streamlit_app.py

import streamlit as st

# Set a custom background
import torch
from langchain import HuggingFacePipeline
from langchain.chains import LLMChain, RetrievalQA
from langchain.document_loaders import (
    DirectoryLoader,
    PyPDFLoader,
    TextLoader,
    UnstructuredPDFLoader,
)
from langchain.embeddings import HuggingFaceEmbeddings, LlamaCppEmbeddings
from langchain.llms import LlamaCpp
from langchain.prompts import PromptTemplate
from langchain.text_splitter import (
    CharacterTextSplitter,
    RecursiveCharacterTextSplitter,
)
from langchain.vectorstores import Chroma
from PIL import Image
from streamlit_extras.add_vertical_space import add_vertical_space

st.set_page_config(page_title="Welcome to our AI Question Answering Bot")

with st.sidebar:
    st.title('🤗💬 QA App')
    st.markdown('''
    ## About
    This app is an LLM-powered chatbot built using:
    - [Streamlit](<https://streamlit.io/>)
    - [HugChat](<https://github.com/Soulter/hugging-chat-api>)
    - Chat Model = llama2-chat-hf 7B 
    - Retreiver model = all-MiniLM-L6-v2
    
    💡 Note: No API key required!
    ''')
    add_vertical_space(5)
    st.write('Made with ❤️ by us')

# logo = Image.open('logo.png')
# st.image(logo, use_column_width=True)


# Introduction
st.markdown("""
Welcome! This is not just any bot, it's a special one equipped with state-of-the-art natural language processing capabilities, and ready to answer your queries.


Ready to explore? Let's get started!

* Step 1: Upload a PDF document.
* Step 2: Type in a question related to your document's content.
* Step 3: Get your answer!

Push clear cache before uploading a new doc !


""")


def write_text_file(content, file_path):
    try:
        with open(file_path, 'wb') as file:
            file.write(content)
        return True
    except Exception as e:
        print(f"Error occurred while writing the file: {e}")
        return False


# Wrap prompt template in a PromptTemplate object

def set_qa_prompt():
    # set prompt template
    prompt_template = """<s>[INST] <<SYS>> Use the following pieces of context closed between $ to answer the question closed between |. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.
    ${context}$ <</SYS>>
    Question: |{question}|
    Answer:[/INST]</s>"""
    prompt = PromptTemplate(
        template=prompt_template, input_variables=["context", "question"]
    )
    return prompt


# Build RetrievalQA object

def build_retrieval_qa(_llm, _prompt, _vectorstore):
    dbqa = RetrievalQA.from_chain_type(llm=_llm,
                                       chain_type='stuff',
                                       retriever=_vectorstore.as_retriever(search_kwargs={'k': 3}),
                                       return_source_documents=True,
                                       chain_type_kwargs={'prompt': _prompt})
    return dbqa


# Instantiate QA object
# @st.cache(allow_output_mutation=True)
# @st.cache_resource()
@st.cache(allow_output_mutation=True)
def setup_dbqa(_texts):
    print("setup_dbqa ...")
    llm = HuggingFacePipeline.from_model_id(
        model_id="NousResearch/Llama-2-13b-chat-hf",
        task="text-generation",
        model_kwargs={
            "max_length": 1500, "load_in_8bit": True},
    )

    embeddings = HuggingFaceEmbeddings(model_name='sentence-transformers/all-MiniLM-L6-v2',
                                       model_kwargs={'device': 'cpu'})

    vectorstore = Chroma.from_documents(texts, embeddings, persist_directory="vectorstore")

    prompt = set_qa_prompt()

    return build_retrieval_qa(llm, prompt, vectorstore)


def load_docs(uploaded_file):
    print("loading docs ...")
    content = uploaded_file.read()
    file_path_aux = "./temp/file.pdf"
    write_text_file(content, file_path_aux)
    file_path = "./temp/"

    loader = DirectoryLoader(file_path,
                             glob="*.pdf",
                             loader_cls=UnstructuredPDFLoader)
    documents = loader.load()

    # Split text from PDF into chunks
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000,
                                                   chunk_overlap=0,
                                                   length_function=len,)
    texts = text_splitter.split_documents(documents)
    return texts


# Set the background image
# Load a PDF file
uploaded_file = st.file_uploader("Choose a PDF file", type="pdf")

if uploaded_file is not None:
    st.write('Loading file')

    texts = load_docs(uploaded_file)
    model = setup_dbqa(texts)

    # Build and persist FAISS vector store

    question = st.text_input('Ask a question:')

    if question:
        # Placeholder for chatbot logic to generate an answer based on the question and the PDF content

        answer = model({'query': question})
        # The below is just a hardcoded response
        print(question)
        print(answer)

        # st.write('Question: ', answer["query"])
        st.write('Question: ', answer["query"])
        st.write('Answer: ', answer["result"])
        st.write('Source documents: ', answer["source_documents"])

# if st.button("Clear cache before loading new document"):
#     # Clears all st.cache_resource caches:
#     st.cache_resource.clear()