Update app.py

app.py

@@ -93,46 +93,58 @@ def configure_retriever(local_files, chunk_size=15000, chunk_overlap=2500):
 directory = "docs" if os.path.exists("docs") else "."
 local_files = [f for f in os.listdir(directory) if f.endswith(".pdf")]
 
-# Setup LLM
-llm = ChatOpenAI(
-    model_name="gpt-4-0125-preview", openai_api_key=openai_api_key, temperature=0.1, streaming=True
-)
+def setup_llm(system_message):
+    # Setup LLM
+    llm = ChatOpenAI(
+        model_name="gpt-4o", openai_api_key=openai_api_key, temperature=0.1, streaming=True
+    )
 
-retriever = configure_retriever(local_files)
+    retriever = configure_retriever(local_files)
 
-template = """Answer the question based only on the following context:
-{context}
+    template = system_message + """
 
-Question: {question}
+        Answer the question based only on the following context in it's original language.
+        {context}
 
-Chat History: {history}
+        Question: {question}
 
-Answer in German Language. If the question is not related to the context, answer with "I don't know". 
-If the user is asking for follow-up questions on the same topic, generate different questions than you already answered.
-"""
+        Original Message: {original_msg}
+
+        Chat History: {history}
+
+        If the question is not related to the context, answer with "I don't know" in the original language.
+        If the user is asking for follow-up questions on the same topic, generate different questions than you already answered.
+        If the user is asking to explain the context, or expand on the context, then provide explanation in the original language.
+        """
 
+    prompt = ChatPromptTemplate.from_template(template)
 
-prompt = ChatPromptTemplate.from_template(template)
+    chain_translate = (
+        llm
+        | StrOutputParser()
+    )
 
+    chain_rag = (
+        { 
+        "context": itemgetter("question") | retriever, 
+        "question": itemgetter("question"),
+        "original_msg": itemgetter("original_msg"),
+        "history": itemgetter("history")
+        }
+        | prompt
+        | llm
+        | StrOutputParser()
+    )
 
-chain_translate = (
-    llm
-    | StrOutputParser()
-)
+    return chain_rag, chain_translate
 
-chain_rag = (
-    { 
-    "context": itemgetter("question") | retriever, 
-    "question": itemgetter("question"),
-    "history": itemgetter("history")
-    }
-    | prompt
-    | llm
-    | StrOutputParser()
-)
+def predict(message, history, system_message):
 
-def predict(message, history):
-    message = chain_translate.invoke(f"Translate this query to English if it is in German otherwise return original contetn: {message}")
+    logging.info(system_message)
+
+    chain_rag, chain_translate = setup_llm(system_message)
+    
+    message_transalated = chain_translate.invoke(f"Translate this query to English if it is in German otherwise return original contetn: {message}")
 
     history_langchain_format = []
     partial_message = ""
@@ -141,7 +153,7 @@ def predict(message, history):
         history_langchain_format.append(HumanMessage(content=human))
         history_langchain_format.append(AIMessage(content=ai))
     history_langchain_format.append(HumanMessage(content=message))
-    for response in chain_rag.stream({"question": message, "history": history_langchain_format}):
+    for response in chain_rag.stream({"question": message_transalated, "original_msg": message, "history": history_langchain_format}):
         partial_message += response
         yield partial_message 
     
@@ -174,12 +186,10 @@ submit_btn="Senden",
 retry_btn="🔄  Wiederholen",
 undo_btn="⏪  Rückgängig",
 clear_btn="🗑️  Löschen",
-examples=[
-"Generate auditing questions about Change Management",
-"Generate auditing questions about Software Maintenance",
-"Generate auditing questions about Data Protection"
+additional_inputs=[
+gr.Textbox("You are an auditor with many years of professional experience and are to develop a questionnaire on the topic of home office in the form of a self-assessment for me. As a basis for the questionnaire, you use standards and best practices (for example, from ISO 27001 and COBIT). The questionnaire should not exceed 20 questions.", label="System Prompt")
 ],
-#cache_examples=True,
+cache_examples=False,
 fill_height=True,
 css=css,
 ).launch(show_api=False)