app.py

import os
import pickle
from nltk.tokenize import sent_tokenize
import nltk
from PyPDF2 import PdfReader
import streamlit as st
from streamlit_extras.add_vertical_space import add_vertical_space
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.vectorstores import FAISS
from langchain.llms import OpenAI
from langchain.chains.question_answering import load_qa_chain
from langchain.callbacks import get_openai_callback

nltk.download('punkt')

# Sidebar contents
with st.sidebar:
    st.title(':orange_book: BinDoc GmbH')
    
    # API key input
    api_key = st.text_input('Enter your OpenAI API Key:', type='password')
    
    if api_key:
        os.environ['OPENAI_API_KEY'] = api_key
    else:
        st.warning('API key is required to proceed.')

    st.markdown("Experience the future of document interaction with the revolutionary")
    st.markdown("**BinDocs Chat App**.")
    st.markdown("Harnessing the power of a Large Language Model and AI technology,")
    st.markdown("this innovative platform redefines PDF engagement,")
    st.markdown("enabling dynamic conversations that bridge the gap between")
    st.markdown("human and machine intelligence.")

    add_vertical_space(3)  # Add more vertical space between text blocks
    st.write('Made with ❤️ by BinDoc GmbH')

def load_pdf(file_path):
    pdf_reader = PdfReader(file_path)
    chunks = []
    for page in pdf_reader.pages:
        text = page.extract_text()
        if text:
            chunks.append(text)
    
    store_name = file_path.name[:-4]
    
    if os.path.exists(f"{store_name}.pkl"):
        with open(f"{store_name}.pkl", "rb") as f:
            VectorStore = pickle.load(f)
    else:
        embeddings = OpenAIEmbeddings()
        VectorStore = FAISS.from_texts(chunks, embedding=embeddings)
        with open(f"{store_name}.pkl", "wb") as f:
            pickle.dump(VectorStore, f)

    return VectorStore

def load_chatbot(max_tokens=120):
    return load_qa_chain(llm=OpenAI(temperature=0.5, max_tokens=max_tokens), chain_type="stuff") 

def display_chat_history(chat_history):
    for chat in chat_history:
        background_color = "#FFA07A" if chat[2] == "new" else "#acf" if chat[0] == "User" else "#caf"
        st.markdown(f"<div style='background-color: {background_color}; padding: 10px; border-radius: 10px; margin: 10px;'>{chat[0]}: {chat[1]}</div>", unsafe_allow_html=True)

def remove_incomplete_sentences(text):
    sentences = sent_tokenize(text)
    complete_sentences = [sent for sent in sentences if sent.endswith(('.', '!', '?'))]
    return ' '.join(complete_sentences)

def remove_redundant_information(text):
    sentences = sent_tokenize(text)
    unique_sentences = list(set(sentences))
    return ' '.join(unique_sentences)

# Define a maximum token limit to avoid infinite loops
MAX_TOKEN_LIMIT = 400

def main():
    st.title("BinDocs Chat App")

    if "chat_history" not in st.session_state:
        st.session_state['chat_history'] = []

    display_chat_history(st.session_state['chat_history'])

    st.write("<!-- Start Spacer -->", unsafe_allow_html=True)
    st.write("<div style='flex: 1;'></div>", unsafe_allow_html=True)
    st.write("<!-- End Spacer -->", unsafe_allow_html=True)

    new_messages_placeholder = st.empty()

    pdf = st.file_uploader("Upload your PDF", type="pdf")

    if pdf is not None:
        query = st.text_input("Ask questions about your PDF file (in any preferred language):")

        if st.button("Ask") or (query and query != st.session_state.get('last_input', '')):
            st.session_state['last_input'] = query
            st.session_state['chat_history'].append(("User", query, "new"))

            loading_message = st.empty()
            loading_message.text('Bot is thinking...')
            
            VectorStore = load_pdf(pdf)
            max_tokens = 100
            chain = load_chatbot(max_tokens=max_tokens)
            docs = VectorStore.similarity_search(query=query, k=2)
            
            with get_openai_callback() as cb:
                response = chain.run(input_documents=docs, question=query)
            
            # Post-processing to remove incomplete sentences and redundant information
            filtered_response = remove_incomplete_sentences(response)
            filtered_response = remove_redundant_information(filtered_response)
            
            # Check if the response ends with a sentence-ending punctuation
            while not filtered_response.strip().endswith(('.', '!', '?')) and max_tokens < MAX_TOKEN_LIMIT:
                max_tokens += 50  # Increase the max_tokens limit
                chain = load_chatbot(max_tokens=max_tokens)
                additional_response = chain.run(input_documents=docs, question=query)
                filtered_response += additional_response  # Append the additional response to the filtered_response

            st.session_state['chat_history'].append(("Bot", filtered_response, "new"))

            # Display new messages at the bottom
            new_messages = st.session_state['chat_history'][-2:]
            for chat in new_messages:
                background_color = "#FFA07A" if chat[2] == "new" else "#acf" if chat[0] == "User" else "#caf"
                new_messages_placeholder.markdown(f"<div style='background-color: {background_color}; padding: 10px; border-radius: 10px; margin: 10px;'>{chat[0]}: {chat[1]}</div>", unsafe_allow_html=True)

            # Scroll to the latest response using JavaScript
            st.write("<script>document.getElementById('response').scrollIntoView();</script>", unsafe_allow_html=True)

            loading_message.empty()

            # Clear the input field by setting the query variable to an empty string
            query = ""

        # Mark all messages as old after displaying
        st.session_state['chat_history'] = [(sender, msg, "old") for sender, msg, _ in st.session_state['chat_history']]

if __name__ == "__main__":
    main()