Update app.py

app.py

@@ -224,17 +224,6 @@ class SmartDocumentRAG:
             return ""
 
     def answer_question(self, query: str) -> str:
-        """
-        Answer the user's question based on processed documents.
-    
-        Features:
-        - Returns document summary if query asks for summary.
-        - Uses semantic search to find relevant context.
-        - Uses QA pipeline with prompt-style input.
-        - Applies confidence threshold to reduce hallucinations.
-        - Returns a fallback message if answer is unreliable.
-        """
-    
         if not query.strip():
             return "❓ Please ask a valid question."
     
@@ -243,43 +232,52 @@ class SmartDocumentRAG:
     
         query_lower = query.lower()
     
-        # Handle summary requests
         if any(word in query_lower for word in ['summary', 'summarize', 'overview', 'about']):
             if self.document_summary:
                 return f"📄 Document Summary:\n\n{self.document_summary}"
             else:
                 return "⚠️ Summary not available. Please process documents first."
     
-        # Find relevant chunks for context
-        context = self.find_relevant_content(query, top_k=3)
+        context = self.find_relevant_content(query, k=5)
+        print(f"Context found (top 5 chunks): {context}")
+    
         if not context:
             return "🔍 Sorry, no relevant information was found for your question. Try rephrasing."
     
         try:
-            # Prepare input for QA pipeline (some QA pipelines accept question and context separately)
-            # For distilbert QA pipeline:
-            result = self.qa_pipeline(question=query, context=context)
+            if self.model_type in ["distilbert-qa", "fallback"]:
+                result = self.qa_pipeline(question=query, context=context)
+                print(f"QA Pipeline output: {result}")
+                answer = result.get('answer', '').strip()
+                score = result.get('score', 0.0)
     
-            answer = result.get('answer', '').strip()
-            score = result.get('score', 0.0)
+                if not answer or score < 0.05:
+                    return "🤔 I couldn't find a confident answer to your question based on the documents."
     
-            # Confidence threshold to prevent hallucination
-            if score < 0.20 or not answer or answer.lower() in ['no answer', '']:
-                return "🤔 I couldn't find a confident answer to your question based on the documents."
+                snippet = context[:300].strip()
+                if len(context) > 300:
+                    snippet += "..."
     
-            # Optional heuristic: check if answer is too generic or unrelated
-            if len(answer) < 3 or (query_lower not in answer.lower() and score < 0.35):
-                return "🤔 I couldn't find a confident answer to your question based on the documents."
+                return f"**Answer:** {answer}\n\n*Context snippet:* {snippet}"
     
-            # Return answer with a snippet of context for transparency
-            snippet = context[:300].strip()
-            if len(context) > 300:
-                snippet += "..."
+            elif self.model_type == "flan-t5":
+                prompt = (
+                    f"Answer the question based on the context below.\n\n"
+                    f"Context:\n{context}\n\n"
+                    f"Question: {query}\nAnswer:"
+                )
+                result = self.qa_pipeline(prompt, max_length=200, num_return_sequences=1)
+                print(f"Generative pipeline output: {result}")
     
-            return f"**Answer:** {answer}\n\n*Context snippet:* {snippet}"
+                answer = result[0]['generated_text'].replace(prompt, '').strip()
+                if not answer:
+                    return "🤔 I couldn't find a confident answer to your question based on the documents."
+                return f"**Answer:** {answer}"
+    
+            else:
+                return "⚠️ Unsupported model type for QA."
     
         except Exception as e:
-            # If model fails, fallback to simple answer or message
             return f"❌ An error occurred while answering your question: {str(e)}"
 
     def extract_direct_answer(self, query: str, context: str) -> str: