Update app.py

app.py

@@ -128,7 +128,7 @@ retriever = ensemble_retriever
 system_prompt = (
     "You are the LA2050 Navigator, an AI-powered chatbot created to help users discover organizations and community initiatives featured in the Goldhirsh Foundation’s LA2050 Ideas Hub. "
     "Your role is to deliver succinct, personalized recommendations, guide users toward supporting these initiatives, and answer questions about the Goldhirsh Foundation, LA2050, and its projects. "
-    "When responding, include the full name of the organization, a brief (1-2 sentence) description, and a link to its website(labeled as Organization website) or social media; (please do not alter the URL). "
+    "When responding, include the full name of the organization, a brief (1-2 sentence) description, and a link to its website (labeled as Organization website) or social media; (please do not alter the URL). "
     "If an organization’s personal website is unavailable, refer to its LA2050 URL. "
     "Prioritize nonprofit organizations designated as 'winners' by the Goldhirsh Foundation and those with multiple proposal submissions. "
     "If a user inquires about the LA2050 grant winners for a specific year, be sure to reference the metadata 'Year' and 'LA 2050 Grant Status' fields—explicitly noting if the organization was awarded the grant that year. "
@@ -136,6 +136,11 @@ system_prompt = (
     "If some information is missing, acknowledge it and direct the user to additional resources. "
     "Maintain a polite, helpful, respectful, and enthusiastic tone at all times. "
     "If the user responds with a follow-up confirmation (e.g., 'yes') after an initial answer, please expand on that topic with further details. "
+    "\n\nIMPORTANT: Answer the question using ONLY the information provided in the following documents. DO NOT invent or include any organizations that are not present in the retrieved evidence. "
+    "Before giving your final answer, perform the following steps: "
+    "Step 1: Identify all organizations mentioned in the retrieved documents. "
+    "Step 2: Check if there are any organizations beyond those provided that could be considered 'new'. "
+    "Step 3: If no additional organizations exist, clearly state that based on the current dataset, these are all the organizations we have information on. "
     "\n\n{context}"
 )
 
@@ -152,6 +157,16 @@ prompt = ChatPromptTemplate.from_messages(
 # Build the chain that will combine documents with the prompt.
 question_answer_chain = create_stuff_documents_chain(ChatOpenAI(model_name="gpt-4o-mini", temperature=0), prompt)
 rag_chain = create_retrieval_chain(retriever, question_answer_chain)
+
+def post_process_answer(answer, retrieved_docs):
+    """
+    Append a disclaimer to the answer confirming that only organizations from the retrieved documents were used.
+    (A more advanced implementation might parse and filter out any hallucinated names.)
+    """
+    # Extract allowed organization names from retrieved docs.
+    allowed_orgs = {doc.metadata.get("Organization", "").strip() for doc in retrieved_docs if doc.metadata.get("Organization", "").strip()}
+    disclaimer = "\n\n[Answer verified against retrieved documents: Only organizations present in the evidence were included. Allowed organizations: " + ", ".join(sorted(allowed_orgs)) + ".]"
+    return answer + disclaimer
 def debug_retrieved_docs(user_input):
     retrieved_docs = retriever.get_relevant_documents(user_input)
     print(f"DEBUG: Retrieved {len(retrieved_docs)} documents.")
@@ -193,7 +208,8 @@ green_theme = gr.themes.Base(
     button_secondary_text_color='#000000'
 )
 
-def message_and_history(message, history):
+
+  def message_and_history(message, history):
     # Initialize conversation with a welcome message if history is empty.
     if not history:
         history = [{"role": "assistant", "content": "<b>LA2050 Navigator:</b><br> Welcome to the LA2050 ideas hub! How can I help you today?"}]
@@ -211,7 +227,7 @@ def message_and_history(message, history):
 
     # Combine the most recent conversation turns, excluding the assistant's prefix.
     conversation_context = "\n".join(
-        [f"{msg['role']}: {msg['content'].replace('<b>LA2050 Navigator:</b><br>', '')}" for msg in history[0:]]
+        [f"{msg['role']}: {msg['content'].replace('<b>LA2050 Navigator:</b><br>', '')}" for msg in history[-2:]]
     )
     retrieved_docs = retriever.get_relevant_documents(conversation_context)
     print(f"DEBUG: Retrieved {len(retrieved_docs)} documents.")
@@ -223,13 +239,15 @@ def message_and_history(message, history):
     try:
         response = rag_chain.invoke(chain_input)
         answer = response["answer"]
+        # Post-process the answer to append a disclaimer verifying the evidence.
+        answer = post_process_answer(answer, retrieved_docs)
     except Exception as e:
         answer = f"An error occurred: {e}"
 
     # Remove the prefix if the model includes it.
     if answer.startswith("<b>LA2050 Navigator:</b><br>"):
         answer = answer[len("<b>LA2050 Navigator:</b><br>"):]
-
+    
     # Initialize the assistant's response with the prefix.
     assistant_response = {"role": "assistant", "content": "<b>LA2050 Navigator:</b><br> "}
     history.append(assistant_response)