Update app.py

app.py

@@ -388,7 +388,7 @@ class RAGIndex:
         return answer
 
     def answer(self, question: str) -> str:
-        """Answer a question using RAG with simplified, clearer prompting."""
+        """Answer a question using RAG with improved generation."""
         if not self.initialized:
             return "❌ Assistant not properly initialized. Please check the logs."
     
@@ -427,26 +427,39 @@ class RAGIndex:
                 f"💡 Try rephrasing your question or adding more detailed documents to the knowledge base."
             )
         
-        # Combine contexts (limit to avoid overwhelming the model)
-        combined_context = " ".join(evidence_parts[:2])[:1000]
+        # Combine contexts - use MORE context for better answers
+        combined_context = " ".join(evidence_parts[:3])[:1500]  # Top 3, up to 1500 chars
         
-        # 3) FIXED: Simple, direct prompt (no complex instructions)
-        answer_prompt = f"""Answer this question using the context below. Be concise and natural.
+        # 3) Better prompt structure for FLAN-T5
+        answer_prompt = f"""Read the context and answer the question in 2-3 complete sentences.
     
     Context: {combined_context}
     
     Question: {question}
     
-    Answer:"""
+    Answer in complete sentences:"""
     
         try:
-            answer_text = self._generate_from_context(answer_prompt, max_new_tokens=150)
+            answer_text = self._generate_from_context(answer_prompt, max_new_tokens=200)
             answer_text = answer_text.strip()
             
-            # Safety check: if model leaked instructions, try simpler prompt
-            if answer_text.startswith("Do NOT") or answer_text.startswith("You are") or len(answer_text) < 10:
-                simple_prompt = f"Context: {combined_context}\n\nQ: {question}\nA:"
-                answer_text = self._generate_from_context(simple_prompt, max_new_tokens=150).strip()
+            # Clean up common artifacts
+            answer_text = answer_text.replace("**", "").replace("##", "").strip()
+            
+            # If answer is poor quality, try alternative approach
+            if (len(answer_text) < 20 or 
+                answer_text.count(":") > 3 or 
+                answer_text.startswith("Do NOT") or
+                "tone tone tone" in answer_text.lower()):
+                
+                # Try extractive approach: just clean and present the best context
+                best_context = evidence_parts[0]
+                # Remove list markers and clean
+                best_context = re.sub(r'^\s*[-*]\s*', '', best_context, flags=re.MULTILINE)
+                best_context = re.sub(r'^\s*\d+\.\s*', '', best_context, flags=re.MULTILINE)
+                # Take first few sentences
+                sentences = best_context.split('.')[:3]
+                answer_text = '. '.join(s.strip() for s in sentences if s.strip()) + '.'
             
         except Exception as e:
             print(f"Generation error: {e}")
@@ -462,21 +475,6 @@ class RAGIndex:
             f"**Sources:** {sources_str}"
         )
 
-        try:
-            answer_text = self._generate_from_context(answer_prompt, max_new_tokens=128)
-        except Exception as e:
-            print(f"Generation error: {e}")
-            return (
-                "There was an error while generating the answer. "
-                "Please try again with a shorter question or different wording."
-            )
-
-        sources_str = ", ".join(sorted(used_sources)) if used_sources else "N/A"
-
-        return (
-            f"**Answer:** {answer_text}\n\n"
-            f"**Sources:** {sources_str}"
-        )
 
 
 # Initialize RAG system