Update app.py

app.py

@@ -391,7 +391,7 @@ def ask_question(question, temperature, top_p, repetition_penalty, web_search, c
                 )
 
                 full_response = generate_chunked_response(model, formatted_prompt)
-                answer = extract_answer(full_response)
+                answer = extract_answer(full_response, instructions)
                 all_answers.append(answer)
                 break
 
@@ -448,7 +448,7 @@ def ask_question(question, temperature, top_p, repetition_penalty, web_search, c
 
     return "An unexpected error occurred. Please try again later."
 
-def extract_answer(full_response):
+def extract_answer(full_response, instructions=None):
     # First, try to split the response at common instruction phrases
     answer_patterns = [
         r"Provide a concise and direct answer to the question without mentioning the web search or these instructions:",
@@ -466,7 +466,7 @@ def extract_answer(full_response):
             full_response = match[-1].strip()
             break
     
-    # Then, remove any remaining instruction-like phrases
+    # Remove any remaining instruction-like phrases
     cleanup_patterns = [
         r"without mentioning the web search or these instructions\.",
         r"Do not include any source information in your answer\.",
@@ -476,7 +476,12 @@ def extract_answer(full_response):
     for pattern in cleanup_patterns:
         full_response = re.sub(pattern, "", full_response, flags=re.IGNORECASE).strip()
     
-    return full_response
+    # Remove the user instructions if present
+    if instructions:
+        instruction_pattern = rf"User Instructions:\s*{re.escape(instructions)}.*?\n"
+        full_response = re.sub(instruction_pattern, "", full_response, flags=re.IGNORECASE | re.DOTALL)
+    
+    return full_response.strip()
 
 # Gradio interface
 with gr.Blocks() as demo: