Update response_generator.py

response_generator.py

@@ -18,12 +18,8 @@ class ResponseGenerator:
             if len(relevant_docs) == 0:
                 return "عذرًا، لم أجد أي معلومات ذات صلة في المستندات."
             
-            # Get the most relevant document (first one)
-            context = self._prepare_context(relevant_docs)
-            
-            # For BERT-based models, we do extractive QA instead of generation
-            # Return the most relevant context as the answer
-            response = self._create_extractive_answer(query, context, relevant_docs)
+            # NEW: Use multi-chunk synthesis for better answers
+            response = self._create_multi_chunk_answer(query, relevant_docs)
             return response
                 
         except Exception as e:
@@ -39,6 +35,46 @@ class ResponseGenerator:
         max_context_length = 300
         return combined_content[:max_context_length]
     
+    # NEW: Multi-chunk answer synthesis
+    def _create_multi_chunk_answer(self, query, relevant_docs):
+        """
+        Create a comprehensive answer by combining top relevant chunks
+        """
+        # Get top 3 chunks
+        top_chunks = relevant_docs.head(3)
+        
+        # Build answer
+        answer_parts = []
+        answer_parts.append("**بناءً على المستندات المتاحة:**\n")
+        
+        for idx, (_, row) in enumerate(top_chunks.iterrows(), 1):
+            content = row['content'].strip()
+            similarity = row.get('similarity_score', 0)
+            
+            # Add chunk with similarity indicator
+            if len(top_chunks) > 1:
+                # NEW: Show similarity score for transparency
+                answer_parts.append(f"\n**[{idx}] مطابقة: {similarity:.0%}**")
+            
+            answer_parts.append(content)
+        
+        # NEW: Add metadata footer
+        answer_parts.append("\n\n---")
+        
+        # NEW: Show number of sources
+        unique_files = top_chunks['path'].nunique()
+        answer_parts.append(f"**عدد المصادر:** {unique_files}")
+        
+        # NEW: Show score range for quality indicator
+        if 'similarity_score' in top_chunks.columns:
+            scores = top_chunks['similarity_score'].values
+            if len(scores) == 1:
+                answer_parts.append(f" | **المطابقة:** {scores[0]:.0%}")
+            else:
+                answer_parts.append(f" | **المطابقة:** {scores[-1]:.0%} - {scores[0]:.0%}")
+        
+        return "\n".join(answer_parts)
+    
     def _create_extractive_answer(self, query, context, relevant_docs):
         """
         Create an answer by extracting relevant information from documents
@@ -57,13 +93,19 @@ class ResponseGenerator:
             else:
                 most_relevant = truncated + "..."
         
+        # NEW: Include similarity score if available
+        similarity_info = ""
+        if 'similarity_score' in relevant_docs.columns:
+            score = relevant_docs.iloc[0]['similarity_score']
+            similarity_info = f"\n**المطابقة:** {score:.0%}"
+        
         # Format the response
         response = f"""بناءً على المستندات المتاحة:
 
 {most_relevant}
 
 ---
-المصدر: {relevant_docs.iloc[0]['path']}"""
+المصدر: {relevant_docs.iloc[0]['path']}{similarity_info}"""
         
         return response
     
@@ -81,4 +123,47 @@ class ResponseGenerator:
             self.tokenizer = AutoTokenizer.from_pretrained(new_model_name)
             logger.info(f"Model updated to {new_model_name}")
         except Exception as e:
-            logger.error(f"Error updating model: {e}")
+            logger.error(f"Error updating model: {e}")
+```
+
+---
+
+## What Was Added (All marked with "# NEW"):
+
+1. ✅ **`_create_multi_chunk_answer()`** - New function that combines multiple chunks
+2. ✅ **Top 3 chunks** - Uses top 3 instead of just 1 for complete context
+3. ✅ **Similarity scores** - Shows match percentage for each chunk (e.g., "مطابقة: 85%")
+4. ✅ **Metadata footer** - Shows number of sources used
+5. ✅ **Score range** - Shows overall quality (e.g., "المطابقة: 72% - 85%")
+6. ✅ **Updated main method** - Now calls `_create_multi_chunk_answer()` instead of old method
+7. ✅ **Similarity in single answer** - Even single-chunk answers show score
+
+---
+
+## Before vs After Examples:
+
+**Before (Single chunk only):**
+```
+بناءً على المستندات المتاحة:
+
+نسبة الحضور المطلوبة هي 75%
+
+---
+المصدر: RAG DATA.txt
+```
+
+**After (Multi-chunk with scores):**
+```
+بناءً على المستندات المتاحة:
+
+[1] مطابقة: 85%
+نسبة الحضور المطلوبة هي 75%
+
+[2] مطابقة: 72%
+الغياب أكثر من 25% يؤدي للحرمان من دخول الامتحان
+
+[3] مطابقة: 68%
+يمكن تقديم عذر طبي للغياب
+
+---
+عدد المصادر: 1 | المطابقة: 68% - 85%