0.6"
   ],
   "concerns": [
-    "Specific operational concern 1 with evidence",
-    "Specific operational concern 2 with evidence",
-    "Specific operational concern 3 with evidence"
+    "Specific operational concern 1 - reference a 🔴 negative aspect with sentiment < 0",
+    "Specific operational concern 2 - reference a 🔴 negative aspect with sentiment < 0",
+    "Specific operational concern 3 - reference a 🔴 negative aspect with sentiment < 0"
   ],
   "recommendations": [
     {{
       "priority": "high",
-      "action": "Specific action to take",
+      "action": "Specific action to fix a negative sentiment aspect",
       "reason": "Why this matters based on review data",
       "evidence": "Supporting data from reviews"
     }},
@@ -272,14 +293,15 @@ OUTPUT FORMAT (JSON):
 }}
 
 IMPORTANT: 
-- Provide at least 5 strengths and 5 recommendations
+- Provide at least 5 strengths (from 🟢 aspects) and 5 recommendations
+- If there are no negative aspects, focus recommendations on improving neutral aspects
 - Reference actual aspects from the data above
 - Ensure all JSON is properly formatted with no trailing commas
 
 Generate manager insights:"""
         
         return prompt
-    
+
     def _summarize_menu_data(
         self, 
         analysis_data: Dict[str, Any],
@@ -289,7 +311,7 @@ Generate manager insights:"""
         """
         Summarize menu analysis for prompts.
         
-        EXPANDED: Now uses top 20 food items and top 10 drinks (was 10/5).
+        UPDATED: New sentiment thresholds (0.6/0 instead of 0.3/-0.3)
         """
         menu_data = analysis_data.get('menu_analysis', {})
         food_items = menu_data.get('food_items', [])[:max_food]
@@ -302,8 +324,8 @@ Generate manager insights:"""
             for item in food_items:
                 sentiment = item.get('sentiment', 0)
                 mentions = item.get('mention_count', 0)
-                # Add sentiment indicator
-                indicator = "🟢" if sentiment > 0.3 else "🟡" if sentiment > -0.3 else "🔴"
+                # NEW thresholds: >= 0.6 positive, >= 0 neutral, < 0 negative
+                indicator = "🟢" if sentiment >= 0.6 else "🟡" if sentiment >= 0 else "🔴"
                 summary.append(f"  {indicator} {item.get('name', 'unknown')}: sentiment {sentiment:+.2f}, {mentions} mentions")
         
         if drinks:
@@ -311,7 +333,8 @@ Generate manager insights:"""
             for drink in drinks:
                 sentiment = drink.get('sentiment', 0)
                 mentions = drink.get('mention_count', 0)
-                indicator = "🟢" if sentiment > 0.3 else "🟡" if sentiment > -0.3 else "🔴"
+                # NEW thresholds: >= 0.6 positive, >= 0 neutral, < 0 negative
+                indicator = "🟢" if sentiment >= 0.6 else "🟡" if sentiment >= 0 else "🔴"
                 summary.append(f"  {indicator} {drink.get('name', 'unknown')}: sentiment {sentiment:+.2f}, {mentions} mentions")
         
         # Add overall stats
@@ -330,7 +353,7 @@ Generate manager insights:"""
         """
         Summarize aspect analysis for prompts.
         
-        EXPANDED: Now uses top 20 aspects (was 10).
+        UPDATED: New sentiment thresholds (0.6/0 instead of 0.3/-0.3)
         """
         aspect_data = analysis_data.get('aspect_analysis', {})
         aspects = aspect_data.get('aspects', [])
@@ -354,7 +377,8 @@ Generate manager insights:"""
         for aspect in aspects:
             sentiment = aspect.get('sentiment', 0)
             mentions = aspect.get('mention_count', 0)
-            indicator = "🟢" if sentiment > 0.3 else "🟡" if sentiment > -0.3 else "🔴"
+            # NEW thresholds: >= 0.6 positive, >= 0 neutral, < 0 negative
+            indicator = "🟢" if sentiment >= 0.6 else "🟡" if sentiment >= 0 else "🔴"
             summary.append(f"  {indicator} {aspect.get('name', 'unknown')}: sentiment {sentiment:+.2f}, {mentions} mentions")
         
         # Add total count

src/agent/unified_analyzer.py

@@ -1,6 +1,11 @@
 """
 Unified Review Analyzer - Single-pass extraction
 Extracts menu items, aspects, and sentiment in ONE API call per batch
+
+UPDATED: New sentiment scale
+- Positive: >= 0.6
+- Neutral: 0 to 0.59
+- Negative: < 0
 """
 
 from typing import List, Dict, Any
@@ -23,7 +28,7 @@ class UnifiedReviewAnalyzer:
     - Customer aspects (service, ambience, etc.)
     - Sentiment for each
     
-    Reduces API calls by 3x!
+    Reduces API calls by 3x compared to separate extraction!
     """
     
     def __init__(self, client: Anthropic, model: str):
@@ -41,9 +46,13 @@ class UnifiedReviewAnalyzer:
         
         Returns:
             {
-                "menu_items": {...},
-                "aspects": {...},
-                "overall_stats": {...}
+                "menu_analysis": {
+                    "food_items": [...],
+                    "drinks": [...]
+                },
+                "aspect_analysis": {
+                    "aspects": [...]
+                }
             }
         """
         print(f"🚀 Unified analysis: {len(reviews)} reviews in batches of {batch_size}...")
@@ -63,65 +72,73 @@ class UnifiedReviewAnalyzer:
             try:
                 batch_result = self._analyze_batch(batch, restaurant_name, start_index=i)
                 
-                # Merge menu items
+                # Merge food items
                 for item in batch_result.get('food_items', []):
-                    name = item['name']
+                    name = item.get('name', '').lower()
+                    if not name:
+                        continue
                     if name in all_food_items:
-                        all_food_items[name]['mention_count'] += item['mention_count']
+                        all_food_items[name]['mention_count'] += item.get('mention_count', 1)
                         all_food_items[name]['related_reviews'].extend(item.get('related_reviews', []))
+                        # Average sentiment
                         old_sent = all_food_items[name]['sentiment']
-                        new_sent = item['sentiment']
-                        all_food_items[name]['sentiment'] = (old_sent + new_sent) / 2
+                        new_sent = item.get('sentiment', 0)
+                        old_count = all_food_items[name]['mention_count'] - item.get('mention_count', 1)
+                        new_count = item.get('mention_count', 1)
+                        all_food_items[name]['sentiment'] = (old_sent * old_count + new_sent * new_count) / (old_count + new_count)
                     else:
                         all_food_items[name] = item
                 
                 # Merge drinks
-                for drink in batch_result.get('drinks', []):
-                    name = drink['name']
+                for item in batch_result.get('drinks', []):
+                    name = item.get('name', '').lower()
+                    if not name:
+                        continue
                     if name in all_drinks:
-                        all_drinks[name]['mention_count'] += drink['mention_count']
-                        all_drinks[name]['related_reviews'].extend(drink.get('related_reviews', []))
+                        all_drinks[name]['mention_count'] += item.get('mention_count', 1)
+                        all_drinks[name]['related_reviews'].extend(item.get('related_reviews', []))
                         old_sent = all_drinks[name]['sentiment']
-                        new_sent = drink['sentiment']
-                        all_drinks[name]['sentiment'] = (old_sent + new_sent) / 2
+                        new_sent = item.get('sentiment', 0)
+                        old_count = all_drinks[name]['mention_count'] - item.get('mention_count', 1)
+                        new_count = item.get('mention_count', 1)
+                        all_drinks[name]['sentiment'] = (old_sent * old_count + new_sent * new_count) / (old_count + new_count)
                     else:
-                        all_drinks[name] = drink
+                        all_drinks[name] = item
                 
                 # Merge aspects
                 for aspect in batch_result.get('aspects', []):
-                    name = aspect['name']
+                    name = aspect.get('name', '').lower()
+                    if not name:
+                        continue
                     if name in all_aspects:
-                        all_aspects[name]['mention_count'] += aspect['mention_count']
+                        all_aspects[name]['mention_count'] += aspect.get('mention_count', 1)
                         all_aspects[name]['related_reviews'].extend(aspect.get('related_reviews', []))
                         old_sent = all_aspects[name]['sentiment']
-                        new_sent = aspect['sentiment']
-                        all_aspects[name]['sentiment'] = (old_sent + new_sent) / 2
+                        new_sent = aspect.get('sentiment', 0)
+                        old_count = all_aspects[name]['mention_count'] - aspect.get('mention_count', 1)
+                        new_count = aspect.get('mention_count', 1)
+                        all_aspects[name]['sentiment'] = (old_sent * old_count + new_sent * new_count) / (old_count + new_count)
                     else:
                         all_aspects[name] = aspect
-                
+                        
             except Exception as e:
-                print(f"   ⚠️  Batch {batch_num} failed: {e}")
+                print(f"   ⚠️ Batch {batch_num} error: {e}")
                 continue
         
-        # Convert to lists and sort
-        food_items_list = sorted(list(all_food_items.values()), 
-                                key=lambda x: x['mention_count'], reverse=True)
-        drinks_list = sorted(list(all_drinks.values()), 
-                            key=lambda x: x['mention_count'], reverse=True)
-        aspects_list = sorted(list(all_aspects.values()), 
-                             key=lambda x: x['mention_count'], reverse=True)
+        # Convert to lists and sort by mention count
+        food_list = sorted(all_food_items.values(), key=lambda x: x.get('mention_count', 0), reverse=True)
+        drinks_list = sorted(all_drinks.values(), key=lambda x: x.get('mention_count', 0), reverse=True)
+        aspects_list = sorted(all_aspects.values(), key=lambda x: x.get('mention_count', 0), reverse=True)
         
-        print(f"✅ Discovered: {len(food_items_list)} food + {len(drinks_list)} drinks + {len(aspects_list)} aspects")
+        print(f"✅ Discovered: {len(food_list)} food + {len(drinks_list)} drinks + {len(aspects_list)} aspects")
         
         return {
             "menu_analysis": {
-                "food_items": food_items_list,
-                "drinks": drinks_list,
-                "total_extracted": len(food_items_list) + len(drinks_list)
+                "food_items": food_list,
+                "drinks": drinks_list
             },
             "aspect_analysis": {
-                "aspects": aspects_list,
-                "total_aspects": len(aspects_list)
+                "aspects": aspects_list
             }
         }
     
@@ -131,14 +148,15 @@ class UnifiedReviewAnalyzer:
         restaurant_name: str,
         start_index: int = 0
     ) -> Dict[str, Any]:
-        """Analyze a single batch - extract EVERYTHING in one call."""
+        """Analyze a single batch of reviews."""
+        
         prompt = self._build_unified_prompt(reviews, restaurant_name, start_index)
         
         try:
             response = call_claude_with_retry(
                 client=self.client,
                 model=self.model,
-                max_tokens=4000,  # Reduced since we're not returning full text
+                max_tokens=4000,
                 temperature=0.3,
                 messages=[{"role": "user", "content": prompt}]
             )
@@ -150,7 +168,7 @@ class UnifiedReviewAnalyzer:
             try:
                 data = json.loads(result_text)
             except json.JSONDecodeError as e:
-                print(f"   ⚠️  JSON parse error: {e}")
+                print(f"   ⚠️ JSON parse error: {e}")
                 return {"food_items": [], "drinks": [], "aspects": []}
             
             # Post-process: Add full review text back using indices
@@ -163,99 +181,13 @@ class UnifiedReviewAnalyzer:
             print(f"❌ Extraction error: {e}")
             return {"food_items": [], "drinks": [], "aspects": []}
     
-    def _map_reviews_to_items(
-        self,
-        data: Dict[str, Any],
-        reviews: List[str],
-        start_index: int
-    ) -> Dict[str, Any]:
-        """
-        Map review indices back to full review text.
-        
-        Claude returns just indices to avoid JSON breaking.
-        We add the full text back here.
-        """
-        # Process food items
-        for item in data.get('food_items', []):
-            review_indices = item.get('related_reviews', [])
-            if isinstance(review_indices, list) and review_indices:
-                # If it's already in full format, skip
-                if isinstance(review_indices[0], dict):
-                    continue
-                
-                # Map indices to full reviews
-                full_reviews = []
-                for idx in review_indices:
-                    if isinstance(idx, int) and 0 <= idx < len(reviews):
-                        full_reviews.append({
-                            "review_index": start_index + idx,
-                            "review_text": reviews[idx],
-                            "sentiment_context": reviews[idx][:200]  # First 200 chars as context
-                        })
-                
-                item['related_reviews'] = full_reviews
-        
-        # Process drinks
-        for drink in data.get('drinks', []):
-            review_indices = drink.get('related_reviews', [])
-            if isinstance(review_indices, list) and review_indices:
-                if isinstance(review_indices[0], dict):
-                    continue
-                
-                full_reviews = []
-                for idx in review_indices:
-                    if isinstance(idx, int) and 0 <= idx < len(reviews):
-                        full_reviews.append({
-                            "review_index": start_index + idx,
-                            "review_text": reviews[idx],
-                            "sentiment_context": reviews[idx][:200]
-                        })
-                
-                drink['related_reviews'] = full_reviews
-        
-        # Process aspects
-        for aspect in data.get('aspects', []):
-            review_indices = aspect.get('related_reviews', [])
-            if isinstance(review_indices, list) and review_indices:
-                if isinstance(review_indices[0], dict):
-                    continue
-                
-                full_reviews = []
-                for idx in review_indices:
-                    if isinstance(idx, int) and 0 <= idx < len(reviews):
-                        full_reviews.append({
-                            "review_index": start_index + idx,
-                            "review_text": reviews[idx],
-                            "sentiment_context": reviews[idx][:200]
-                        })
-                
-                aspect['related_reviews'] = full_reviews
-        
-        return data
-    
-    def _normalize_data(self, data: Dict[str, Any]) -> Dict[str, Any]:
-        """Normalize all names to lowercase."""
-        for item in data.get('food_items', []):
-            if 'name' in item:
-                item['name'] = item['name'].lower()
-        
-        for drink in data.get('drinks', []):
-            if 'name' in drink:
-                drink['name'] = drink['name'].lower()
-        
-        for aspect in data.get('aspects', []):
-            if 'name' in aspect:
-                aspect['name'] = aspect['name'].lower()
-        
-        return data
-    
     def _build_unified_prompt(
         self,
         reviews: List[str],
         restaurant_name: str,
         start_index: int
     ) -> str:
-        """Build unified extraction prompt."""
+        """Build unified extraction prompt with NEW SENTIMENT SCALE."""
         numbered_reviews = []
         for i, review in enumerate(reviews):
             numbered_reviews.append(f"[Review {i}]: {review}")
@@ -272,20 +204,30 @@ YOUR TASK - Extract THREE things simultaneously:
 2. **ASPECTS** (what customers care about: service, ambience, etc.)
 3. **SENTIMENT** for each
 
+SENTIMENT SCALE (IMPORTANT):
+- **Positive (0.6 to 1.0):** Customer clearly enjoyed/praised this item or aspect
+- **Neutral (0.0 to 0.59):** Mixed feelings, okay but not exceptional, or simply mentioned without strong opinion
+- **Negative (-1.0 to -0.01):** Customer complained, criticized, or expressed disappointment
+
+Examples:
+- "The pasta was absolutely divine!" → 0.85 (Positive)
+- "The pasta was decent, nothing special" → 0.3 (Neutral)
+- "The pasta was undercooked and bland" → -0.6 (Negative)
+
 RULES:
 
 **MENU ITEMS:**
 - Specific items only: "salmon sushi", "miso soup", "sake"
 - Separate food from drinks
 - Lowercase names
-- Calculate sentiment per item
+- Calculate sentiment per item using the scale above
 
 **ASPECTS:**
 - What customers discuss: "service speed", "food quality", "ambience", "value"
 - Be specific: "service speed" not just "service"
 - Cuisine-specific welcome: "freshness", "authenticity", "presentation"
 - Lowercase names
-- Calculate sentiment per aspect
+- Calculate sentiment per aspect using the scale above
 
 **REVIEW LINKING:**
 - For EACH item/aspect, list which review NUMBERS mention it
@@ -298,7 +240,7 @@ OUTPUT (JSON) - IMPORTANT: Return ONLY review indices, NOT full text:
     {{
       "name": "salmon aburi sushi",
       "mention_count": 2,
-      "sentiment": 0.9,
+      "sentiment": 0.85,
       "category": "sushi",
       "related_reviews": [0, 5]
     }}
@@ -307,7 +249,7 @@ OUTPUT (JSON) - IMPORTANT: Return ONLY review indices, NOT full text:
     {{
       "name": "sake",
       "mention_count": 1,
-      "sentiment": 0.8,
+      "sentiment": 0.7,
       "category": "alcohol",
       "related_reviews": [3]
     }}
@@ -316,7 +258,7 @@ OUTPUT (JSON) - IMPORTANT: Return ONLY review indices, NOT full text:
     {{
       "name": "service speed",
       "mention_count": 3,
-      "sentiment": 0.6,
+      "sentiment": 0.65,
       "description": "How quickly food arrives",
       "related_reviews": [1, 2, 7]
     }}
@@ -328,7 +270,68 @@ CRITICAL:
 - DO NOT include review text or quotes
 - This prevents JSON parsing errors and saves tokens
 - Output ONLY valid JSON, no other text
+- Use the sentiment scale: >= 0.6 positive, 0-0.59 neutral, < 0 negative
 
 Extract everything:"""
         
-        return prompt
+        return prompt
+    
+    def _map_reviews_to_items(
+        self,
+        data: Dict[str, Any],
+        reviews: List[str],
+        start_index: int
+    ) -> Dict[str, Any]:
+        """
+        Map review indices back to full review text.
+        
+        Claude returns just indices to avoid JSON breaking.
+        We add the full text back here.
+        """
+        for item in data.get('food_items', []):
+            indices = item.get('related_reviews', [])
+            item['related_reviews'] = []
+            for idx in indices:
+                if isinstance(idx, int) and 0 <= idx < len(reviews):
+                    item['related_reviews'].append({
+                        'review_index': start_index + idx,
+                        'review_text': reviews[idx]
+                    })
+        
+        for item in data.get('drinks', []):
+            indices = item.get('related_reviews', [])
+            item['related_reviews'] = []
+            for idx in indices:
+                if isinstance(idx, int) and 0 <= idx < len(reviews):
+                    item['related_reviews'].append({
+                        'review_index': start_index + idx,
+                        'review_text': reviews[idx]
+                    })
+        
+        for aspect in data.get('aspects', []):
+            indices = aspect.get('related_reviews', [])
+            aspect['related_reviews'] = []
+            for idx in indices:
+                if isinstance(idx, int) and 0 <= idx < len(reviews):
+                    aspect['related_reviews'].append({
+                        'review_index': start_index + idx,
+                        'review_text': reviews[idx]
+                    })
+        
+        return data
+    
+    def _normalize_data(self, data: Dict[str, Any]) -> Dict[str, Any]:
+        """Normalize names to lowercase."""
+        for item in data.get('food_items', []):
+            if 'name' in item:
+                item['name'] = item['name'].lower()
+        
+        for drink in data.get('drinks', []):
+            if 'name' in drink:
+                drink['name'] = drink['name'].lower()
+        
+        for aspect in data.get('aspects', []):
+            if 'name' in aspect:
+                aspect['name'] = aspect['name'].lower()
+        
+        return data

src/ui/gradio_app.py

@@ -5,11 +5,17 @@ Professional UI with cards, plain English summaries, polished layout
 Hackathon: Anthropic MCP 1st Birthday - Track 2 (Productivity)
 Author: Tushar Pingle
 
-VERSION 4.0 FIXES:
-1. Fixed Q&A "proxies" error with Anthropic SDK
-2. Fixed PDF download functionality
-3. Improved summaries to be more detailed and actionable
-4. Multi-platform support (OpenTable + Google Maps)
+VERSION 4.1 UPDATES:
+1. NEW SENTIMENT SCALE:
+   - 🟢 Positive: >= 0.6 (customers clearly enjoyed/praised)
+   - 🟡 Neutral: 0 to 0.59 (mixed feelings, average, okay)
+   - 🔴 Negative: < 0 (complaints, criticism, disappointment)
+
+2. Updated all thresholds throughout the app for consistency
+3. Improved Q&A prompt for balanced answers (pros AND cons)
+4. Fixed PDF style conflicts with RIA prefix
+5. Fixed Q&A "proxies" error with Anthropic SDK
+6. Multi-platform support (OpenTable + Google Maps)
 """
 
 import gradio as gr
@@ -285,13 +291,13 @@ def generate_trend_insight(trend_data: List[Dict], restaurant_name: str) -> str:
     
     insight = f"**{restaurant_name}** has an average rating of **{avg_rating:.1f} stars** "
     
-    if avg_sentiment > 0.3:
+    if avg_sentiment >= 0.6:
         insight += "with **positive sentiment**. "
         if avg_rating >= 4.0:
             insight += "✅ Ratings and sentiment are aligned!"
         else:
             insight += "🤔 Sentiment is positive but ratings are moderate."
-    elif avg_sentiment < -0.1:
+    elif avg_sentiment < 0:
         insight += "but with **concerning sentiment**. "
         if avg_rating >= 4.0:
             insight += "⚠️ **Warning:** High ratings but negative sentiment detected."
@@ -377,11 +383,10 @@ def translate_menu_performance(menu: dict, restaurant_name: str) -> str:
     if not all_items:
         return f"*No menu data available for {restaurant_name} yet.*"
     
-    # Count categories
-    stars = len([i for i in all_items if i.get('sentiment', 0) > 0.5])
-    good = len([i for i in all_items if 0.2 < i.get('sentiment', 0) <= 0.5])
-    mixed = len([i for i in all_items if -0.2 <= i.get('sentiment', 0) <= 0.2])
-    concerns = len([i for i in all_items if i.get('sentiment', 0) < -0.2])
+    # Count categories - NEW thresholds: >= 0.6 positive, 0-0.59 neutral, < 0 negative
+    stars = len([i for i in all_items if i.get('sentiment', 0) >= 0.6])
+    good = len([i for i in all_items if 0 <= i.get('sentiment', 0) < 0.6])
+    concerns = len([i for i in all_items if i.get('sentiment', 0) < 0])
     
     # Simple summary
     summary = f"""### 🍽️ Menu Overview for {restaurant_name}
@@ -390,10 +395,9 @@ def translate_menu_performance(menu: dict, restaurant_name: str) -> str:
 
 | Category | Count |
 |----------|-------|
-| 🌟 Customer Favorites | {stars} |
-| ✅ Performing Well | {good} |
-| 🟡 Mixed Reviews | {mixed} |
-| ⚠️ Needs Attention | {concerns} |
+| 🟢 Positive (≥0.6) | {stars} |
+| 🟡 Neutral (0 to 0.59) | {good} |
+| 🔴 Negative (<0) | {concerns} |
 
 👇 **Select an item from the dropdown below to see detailed customer feedback.**
 """
@@ -410,10 +414,10 @@ def translate_aspect_performance(aspects: dict, restaurant_name: str) -> str:
     if not aspect_list:
         return f"*No aspect data available for {restaurant_name} yet.*"
     
-    # Count categories
-    strengths = len([a for a in aspect_list if a.get('sentiment', 0) > 0.3])
-    neutral = len([a for a in aspect_list if -0.3 <= a.get('sentiment', 0) <= 0.3])
-    weaknesses = len([a for a in aspect_list if a.get('sentiment', 0) < -0.3])
+    # Count categories - NEW thresholds: >= 0.6 positive, 0-0.59 neutral, < 0 negative
+    strengths = len([a for a in aspect_list if a.get('sentiment', 0) >= 0.6])
+    neutral = len([a for a in aspect_list if 0 <= a.get('sentiment', 0) < 0.6])
+    weaknesses = len([a for a in aspect_list if a.get('sentiment', 0) < 0])
     
     # Simple summary
     summary = f"""### 📊 Customer Experience Overview for {restaurant_name}
@@ -422,9 +426,9 @@ def translate_aspect_performance(aspects: dict, restaurant_name: str) -> str:
 
 | Category | Count |
 |----------|-------|
-| 💪 Strengths | {strengths} |
-| 🟡 Neutral | {neutral} |
-| 📉 Needs Work | {weaknesses} |
+| 🟢 Strengths (≥0.6) | {strengths} |
+| 🟡 Neutral (0 to 0.59) | {neutral} |
+| 🔴 Weaknesses (<0) | {weaknesses} |
 
 👇 **Select an aspect from the dropdown below to see detailed customer feedback.**
 """
@@ -456,7 +460,8 @@ def generate_chart(items: list, title: str) -> Optional[str]:
         names = [f"{item.get('name', '?')[:18]} ({item.get('mention_count', 0)})" for item in sorted_items]
         sentiments = [item.get('sentiment', 0) for item in sorted_items]
         
-        colors = [POSITIVE if s > 0.3 else NEUTRAL if s > -0.3 else NEGATIVE for s in sentiments]
+        # NEW thresholds: >= 0.6 positive, >= 0 neutral, < 0 negative
+        colors = [POSITIVE if s >= 0.6 else NEUTRAL if s >= 0 else NEGATIVE for s in sentiments]
         
         fig, ax = plt.subplots(figsize=(10, max(5, len(names) * 0.5)))
         fig.patch.set_facecolor(BG_COLOR)
@@ -538,7 +543,8 @@ def get_item_detail(item_name: str, state: dict) -> str:
             summary = item.get('summary', '')
             related_reviews = item.get('related_reviews', [])
             
-            emoji = "🟢" if sentiment > 0.3 else "🟡" if sentiment > -0.3 else "🔴"
+            # NEW thresholds: >= 0.6 positive, >= 0 neutral, < 0 negative
+            emoji = "🟢" if sentiment >= 0.6 else "🟡" if sentiment >= 0 else "🔴"
             
             detail = f"""### {clean_name.title()}
 
@@ -562,16 +568,14 @@ def get_item_detail(item_name: str, state: dict) -> str:
                     if text and len(text) > 20:
                         detail += f"> *\"{text[:200]}{'...' if len(text) > 200 else ''}\"*\n\n"
             
-            # Add actionable insight
+            # Add actionable insight - NEW thresholds
             detail += "\n**🎯 Recommended Action:**\n"
-            if sentiment > 0.5:
+            if sentiment >= 0.6:
                 detail += f"This is a **star performer**! Consider featuring {clean_name.title()} in promotions and training staff to recommend it."
-            elif sentiment > 0.2:
-                detail += f"Customers generally like {clean_name.title()}. Monitor feedback and maintain quality."
-            elif sentiment > -0.2:
-                detail += f"Mixed feedback on {clean_name.title()}. Review recent complaints and consider recipe adjustments."
+            elif sentiment >= 0:
+                detail += f"Customers have neutral/mixed feelings about {clean_name.title()}. Monitor feedback and look for improvement opportunities."
             else:
-                detail += f"⚠️ **Urgent:** {clean_name.title()} has significant negative feedback. Review preparation process and consider temporary removal while issues are addressed."
+                detail += f"⚠️ **Attention Needed:** {clean_name.title()} has negative feedback. Review preparation process and address customer complaints."
             
             return detail
     
@@ -593,7 +597,8 @@ def get_aspect_detail(aspect_name: str, state: dict) -> str:
             summary = aspect.get('summary', '')
             related_reviews = aspect.get('related_reviews', [])
             
-            emoji = "🟢" if sentiment > 0.3 else "🟡" if sentiment > -0.3 else "🔴"
+            # NEW thresholds: >= 0.6 positive, >= 0 neutral, < 0 negative
+            emoji = "🟢" if sentiment >= 0.6 else "🟡" if sentiment >= 0 else "🔴"
             
             detail = f"""### {clean_name.title()}
 
@@ -617,16 +622,14 @@ def get_aspect_detail(aspect_name: str, state: dict) -> str:
                     if text and len(text) > 20:
                         detail += f"> *\"{text[:200]}{'...' if len(text) > 200 else ''}\"*\n\n"
             
-            # Add actionable insight
+            # Add actionable insight - NEW thresholds
             detail += "\n**🎯 Recommended Action:**\n"
-            if sentiment > 0.5:
+            if sentiment >= 0.6:
                 detail += f"**{clean_name.title()}** is a major strength! Maintain current standards and use in marketing."
-            elif sentiment > 0.2:
-                detail += f"**{clean_name.title()}** is performing well. Look for opportunities to make it exceptional."
-            elif sentiment > -0.2:
-                detail += f"**{clean_name.title()}** has mixed reviews. Identify specific pain points and address them."
+            elif sentiment >= 0:
+                detail += f"**{clean_name.title()}** has neutral/mixed reviews. Identify specific areas to improve and make it exceptional."
             else:
-                detail += f"⚠️ **Priority Issue:** **{clean_name.title()}** is hurting the business. Immediate action needed - consider staff training, process changes, or operational review."
+                detail += f"⚠️ **Priority Issue:** **{clean_name.title()}** needs attention. Address customer complaints and consider staff training or process changes."
             
             return detail
     
@@ -799,9 +802,10 @@ def generate_pdf_report(state: dict) -> Optional[str]:
         all_sentiments = [item.get('sentiment', 0) for item in all_menu]
         avg_sentiment = sum(all_sentiments) / len(all_sentiments) if all_sentiments else 0
         
-        sent_label = "Excellent" if avg_sentiment > 0.5 else "Good" if avg_sentiment > 0.3 else "Positive" if avg_sentiment > 0 else "Mixed" if avg_sentiment > -0.3 else "Needs Attention"
-        sent_color = POSITIVE if avg_sentiment > 0.3 else WARNING if avg_sentiment > -0.3 else NEGATIVE
-        sent_bg = POSITIVE_LIGHT if avg_sentiment > 0.3 else WARNING_LIGHT if avg_sentiment > -0.3 else NEGATIVE_LIGHT
+        # NEW thresholds: >= 0.6 positive, >= 0 neutral, < 0 negative
+        sent_label = "Excellent" if avg_sentiment >= 0.8 else "Positive" if avg_sentiment >= 0.6 else "Neutral" if avg_sentiment >= 0 else "Needs Attention"
+        sent_color = POSITIVE if avg_sentiment >= 0.6 else WARNING if avg_sentiment >= 0 else NEGATIVE
+        sent_bg = POSITIVE_LIGHT if avg_sentiment >= 0.6 else WARNING_LIGHT if avg_sentiment >= 0 else NEGATIVE_LIGHT
         
         # Sentiment box
         sent_data = [[f"Overall Sentiment: {avg_sentiment:+.2f}", sent_label]]
@@ -827,7 +831,8 @@ def generate_pdf_report(state: dict) -> Optional[str]:
             for item in top_items:
                 elements.append(Paragraph(f"    • {item.get('name', '?').title()} (sentiment: {item.get('sentiment', 0):+.2f})", styles['RIABullet']))
         
-        concern_items = [i for i in all_menu if i.get('sentiment', 0) < -0.2]
+        # NEW threshold: < 0 for concerns
+        concern_items = [i for i in all_menu if i.get('sentiment', 0) < 0]
         if concern_items:
             elements.append(Spacer(1, 10))
             elements.append(Paragraph("⚠️ <b>Items Needing Attention:</b>", styles['RIABody']))
@@ -836,20 +841,18 @@ def generate_pdf_report(state: dict) -> Optional[str]:
         
         elements.append(Spacer(1, 15))
         
-        # Summary stats
-        stars = len([i for i in all_menu if i.get('sentiment', 0) > 0.5])
-        good = len([i for i in all_menu if 0.2 < i.get('sentiment', 0) <= 0.5])
-        mixed = len([i for i in all_menu if -0.2 <= i.get('sentiment', 0) <= 0.2])
-        concerns = len([i for i in all_menu if i.get('sentiment', 0) < -0.2])
+        # Summary stats - NEW thresholds
+        positive = len([i for i in all_menu if i.get('sentiment', 0) >= 0.6])
+        neutral = len([i for i in all_menu if 0 <= i.get('sentiment', 0) < 0.6])
+        negative = len([i for i in all_menu if i.get('sentiment', 0) < 0])
         
         summary_data = [
             ['Metric', 'Value', 'Details'],
             ['Reviews Analyzed', str(len(trend_data)), f'From {source}'],
             ['Menu Items', str(len(all_menu)), f'{len(food_items)} food, {len(drinks)} drinks'],
-            ['Customer Favorites', str(stars), 'Sentiment > 0.5'],
-            ['Performing Well', str(good), 'Sentiment 0.2 - 0.5'],
-            ['Mixed Reviews', str(mixed), 'Sentiment -0.2 - 0.2'],
-            ['Needs Attention', str(concerns), 'Sentiment < -0.2'],
+            ['🟢 Positive', str(positive), 'Sentiment ≥ 0.6'],
+            ['🟡 Neutral', str(neutral), 'Sentiment 0 to 0.59'],
+            ['🔴 Negative', str(negative), 'Sentiment < 0'],
         ]
         summary_table = Table(summary_data, colWidths=[2*inch, 1.3*inch, 2.5*inch])
         summary_table.setStyle(TableStyle([
@@ -881,7 +884,8 @@ def generate_pdf_report(state: dict) -> Optional[str]:
             menu_data = [['#', 'Item', 'Sentiment', 'Mentions', 'Status']]
             for i, item in enumerate(sorted_menu, 1):
                 sentiment = item.get('sentiment', 0)
-                status = '✓ Positive' if sentiment > 0.3 else '~ Mixed' if sentiment > -0.3 else '✗ Negative'
+                # NEW thresholds: >= 0.6 positive, >= 0 neutral, < 0 negative
+                status = '✓ Positive' if sentiment >= 0.6 else '~ Neutral' if sentiment >= 0 else '✗ Negative'
                 menu_data.append([str(i), item.get('name', '?').title()[:22], f"{sentiment:+.2f}", str(item.get('mention_count', 0)), status])
             
             menu_table = Table(menu_data, colWidths=[0.4*inch, 2.2*inch, 1*inch, 0.9*inch, 1.1*inch])
@@ -909,7 +913,8 @@ def generate_pdf_report(state: dict) -> Optional[str]:
             aspect_data = [['#', 'Aspect', 'Sentiment', 'Mentions', 'Status']]
             for i, aspect in enumerate(sorted_aspects, 1):
                 sentiment = aspect.get('sentiment', 0)
-                status = '✓ Strength' if sentiment > 0.3 else '~ Neutral' if sentiment > -0.3 else '✗ Weakness'
+                # NEW thresholds: >= 0.6 positive, >= 0 neutral, < 0 negative
+                status = '✓ Strength' if sentiment >= 0.6 else '~ Neutral' if sentiment >= 0 else '✗ Weakness'
                 aspect_data.append([str(i), aspect.get('name', '?').title()[:22], f"{sentiment:+.2f}", str(aspect.get('mention_count', 0)), status])
             
             aspect_table = Table(aspect_data, colWidths=[0.4*inch, 2.2*inch, 1*inch, 0.9*inch, 1.1*inch])
@@ -1049,9 +1054,10 @@ def generate_pdf_report(state: dict) -> Optional[str]:
         # Sort by sentiment to get best positive and worst negative
         for review in sorted(all_related_reviews, key=lambda x: x['sentiment'], reverse=True):
             text = review['text']
-            if review['sentiment'] > 0.2 and len(positive_reviews) < 3:
+            # NEW thresholds: >= 0.6 for positive, < 0 for negative
+            if review['sentiment'] >= 0.6 and len(positive_reviews) < 3:
                 positive_reviews.append(text[:180])
-            elif review['sentiment'] < -0.2 and len(negative_reviews) < 3:
+            elif review['sentiment'] < 0 and len(negative_reviews) < 3:
                 negative_reviews.append(text[:180])
         
         elements.append(Paragraph("✅ Positive Feedback", styles['RIASubHeader']))
@@ -1273,6 +1279,9 @@ Instructions:
 - If reviews mention specific examples, include them
 - Keep your answer helpful and concise (3-5 sentences)
 - If the reviews don't contain relevant information, say so honestly
+- Provide BALANCED answers - mention both pros AND cons when relevant
+- If customers have mixed opinions, acknowledge both positive and negative feedback
+- Don't oversell or undersell - be honest about what customers actually said
 
 Answer:"""