conversation.py

from langchain.document_loaders import TextLoader
import pinecone
from langchain.vectorstores import Pinecone
import os
from transformers import AutoTokenizer, AutoModel
from langchain.agents.agent_toolkits import create_conversational_retrieval_agent
from langchain.agents.agent_toolkits import create_retriever_tool
from langchain.chat_models import ChatOpenAI
import torch
from langchain.agents.openai_functions_agent.agent_token_buffer_memory import (AgentTokenBufferMemory)
from langchain.agents.openai_functions_agent.base import OpenAIFunctionsAgent
from langchain.schema.messages import SystemMessage
from langchain.prompts import MessagesPlaceholder
import gradio as gr
import time
from db_func import insert_one

def get_bert_embeddings(sentence):
    embeddings = []
    input_ids = tokenizer.encode(sentence, return_tensors="pt")
    with torch.no_grad():
        output = model(input_ids)
    embedding = output.last_hidden_state[:,0,:].numpy().tolist()
    return embedding

model_name = "BAAI/bge-base-en-v1.5"
model = AutoModel.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
prompt_file = open("prompts/version_2.txt", "r").read()

pinecone.init(
    api_key=os.getenv("PINECONE_API_KEY"),  # find at app.pinecone.io
    environment=os.getenv("PINECONE_ENV"),  # next to api key in console
)

index_name = "ophtal-knowledge-base"
index = pinecone.Index(index_name)
vectorstore = Pinecone(index, get_bert_embeddings, "text")
retriever = vectorstore.as_retriever()
tool = create_retriever_tool(
    retriever,
    "search_ophtal-knowledge-base",
    "Searches and returns documents regarding the ophtal-knowledge-base.",
)
tools = [tool]
system_message = SystemMessage(content=prompt_file)
memory_key='history'

llm = ChatOpenAI(openai_api_key=os.getenv("OPENAI_API_KEY"), model="gpt-4", temperature=0.2)
prompt = OpenAIFunctionsAgent.create_prompt(
    system_message=system_message,
    extra_prompt_messages=[MessagesPlaceholder(variable_name=memory_key)],
)

agent_executor = create_conversational_retrieval_agent(llm, tools, verbose=False, prompt=prompt)
user_name = None

def run(input_):
    output = agent_executor({"input": input_})
    output_text = output["output"]
    source_text = ""
    doc_text = ""    

    if len(output["intermediate_steps"])>0:
        documents = output["intermediate_steps"][0][1]
        sources = []
        docs = []

        for doc in documents:
            if doc.metadata["source"] not in sources:
                sources.append(doc.metadata["source"])
                docs.append(doc.page_content)

        for i in range(len(sources)):
            temp = sources[i].replace('.pdf',  '').replace('.txt', '').replace("AAO", "").replace("2022-2023", "").replace("data/book", "").replace("text", "").replace("  ", " ")
            source_text += f"{i+1}. {temp}\n"
            doc_text += f"{i+1}. {docs[i]}\n"

        output_text =  f"{output_text} \n\nSources: \n{source_text}\n\nDocuments: \n{doc_text}"
    
    doc_to_insert = {
        "user": user_name,
        "input": input_,
        "output": output_text,
        "source": source_text,
        "documents": doc_text
    }

    insert_one(doc_to_insert)

    return output_text

def make_conversation(message, history):
    text_ = run(message)
    for i in range(len(text_)):
        time.sleep(0.001)
        yield text_[: i+1]
    
def auth_function(username, password):
    user_name = username
    return username == password