app.py

import gradio as gr
import random
import time
import openai

import os

openai.api_type = os.environ['OPENAI_API_TYPE']
openai.api_key = os.environ['OPENAI_API_KEY']
openai.api_base =  os.environ['OPENAI_API_BASE']
openai.api_version = os.environ['OPENAI_API_VERSION']

######################## Input TASK 1A  ########################
from langchain.document_loaders import PyPDFLoader, OnlinePDFLoader
from langchain.embeddings.sentence_transformer import SentenceTransformerEmbeddings
from langchain.text_splitter import CharacterTextSplitter

linkToPDF = os.environ['ONLINE_PDF_URL']

loader = OnlinePDFLoader(linkToPDF)
documents = loader.load()

chuck_size = 1000
chuck_overlap = 200
text_splitter = CharacterTextSplitter(chunk_size=chuck_size, chunk_overlap=chuck_overlap)
docs = text_splitter.split_documents(documents)
###########################END OF TASK 1A ##################################

######################## Input TASK 1B  ######3##############################
from langchain.vectorstores import Chroma

# create the open-source embedding function
embedding_function = SentenceTransformerEmbeddings(model_name="all-MiniLM-L6-v2")

# create simple ids - Index
ids = [str(i) for i in range(1, len(docs) + 1)]

# load it into Chroma
db = Chroma.from_documents(docs, embedding_function, ids=ids, collection_metadata={"hnsw:space": "cosine"} )
###########################END OF TASK 1B ##################################

###########################INPUT OF TASK 3B ##################################
# We try to limit the number of characters for the context
contextCharsLimit = 3072
promptStart = "Answer questions truthfully based on the information in sources provided below. \n If you cannot find the answer to a question based on the sources below, respond by saying “I apologize, but I am unable to provide an answer to your question, which is out of the scope of the document uploaded. Thank you! \n Sources:\n"

# We try to construct a function which can return the system prompt based on user query and fit in context into system prompt
def construct_system_prompt_chromadb (userQuery):
  ### find the number of relevant documents
  docs = db.similarity_search(userQuery)
  context = []

  ### append all relevant documents pagecontent into the context array
  for match in docs:
    context.append(match.page_content)

  ### loop for the context, check if the chars of context > limit, if not insert the pagecontent into the prompt with "-" or "\n" separator
  for i in range(1, len(context)):
    if len("-".join(context[:i])) >= contextCharsLimit:
      responsePrompt = promptStart + "-".join(context[:i-1])
    elif i == len(context)-1:
      responsePrompt = promptStart + "-".join(context)

  ## return the response rpompt
  return responsePrompt
########################### END OF TASK 3B ##################################



systemMessageContent = "" # System prompt we talked before - e.g., "You are a teaching assistant of a programming course CS1117. Try to answer student's question on python only"
systemMessage = {"role": "system", "content": systemMessageContent}

userMessageContent = "" # place holder
chatbotMessageContent = ""  # place holder

temperature = 0.8
top_p = 0.95
max_tokens = 800
numOfHistory = 5 # Add in the number history windows here



with gr.Blocks() as simpleChatDemo:
    inputMessages = gr.State([systemMessage])
#    inputMessages.append(systemMessage)

    # Chatbot interface
    chatbot = gr.Chatbot()

    # Message is a Text Box
    msg = gr.Textbox()

    # Clear Button on to clear up the msg and chatbot
    clear = gr.ClearButton([msg, chatbot])


    def respond(userMessageInput, inputMessagesHistory, chatbot_history):

        ## Construct the system message content based on the prompt function -- i.e., the input messages [0] will change based on system message now
        systemMessageContent = construct_system_prompt_chromadb(userMessageInput) # Change the system content to the function
        systemMessage = {"role": "system", "content": systemMessageContent}
        inputMessagesHistory[0] = systemMessage


        userMessageContent = userMessageInput
        userMessage = {"role": "user", "content": userMessageContent}

        inputMessagesHistory.append(userMessage)

        if len(inputMessagesHistory)  > numOfHistory + 1:
          numOutstandingMessages = len(inputMessagesHistory) - (numOfHistory + 1)
          inputMessagesHistory = [inputMessagesHistory[0], *inputMessagesHistory[1+numOutstandingMessages :]]
        print(inputMessages)
        completion = openai.ChatCompletion.create(engine="chatgpt", messages=inputMessagesHistory, temperature=temperature, top_p = top_p, max_tokens = max_tokens)

        chatbotMessageContent =  completion.choices[0].message.content
        chatbotMessage = {"role": "assistant", "content": chatbotMessageContent }
        inputMessagesHistory.append(chatbotMessage)

        # chat history is main list of [(user message string, bot message string)]
        chatbot_history.append((userMessageContent, chatbotMessageContent))

        time.sleep(2)
        # return with clear up the message box, and put the new messages into the chat_history
        return "", inputMessagesHistory, chatbot_history,

    # when the textbox click submit, i.e., enter, the function will be called (function, [input parameters], [output response])
    msg.submit(respond, [msg, inputMessages, chatbot], [msg, inputMessages, chatbot])

simpleChatDemo.launch(share=True)