Update app.py

app.py

@@ -80,9 +80,12 @@ def generate_chunked_response(model, prompt, max_tokens=500, max_chunks=5):
     full_response = ""
     for i in range(max_chunks):
         chunk = model(prompt + full_response, max_new_tokens=max_tokens)
-        full_response += chunk
-        if chunk.strip().endswith((".", "!", "?")):
+        chunk = chunk.strip()
+        # Check for final sentence endings
+        if chunk.endswith((".", "!", "?")):
+            full_response += chunk
             break
+        full_response += chunk
     return full_response.strip()
 
 def response(database, model, question):
@@ -95,7 +98,7 @@ def response(database, model, question):
     formatted_prompt = prompt_val.format(context=context_str, question=question)
     
     ans = generate_chunked_response(model, formatted_prompt)
-    return ans  # Only return the answer
+    return ans  # Only return the final answer
 
 def update_vectors(files, use_recursive_splitter):
     if not files: