Added Cohere reranker

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_vectorstore, load_conversational_retrievel_chain
+from utils import load_retriver,load_vectorstore, load_conversational_retrievel_chain
 
 load_dotenv()
 # Getting current timestamp to keep track of historical conversations
@@ -54,11 +54,11 @@ if "vector_db" not in st.session_state:
 # if "text_chunks" not in st.session_state:
 #     st.session_state["text_chunks"] = load_text_chunks(text_chunks_pkl_dir=all_docs_pkl_directory)
 
-if "ensemble_retriver" not in st.session_state:
-    st.session_state["ensemble_retriver"] = load_ensemble_retriver(embeddings=st.session_state["embeddings"], chroma_vectorstore=st.session_state["vector_db"] )
+if "retriever" not in st.session_state:
+    st.session_state["retriever"] = load_retriver(embeddings=st.session_state["embeddings"], chroma_vectorstore=st.session_state["vector_db"] )
 
 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"])    
+    st.session_state["conversation_chain"] = load_conversational_retrievel_chain(retriever=st.session_state["retriever"], llm=st.session_state["llm"])    
 
 
 
@@ -83,8 +83,12 @@ title1 = """
 """
 
 def clear_chat_history():
+        """
+        Clear chat and start new chat
+        """
         st.session_state.messages = [{"role": "assistant", "content": "How may I assist you today?"}]
 
+#loading OPM logo
 file_ = open("opm_logo.png", "rb")
 contents = file_.read()
 data_url = base64.b64encode(contents).decode("utf-8")
@@ -215,11 +219,6 @@ if st.session_state["vector_db"] and st.session_state["llm"]:
                 for item in response:
                     full_response += item
                 placeholder.markdown(full_response)
-                # The following logic will work in the way given below.
-                # -- Check if intermediary steps are present in the output of the given prompt.
-                # -- 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 response:
                     st.text("-------------------------------------")
                     docs= st.session_state["ensemble_retriver"].get_relevant_documents(prompt)   

utils.py

@@ -33,6 +33,9 @@ from langchain.agents import load_tools
 from langchain.chat_models import ChatOpenAI
 from langchain.retrievers.multi_query import MultiQueryRetriever
 from langchain.chains import RetrievalQA
+from langchain.retrievers import ContextualCompressionRetriever
+from langchain.retrievers.document_compressors import CohereRerank
+
 import logging
 
 
@@ -60,6 +63,10 @@ def build_embedding_model():
     return embeddings
 
 def unzip_opm():
+    '''
+    This function is used to unzip the documents file. This is required if there is no extisting vector database
+    created and wanted to build from the scratch
+    '''
     # Specify the path to your ZIP file
     zip_file_path = r'OPM_Files/OPM_Retirement_backup-20230902T130906Z-001.zip'
 
@@ -90,7 +97,9 @@ def unzip_opm():
 
 def count_files_by_type(folder_path):
     '''
-    Counting files by file type in the specified folder
+    Counting files by file type in the specified folder.
+    This is required if there is no extisting vector database
+    created and wanted to build from the scratch
     '''
     file_count_by_type = defaultdict(int)
     
@@ -103,7 +112,9 @@ def count_files_by_type(folder_path):
 
 def generate_file_count_table(file_count_by_type):
     '''
-    Generate a table files count file type
+    Generate a table files count file type.
+    This is required if there is no extisting vector database
+    created and wanted to build from the scratch
     '''
     data = {"File Type": [], "Number of Files": []}
     for extension, count in file_count_by_type.items():
@@ -117,6 +128,8 @@ def generate_file_count_table(file_count_by_type):
 def move_files_to_folders(folder_path):
     '''
     Move files to respective folder. Example, PDF docs to PDFs folder, HTML docs to HTMLs folder.
+    This is required if there is no extisting vector database
+    created and wanted to build from the scratch
     '''
     for root, _, files in os.walk(folder_path):
         for file in files:
@@ -144,7 +157,9 @@ def load_vectorstore(persist_directory, embeddings):
         2) create text chunks
         3) Index it and store it in a Chroma DB
         4) Peform the same for HTML files
-        5) Store the final chroma db in the disk
+        5) Store the final chroma db in the disk.
+        This is required if there is no extisting vector database
+        created and wanted to build from the scratch
     '''
     if os.path.exists(persist_directory):
         print("Using existing vectore store for these documents.")
@@ -214,6 +229,7 @@ def load_vectorstore(persist_directory, embeddings):
 
 def load_text_chunks(text_chunks_pkl_dir):
     '''
+    We need to get all the text chunks as it is required for bm25 retriever incase we are using it for creating enemble retriever 
     Loading the pickle file that holds all the documents from the disk.
     If it does not exist, create new one.
     Text documents are required to create BM25 Retriever. But loading all the documents in
@@ -253,68 +269,59 @@ def load_text_chunks(text_chunks_pkl_dir):
             pickle.dump(all_texts, file)
         print("Text chunks are created and cached")
         
-def load_ensemble_retriver(embeddings, chroma_vectorstore):
-    """Load ensemble retiriever with BM25 and Chroma as individual retrievers"""
-    # bm25_retriever = BM25Retriever.from_documents(text_chunks)
-    # bm25_retriever.k = 2
-    chroma_retriever = chroma_vectorstore.as_retriever(search_kwargs={"k": 10})  
+def load_retriver(text_chunks, embeddings, chroma_vectorstore):
+    """Load cohere rerank method for retrieval"""
+    bm25_retriever = BM25Retriever.from_documents(text_chunks)
+    bm25_retriever.k = 2
+    chroma_retriever = chroma_vectorstore.as_retriever(search_kwargs={"k": 3})  
+    # ensemble_retriever = EnsembleRetriever(retrievers=[bm25_retriever, chroma_retriever], weights=[0.3, 0.7])
+    logging.basicConfig()
+    logging.getLogger('langchain.retrievers.multi_query').setLevel(logging.INFO)
+    multi_query_retriever = MultiQueryRetriever.from_llm(retriever=chroma_retriever,
+                                                              llm=ChatOpenAI(temperature=0))
+    compressor = CohereRerank()
+    compression_retriever = ContextualCompressionRetriever(
+        base_compressor=compressor,
+        base_retriever=multi_query_retriever)
+    return compression_retriever
+
+
+def load_retriver(text_chunks, embeddings, chroma_vectorstore):
+    """Load cohere rerank method for retrieval"""
+    bm25_retriever = BM25Retriever.from_documents(text_chunks)
+    bm25_retriever.k = 2
+    chroma_retriever = chroma_vectorstore.as_retriever(search_kwargs={"k": 3})  
     # ensemble_retriever = EnsembleRetriever(retrievers=[bm25_retriever, chroma_retriever], weights=[0.3, 0.7])
     logging.basicConfig()
     logging.getLogger('langchain.retrievers.multi_query').setLevel(logging.INFO)
-    retriever_from_llm = MultiQueryRetriever.from_llm(retriever=chroma_retriever,
+    multi_query_retriever = MultiQueryRetriever.from_llm(retriever=chroma_retriever,
                                                               llm=ChatOpenAI(temperature=0))
-    return retriever_from_llm
+    compressor = CohereRerank()
+    compression_retriever = ContextualCompressionRetriever(
+        base_compressor=compressor,
+        base_retriever=multi_query_retriever)
+    return compression_retriever
 
 
 def load_conversational_retrievel_chain(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
-  # string_dialogue = "You are a helpful assistant. You do not respond as 'User' or pretend to be 'User'. You only respond once as 'Assistant'."
-    # _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.
-    # Your answer should in English language only.
-    # 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,
-    # )
-    template = """You are a helpful assistant. You do not respond as 'User' or pretend to be 'User'. You only respond once as 'Assistant'.
-    Use the following pieces of context to answer the question at the end. If you don't know the answer,\
-    just say that you don't know, don't try to make up an answer.
-
-    {context}
-
-    {history}
-    Question: {question}
-    Helpful Answer:"""
-
-    prompt = PromptTemplate(input_variables=["history", "context", "question"], template=template)
+    Create RetrievalQA chain with memory
+    '''
+    # template = """You are a helpful assistant. You do not respond as 'User' or pretend to be 'User'. You only respond once as 'Assistant'.
+    # Use the following pieces of context to answer the question at the end. If you don't know the answer, just say that you don't know, don't try to make up an answer.
+    # Only include information found in the results and don't add any additional information.
+    # Make sure the answer is correct and don't output false content.
+    # If the text does not relate to the query, simply state 'Text Not Found in the Document'. Ignore outlier,
+    # search results which has nothing to do with the question. Only answer what is asked.
+    # The answer should be short and concise. Answer step-by-step.
+
+    # {context}
+
+    # {history}
+    # Question: {question}
+    # Helpful Answer:"""
+
+    # prompt = PromptTemplate(input_variables=["history", "context", "question"], template=template)
     memory = ConversationBufferMemory(input_key="question", memory_key="history")
 
     qa = RetrievalQA.from_chain_type(
@@ -325,3 +332,4 @@ def load_conversational_retrievel_chain(retriever, llm):
         chain_type_kwargs={"memory": memory},
     )
     return qa
+