Update functions.py

functions.py

@@ -26,7 +26,7 @@ from pyvis.network import Network
 import torch
 from pydub import AudioSegment
 from langchain.docstore.document import Document
-from langchain.embeddings import HuggingFaceEmbeddings,HuggingFaceInstructEmbeddings
+from langchain.embeddings import HuggingFaceEmbeddings, HuggingFaceBgeEmbeddings, HuggingFaceInstructEmbeddings
 from langchain.vectorstores import FAISS
 from langchain.text_splitter import RecursiveCharacterTextSplitter
 from langchain.chat_models import ChatOpenAI
@@ -34,7 +34,13 @@ from langchain.callbacks import StdOutCallbackHandler
 from langchain.chains import ConversationalRetrievalChain, QAGenerationChain, LLMChain
 from langchain.memory import ConversationBufferMemory
 from langchain.chains.question_answering import load_qa_chain
-from langchain.chains.conversational_retrieval.prompts import CONDENSE_QUESTION_PROMPT
+
+from langchain.callbacks import StreamlitCallbackHandler
+from langchain.agents import OpenAIFunctionsAgent, AgentExecutor
+from langchain.agents.agent_toolkits import create_retriever_tool
+from langchain.agents.openai_functions_agent.agent_token_buffer_memory import (
+    AgentTokenBufferMemory,
+)
 
 from langchain.prompts.chat import (
     ChatPromptTemplate,
@@ -97,6 +103,7 @@ def load_prompt():
         
     return prompt
 
+
 ###################### Functions #######################################################################################
 
 # @st.cache_data
@@ -528,6 +535,30 @@ def generate_eval(raw_text, N, chunk):
 
     return eval_set_full
 
+@st.cache_resource
+def create_prompt_and_llm():
+    '''Create prompt'''
+
+    llm = ChatOpenAI(temperature=0, streaming=True, model="gpt-4")
+
+    message = SystemMessage(
+        content=(
+            "You are a helpful chatbot who is tasked with answering questions acuurately about earnings call transcript provided. "
+            "Unless otherwise explicitly stated, it is probably fair to assume that questions are about the earnings call transcript. "
+            "If there is any ambiguity, you probably assume they are about that."
+            "Do not use any information not provided in the earnings context and remember you are a to speak like a finance expert."
+            "If you don't know the answer, just say 'There is no relevant answer in the given earnings call transcript'" 
+            "don't try to make up an answer"
+        )
+    )
+
+    prompt = OpenAIFunctionsAgent.create_prompt(
+        system_message=message,
+        extra_prompt_messages=[MessagesPlaceholder(variable_name="history")],
+    )
+
+    return prompt, llm
+    
 @st.cache_resource
 def gen_embeddings(embedding_model):
 
@@ -539,14 +570,21 @@ def gen_embeddings(embedding_model):
                                            query_instruction='Represent the Financial question for retrieving supporting paragraphs: ',
                                            embed_instruction='Represent the Financial paragraph for retrieval: ')
 
-    else:
+    elif 'mpnet' in embedding_model:
         
         embeddings = HuggingFaceEmbeddings(model_name=embedding_model)
 
+    elif 'FlagEmbedding' in embedding_model:
+
+        encode_kwargs = {'normalize_embeddings': True}
+        embeddings = HuggingFaceBgeEmbeddings(model_name=embedding_model,
+                                                   encode_kwargs = encode_kwargs
+                                                   )
+
     return embeddings
     
 @st.cache_data
-def process_corpus(corpus, title, embedding_model, chunk_size=1000, overlap=50):
+def create_vectorstore(corpus, title, embedding_model, chunk_size=1000, overlap=50):
 
     '''Process text for Semantic Search'''
     
@@ -560,34 +598,36 @@ def process_corpus(corpus, title, embedding_model, chunk_size=1000, overlap=50):
 
     return vectorstore
     
-def embed_text(query,_docsearch):
+def create_memory_and_agent(query,_docsearch):
     
     '''Embed text and generate semantic search scores'''
+
+    #create vectorstore
+    vectorstore = _docsearch.as_retriever(search_kwargs={"k": 4})
+
+    #create retriever tool
+    tool = create_retriever_tool(
+    vectorstore,
+    "earnings_call_search",
+    "Searches and returns documents using the earnings context provided as a source, relevant to the user input question.",
+    )
+
+    tools = [tool]
+
+    prompt,llm = create_prompt_and_llm()
+
+    agent = OpenAIFunctionsAgent(llm=llm, tools=tools, prompt=prompt)
     
-    # llm = OpenAI(temperature=0)
-    chat_llm = ChatOpenAI(streaming=True,
-                          model_name = 'gpt-4',
-                          callbacks=[StdOutCallbackHandler()], 
-                          verbose=True, 
-                          temperature=0
-                         )
-
-    # chain = RetrievalQA.from_chain_type(llm=chat_llm, chain_type="stuff", 
-    #                              retriever=_docsearch.as_retriever(), 
-    #                              return_source_documents=True)
-    
-    question_generator = LLMChain(llm=chat_llm, prompt=CONDENSE_QUESTION_PROMPT) 
-    doc_chain = load_qa_chain(llm=chat_llm,chain_type="stuff",prompt=load_prompt()) 
-    chain = ConversationalRetrievalChain(retriever=_docsearch.as_retriever(search_kwags={"k": 3}), 
-                                     question_generator=question_generator, 
-                                     combine_docs_chain=doc_chain, 
-                                     memory=memory, 
-                                     return_source_documents=True, 
-                                     get_chat_history=lambda h :h)
-    
-    answer = chain({"question": query})
-
-    return answer
+    agent_executor = AgentExecutor(
+        agent=agent,
+        tools=tools,
+        verbose=True,
+        return_intermediate_steps=True,
+    )
+    
+    memory = AgentTokenBufferMemory(llm=llm)
+        
+    return memory, agent_executor
 
 @st.cache_data
 def gen_sentiment(text):