app.py

import base64
from langchain.chains.summarize import load_summarize_chain
from langchain.docstore.document import Document
from langchain.document_loaders.pdf import PyMuPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from PyPDF2 import PdfReader
import re
import streamlit as st
import sys
import time
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForCausalLM
from transformers import pipeline

# notes
# https://huggingface.co/docs/transformers/pad_truncation


# file loader and preprocessor
def file_preprocessing(file, skipfirst, skiplast):
    loader = PyMuPDFLoader(file)
    pages = loader.load_and_split()
    # skip page(s)
    if (skipfirst == 1) & (skiplast == 0):
        del pages[0]
    elif (skipfirst == 0) & (skiplast == 1):
        del pages[-1]
    elif (skipfirst == 1) & (skiplast == 1):
        del pages[0]
        del pages[-1]
    else:
        pages = pages
    # https://stackoverflow.com/questions/76431655/langchain-pypdfloader
    content = ""
    for page in pages:
        content = content + page.page_content
    content = re.sub("-\n", "", content)
    print("\n###### New article ######\n")
    print("Input text:\n")
    print(content)
    print("\nChunking...")
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=1000,  # number of characters
        chunk_overlap=100,
        length_function=len,
        separators=["\n\n", "\n", " ", ""],  # default list
    )
    # https://dev.to/eteimz/understanding-langchains-recursivecharactertextsplitter-2846
    texts = text_splitter.split_text(content)
    print("Number of tokens: " + str(len(texts)))
    print("\nFirst three tokens:\n")
    print(texts[0])
    print("")
    print(texts[1])
    print("")
    print(texts[2])
    print("")
    final_texts = ""
    for text in texts:
        final_texts = final_texts + text
    return texts, final_texts


# function to count words in the input
def preproc_count(filepath, skipfirst, skiplast):
    texts, input_text = file_preprocessing(filepath, skipfirst, skiplast)
    input_text = input_text.replace("-", "")
    text_length = len(re.findall(r"\w+", input_text))
    print("Input word count: " f"{text_length:,}")
    return texts, input_text, text_length


# function to covert (bart) summary to sentence case
def convert_to_sentence_case(text):
    sentences = re.split(r"(?<=[.!?])\s+", text)
    formatted_sentences = [sentence.capitalize() for sentence in sentences]
    return " ".join(formatted_sentences)


# llm pipeline
def llm_pipeline(tokenizer, base_model, input_text, model_source):
    pipe_sum = pipeline(
        "summarization",
        model=base_model,
        tokenizer=tokenizer,
        max_length=300,
        min_length=200,
        truncation=True,
    )
    print("Model source: %s" % (model_source))
    print("Summarizing...")
    result = pipe_sum(input_text)
    summary = result[0]["summary_text"]
    print("Summarization finished\n")
    print("Summary text:\n")
    print(summary)
    print("")
    return summary


# function to count words in the summary
def postproc_count(summary):
    text_length = len(re.findall(r"\w+", summary))
    print("Summary word count: " f"{text_length:,}")
    return text_length


# function to clean summary text
def clean_summary_text(summary):
    # remove whitespace
    summary_clean_1 = summary.strip()
    # remove spaces before punctuation (bart)
    summary_clean_2 = re.sub(r'\s([,.():;?!"](?:\s|$))', r"\1", summary_clean_1)
    # convert to sentence case
    summary_clean_3 = convert_to_sentence_case(summary_clean_2)
    return summary_clean_3


@st.cache_data(ttl=60 * 60)
# function to display the PDF
def displayPDF(file):
    with open(file, "rb") as f:
        base64_pdf = base64.b64encode(f.read()).decode("utf-8")
    # embed pdf in html
    pdf_display = f'<iframe src="data:application/pdf;base64,{base64_pdf}" width="100%" height="600" type="application/pdf"></iframe>'
    # display file
    st.markdown(pdf_display, unsafe_allow_html=True)


# streamlit code
st.set_page_config(layout="wide")


def main():
    st.title("RASA: Research Article Summarization App")
    uploaded_file = st.file_uploader("Upload your PDF file", type=["pdf"])
    if uploaded_file is not None:
        st.subheader("Options")
        col1, col2, col3, col4 = st.columns([1, 1, 1, 2])
        with col1:
            model_source_names = ["Cached model", "Download model"]
            model_source = st.radio(
                "For development:",
                model_source_names,
                help="Defaults to a cached model; downloading will take longer",
            )
        with col2:
            model_names = [
                "T5-Small",
                "BART",
            ]
            selected_model = st.radio(
                "Select a model to use:",
                model_names,
                help="Defauls to T5-Small; for most articles it summarizes better than BART",
            )
            if selected_model == "BART":
                checkpoint = "ccdv/lsg-bart-base-16384-pubmed"
                tokenizer = AutoTokenizer.from_pretrained(
                    checkpoint,
                    truncation=True,
                    model_max_length=1000,
                    trust_remote_code=True,
                )
                if model_source == "Download model":
                    base_model = AutoModelForSeq2SeqLM.from_pretrained(
                        checkpoint,
                        torch_dtype=torch.float32,
                        trust_remote_code=True,
                    )
                else:
                    base_model = "model_cache/models--ccdv--lsg-bart-base-16384-pubmed/snapshots/4072bc1a7a94e2b4fd860a5fdf1b71d0487dcf15"
            else:
                checkpoint = "MBZUAI/LaMini-Flan-T5-77M"
                tokenizer = AutoTokenizer.from_pretrained(
                    checkpoint,
                    truncation=True,
                    legacy=False,
                    model_max_length=1000,
                )
                if model_source == "Download model":
                    base_model = AutoModelForSeq2SeqLM.from_pretrained(
                        checkpoint,
                        torch_dtype=torch.float32,
                    )
                else:
                    base_model = "model_cache/models--MBZUAI--LaMini-Flan-T5-77M/snapshots/c5b12d50a2616b9670a57189be20055d1357b474"
        with col3:
            st.write("Skip any pages?")
            skipfirst = st.checkbox(
                "Skip first page", help="Select if your PDF has a cover page"
            )
            skiplast = st.checkbox("Skip last page")
        with col4:
            st.write("Background information (links open in a new window)")
            st.write(
                "Model class: [T5-Small](https://huggingface.co/docs/transformers/main/en/model_doc/t5)"
                "&nbsp;&nbsp;|&nbsp;&nbsp;Model: [LaMini-Flan-T5-77M](https://huggingface.co/MBZUAI/LaMini-Flan-T5-77M)"
            )
            st.write(
                "Model class: [BART](https://huggingface.co/docs/transformers/main/en/model_doc/bart)"
                "&nbsp;&nbsp;|&nbsp;&nbsp;Model: [lsg-bart-base-16384-pubmed](https://huggingface.co/ccdv/lsg-bart-base-16384-pubmed)"
            )
        if st.button("Summarize"):
            col1, col2 = st.columns(2)
            filepath = "data/" + uploaded_file.name
            with open(filepath, "wb") as temp_file:
                temp_file.write(uploaded_file.read())
            with col1:
                texts, input_text, preproc_text_length = preproc_count(
                    filepath, skipfirst, skiplast
                )
                st.info(
                    "Uploaded PDF&nbsp;&nbsp;|&nbsp;&nbsp;Number of words: "
                    f"{preproc_text_length:,}"
                )
                pdf_viewer = displayPDF(filepath)
            with col2:
                start = time.time()
                with st.spinner("Summarizing..."):
                    summary = llm_pipeline(
                        tokenizer, base_model, input_text, model_source
                    )
                    postproc_text_length = postproc_count(summary)
                end = time.time()
                duration = end - start
                print("Duration: " f"{duration:.0f}" + " seconds")
                st.info(
                    "PDF Summary&nbsp;&nbsp;|&nbsp;&nbsp;Number of words: "
                    f"{postproc_text_length:,}"
                    + "&nbsp;&nbsp;|&nbsp;&nbsp;Summarization time: "
                    f"{duration:.0f}" + " seconds"
                )
                if selected_model == "BART":
                    summary_cleaned = clean_summary_text(summary)
                    st.success(summary_cleaned)
                    with st.expander("Raw output"):
                        st.write(summary)
                else:
                    st.success(summary)
            col1 = st.columns(1)
            url = "https://dev.to/eteimz/understanding-langchains-recursivecharactertextsplitter-2846"
            st.info("Additional information")
            st.write("\n[RecursiveCharacterTextSplitter](%s) parameters used:" % url)
            st.write("&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;chunk_size=1000")
            st.write(
                "&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;chunk_overlap=100"
            )
            st.write(
                "&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;length_function=len"
            )
            st.write("")
            st.write("Number of tokens generated: " + str(len(texts)))
            st.write("")
            st.write("First three tokens:")
            st.write("----")
            st.write(texts[0])
            st.write("----")
            st.write(texts[1])
            st.write("----")
            st.write(texts[2])


st.markdown(
    """<style>
div[class*="stRadio"] > label > div[data-testid="stMarkdownContainer"] > p {
    font-size: 1rem;
    font-weight: 400;
}
div[class*="stMarkdown"] > div[data-testid="stMarkdownContainer"] > p {
    margin-bottom: -15px;
}
div[class*="stCheckbox"] > label[data-baseweb="checkbox"] {
    margin-bottom: -15px;
}
body > a {
    text-decoration: underline;
}
    </style>
    """,
    unsafe_allow_html=True,
)


if __name__ == "__main__":
    main()