Update utils and app files

Removed agent from the chain

app.py

@@ -30,7 +30,7 @@ from langchain.agents.agent_toolkits import create_conversational_retrieval_agen
 from langchain.utilities import SerpAPIWrapper
 
 from utils import build_embedding_model, build_llm
-from utils import load_ensemble_retriver, load_text_chunks, load_vectorstore, load_conversational_retrievel_agent
+from utils import load_ensemble_retriver, load_text_chunks, load_vectorstore, load_conversational_retrievel_chain
 
 load_dotenv()
 # Getting current timestamp to keep track of historical conversations
@@ -57,8 +57,8 @@ if "text_chunks" not in st.session_state:
 if "ensemble_retriver" not in st.session_state:
     st.session_state["ensemble_retriver"] = load_ensemble_retriver(text_chunks=st.session_state["text_chunks"], embeddings=st.session_state["embeddings"], chroma_vectorstore=st.session_state["vector_db"] )
 
-if "agent_executor" not in st.session_state:
-    st.session_state["agent_executor"] = load_conversational_retrievel_agent(retriever=st.session_state["ensemble_retriver"], llm=st.session_state["llm"])    
+if "conversation_chain" not in st.session_state:
+    st.session_state["conversation_chain"] = load_conversational_retrievel_chain(retriever=st.session_state["ensemble_retriver"], llm=st.session_state["llm"])
 
 
 
@@ -192,9 +192,9 @@ if st.session_state["vector_db"] and st.session_state["llm"]:
                     st.write(message)
 
 
-    def generate_llm_response(agent_executor, prompt_input):
-        result = agent_executor({"input": prompt_input})
-        return [result['output'], result['intermediate_steps']]
+    def generate_llm_response(conversation_chain, prompt_input):
+        output= conversation_chain({'question': prompt_input})
+        return output['answer']
 
 
     # User-provided prompt
@@ -208,10 +208,10 @@ if st.session_state["vector_db"] and st.session_state["llm"]:
         with st.chat_message("assistant"):
             with st.spinner("Searching..."):
                 start = timeit.default_timer()
-                response = generate_llm_response(agent_executor=st.session_state["agent_executor"], prompt_input=prompt)
+                response = generate_llm_response(conversation_chain=st.session_state["conversation_chain"], prompt_input=prompt)
                 placeholder = st.empty()
                 full_response = ''
-                for item in response[0]:
+                for item in response:
                     full_response += item
                 placeholder.markdown(full_response)
                 # The following logic will work in the way given below.
@@ -219,7 +219,7 @@ if st.session_state["vector_db"] and st.session_state["llm"]:
                 # -- If not, we can conclude that, agent has used internet search as tool.
                 # -- Check if intermediary steps are present in the output of the prompt.
                 # -- If intermediary steps are present, it means agent has used exising custom knowledge base for iformation retrival and therefore we need to give souce docs as output along with LLM's reponse.
-                if len(response[1])>0:
+                if response:
                     st.text("-------------------------------------")
                     docs= st.session_state["ensemble_retriver"].get_relevant_documents(prompt)   
                     source_doc_list= []  
@@ -237,8 +237,6 @@ if st.session_state["vector_db"] and st.session_state["llm"]:
                             st.write("---")  # Add a separator between entries
                     message = {"role": "assistant", "content": full_response, "Source":merged_source_doc}
                     st.session_state.messages.append(message)
-                else:
-                    message = {"role": "assistant", "content": full_response, "Source":""}
-                    st.session_state.messages.append(message)
+                    
         end = timeit.default_timer()
         print(f"Time to retrieve response: {end - start}")

utils.py

@@ -257,23 +257,24 @@ def load_ensemble_retriver(text_chunks, embeddings, chroma_vectorstore):
     return ensemble_retriever
 
 
-def load_conversational_retrievel_agent(retriever, llm):
-    '''Load Conversational Retrievel agent with following tasks as tools,
-    1) OPM Knowledge base query
-    2) INternet search with SerpAPI
-    This agent combines RAG, chat interfaces, agents.
-    '''
-    retriever_tool = create_retriever_tool(
-    retriever, 
-    "Search_US_Office_of_Personnel_Management_Document",
-    "Searches and returns documents regarding the U.S. Office of Personnel Management (OPM).")
-    search_api = SerpAPIWrapper()
-    search_api_tool = Tool(
-            name = "Current_Search",
-            func=search_api.run,
-            description="useful for when you need to answer questions about current events or the current state of the world"
-        )
-    tools = [retriever_tool]
-    agent_executor = create_conversational_retrieval_agent(llm, tools, verbose=True, max_token_limit=512)
-    return agent_executor
+def load_conversational_retrievel_chain(retriever, llm):
+    '''Load Conversational Retrievel chain,'''
+    _template= """
+    You are a helpful assistant. You do not respond as 'User' or pretend to be 'User'. You only respond once as 'Assistant'.
+    Given the following conversation and a follow up question, rephrase the follow up question to be a standalone question, in its original language.
+    Chat History:
+    {chat_history}
+    Follow Up Input: {question}
+    Standalone question:"""
+
+    CONDENSE_QUESTION_PROMPT = PromptTemplate.from_template(_template)
+    memory = ConversationBufferMemory(return_messages=True,memory_key="chat_history")
+    conversation_chain = ConversationalRetrievalChain.from_llm(
+        llm=st.session_state["llm"],
+        retriever=st.session_state["ensemble_retriver"],
+        condense_question_prompt=CONDENSE_QUESTION_PROMPT,
+        memory=memory,
+        verbose=True,
+    )
+    return conversation_chain