Update app.py

app.py

@@ -389,16 +389,30 @@ chain_neo4j = (
 #         return "Sorry, I encountered an error while processing your request.", []
 
 
-# # Short Prompt Template for Phi-3.5 Proprietary Model
+# Short Prompt Template for Phi-3.5 Proprietary Model
 
-# phi_short_template = f"""
-# As an expert on Birmingham, Alabama, I will provide concise, accurate, and informative responses to your queries based on 128 token limit . Given the sunny weather today, {current_date}, feel free to ask me anything you need to know about the city.
-# Provide only the direct answer to the question without any follow-up questions.
-# {{context}}
-# Question: {{question}}
-# Answer:
-# """
+phi_short_template = f"""
+As an expert on Birmingham, Alabama, I will provide concise, accurate, and informative responses to your queries based on 128 token limit . Given the sunny weather today, {current_date}, feel free to ask me anything you need to know about the city.
+Provide only the direct answer to the question without any follow-up questions.
+{{context}}
+Question: {{question}}
+Answer:
+"""
+
+
+import re
 
+def clean_response(response_text):
+    # Remove any metadata-like information and focus on the main content
+    # Removes "Document(metadata=...)" and other similar patterns
+    cleaned_response = re.sub(r'Document\(metadata=.*?\),?\s*', '', response_text, flags=re.DOTALL)
+    cleaned_response = re.sub(r'page_content=".*?"\),?', '', cleaned_response, flags=re.DOTALL)
+    cleaned_response = re.sub(r'\[.*?\]', '', cleaned_response, flags=re.DOTALL)  # Remove content in brackets
+    cleaned_response = re.sub(r'\s+', ' ', cleaned_response).strip()
+    #Remove any unwanted follow-up questions or unnecessary text
+    cleaned_response = re.sub(r'Question:.*\nAnswer:', '', response_text, flags=re.DOTALL).strip()
+    return cleaned_response
+    
 
 import re
 
@@ -418,24 +432,20 @@ def generate_answer(message, choice, retrieval_mode, selected_model):
     logging.debug(f"generate_answer called with choice: {choice} and retrieval_mode: {retrieval_mode}")
 
     try:
-        # Handle hotel-related queries
         if "hotel" in message.lower() or "hotels" in message.lower() and "birmingham" in message.lower():
             response = fetch_google_hotels()
             return response, extract_addresses(response)
         
-        # Handle restaurant-related queries
         if "restaurant" in message.lower() or "restaurants" in message.lower() and "birmingham" in message.lower():
             response = fetch_yelp_restaurants()
             return response, extract_addresses(response)
 
-        # Handle flight-related queries
         if "flight" in message.lower() or "flights" in message.lower() and "birmingham" in message.lower():
             response = fetch_google_flights()
             return response, extract_addresses(response)
 
         if retrieval_mode == "VDB":
             if selected_model == chat_model:
-                # Use GPT-4o with its vector store and template
                 retriever = gpt_retriever
                 prompt_template = QA_CHAIN_PROMPT_1 if choice == "Details" else QA_CHAIN_PROMPT_2
                 context = retriever.get_relevant_documents(message)
@@ -451,28 +461,24 @@ def generate_answer(message, choice, retrieval_mode, selected_model):
                 return response['result'], extract_addresses(response['result'])
 
             elif selected_model == phi_pipe:
-                # Use Phi-3.5 with its vector store and a simplified prompt
                 retriever = phi_retriever
                 context = retriever.get_relevant_documents(message)
-                prompt = f"""
-                Here is the information :
-                {context}
-                {message}
-                """
+                prompt = phi_short_template.format(context=context, question=message)
 
                 logging.debug(f"Phi-3.5 Prompt: {prompt}")
 
                 response = selected_model(prompt, **{
-                    "max_new_tokens": 128,
+                    "max_new_tokens": 128,  # Increased to handle longer responses
                     "return_full_text": False,
-                    "temperature": 0.5,
-                    "do_sample": False,
+                    "temperature": 0.7,  # Adjusted to avoid cutting off
+                    "do_sample": True,   # Allow sampling to increase response diversity
                 })
 
                 if response:
                     generated_text = response[0]['generated_text']
                     logging.debug(f"Phi-3.5 Response: {generated_text}")
-                    return generated_text, extract_addresses(generated_text)
+                    cleaned_response = clean_response(generated_text)
+                    return cleaned_response, extract_addresses(cleaned_response)
                 else:
                     logging.error("Phi-3.5 did not return any response.")
                     return "No response generated.", []