app.py

# Import required modules
import gradio as gr
import urllib.request
import fitz
import re
import numpy as np
import tensorflow_hub as hub
from sklearn.neighbors import NearestNeighbors
from transformers import AutoTokenizer, AutoModelForCausalLM
import transformers
import torch

# Load the Falcon model
model = "tiiuae/falcon-40b-instruct"
tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto",
)

# Load the PDF-GPT model
recommender = SemanticSearch()

# Define chat function
def chat():
    with gr.Interface(
        question_answer,
        [
            gr.inputs.Textbox(placeholder="Chat History", type="text", label="Chat History", lines=20),
            gr.inputs.Textbox(placeholder="Enter PDF URL here", type="text", label="URL"),
            gr.inputs.File(label="Or upload your PDF here"),
            gr.inputs.Textbox(placeholder="Enter your question here", type="text", label="Question"),
        ],
        gr.outputs.Textbox(placeholder="Chat History", type="text", label="Chat History", lines=20),
        title="Falcon-PDF Chatbot",
        description="A chatbot that can read and answer questions about a PDF document using the Falcon model",
        layout="vertical",
    ) as interface:
        with gr.Row():
            chatbot = gr.Chatbot(placeholder="Chat History", lines=20)
        with gr.Row():
            inputs = gr.Textbox(placeholder="Hello Falcon !!", label="Type an input and press Enter", max_lines=3)
            url = gr.Textbox(placeholder="Enter PDF URL here", label="URL")
            file = gr.File(label="Or upload your PDF here")
            question = gr.Textbox(placeholder="Enter your question here", label="Question")
            chat_button = gr.Button(label="Chat")

        chat_button.on_click(question_answer, [chatbot, url, file, question])

        with gr.Row():
            retry_button = gr.Button("♻️ Retry last turn")
            delete_turn_button = gr.Button("🧽 Delete last turn")
            clear_chat_button = gr.Button("✨ Delete all history")

        retry_button.on_click(retry_last_turn, [chatbot])
        delete_turn_button.on_click(delete_last_turn, [chatbot])
        clear_chat_button.on_click(clear_chat_history, [chatbot])

    # Launch the Gradio interface
    interface.launch()

def retry_last_turn(chat_history):
    """Handles retrying the last turn."""
    if len(chat_history) > 0:
        # Get the last question from the chat history
        last_question = chat_history[-1][0]
        # Remove the last turn from the chat history
        chat_history = chat_history[:-1]
        # Retry the last question
        question_answer(chat_history, last_question)
    else:
        print("Chat history is empty.")
    return chat_history

def delete_last_turn(chat_history):
    """Handles deleting the last turn."""
    if len(chat_history) > 0:
        # Remove the last turn from the chat history
        chat_history = chat_history[:-1]
    else:
        print("Chat history is empty.")
    return chat_history

def clear_chat_history(chat_history):
    """Handles clearing the chat history."""
    # Clear the chat history
    chat_history = []
    return chat_history


def download_pdf(url, output_path):
    """Download a PDF from a URL and save it to the specified output path."""
    urllib.request.urlretrieve(url, output_path)

def preprocess(text):
    """Preprocess a text by replacing newline characters with spaces and reducing multiple spaces to single spaces."""
    text = text.replace('\n', ' ')
    text = re.sub('\s+', ' ', text)
    return text

def pdf_to_text(path, start_page=1, end_page=None):
    """Extract text from a PDF file from the specified start page to the end page."""
    doc = fitz.open(path)
    total_pages = doc.page_count

    if end_page is None:
        end_page = total_pages

    text_list = []

    for i in range(start_page-1, end_page):
        text = doc.load_page(i).get_text("text")
        text = preprocess(text)
        text_list.append(text)

    doc.close()
    return text_list

def text_to_chunks(texts, word_length=150, start_page=1):
    """Split a list of texts into chunks with the specified word length."""
    text_toks = [t.split(' ') for t in texts]
    chunks = []

    for idx, words in enumerate(text_toks):
        for i in range(0, len(words), word_length):
            chunk = words[i:i+word_length]
            if (i+word_length) > len(words) and (len(chunk) < word_length) and (len(text_toks) != (idx+1)):
                text_toks[idx+1] = chunk + text_toks[idx+1]
                continue
            chunk = ' '.join(chunk).strip()
            chunk = f'[Page no. {idx+start_page}]' + ' ' + '"' + chunk + '"'
            chunks.append(chunk)
    return chunks

class SemanticSearch:
    """A class for performing semantic search using the Universal Sentence Encoder."""
    def __init__(self):
        self.use = hub.load('https://tfhub.dev/google/universal-sentence-encoder/4')
        self.fitted = False

    def fit(self, data, batch=1000, n_neighbors=5):
        """Fit the model to the data."""
        self.data = data
        self.embeddings = self.get_text_embedding(data, batch=batch)
        n_neighbors = min(n_neighbors, len(self.embeddings))
        self.nn = NearestNeighbors(n_neighbors=n_neighbors)
        self.nn.fit(self.embeddings)
        self.fitted = True

    def __call__(self, text, return_data=True):
        """Find the nearest neighbors to a text."""
        inp_emb = self.use([text])
        neighbors = self.nn.kneighbors(inp_emb, return_distance=False)[0]

        if return_data:
            return [self.data[i] for i in neighbors]
        else:
            return neighbors

    def get_text_embedding(self, texts, batch=1000):
        """Get the embeddings of a list of texts."""
        embeddings = []
        for i in range(0, len(texts), batch):
            text_batch = texts[i:(i+batch)]
            emb_batch = self.use(text_batch)
            embeddings.append(emb_batch)
        embeddings = np.vstack(embeddings)
        return embeddings

def load_recommender(path, start_page=1):
    """Load a recommender model with a PDF file."""
    global recommender
    texts = pdf_to_text(path, start_page=start_page)
    chunks = text_to_chunks(texts, start_page=start_page)
    recommender.fit(chunks)
    return 'Corpus Loaded.'


def generate_answer(question):
    topn_chunks = recommender(question)
    prompt = ""
    prompt += 'search results:\n\n'
    for c in topn_chunks:
        prompt += c + '\n\n'
        
    prompt += f"Query: {question}\nAnswer:"
    
    sequences = pipeline(
        prompt,
        max_length=200,
        do_sample=True,
        top_k=10,
        num_return_sequences=1,
        eos_token_id=tokenizer.eos_token_id,
    )

    return sequences[0]['generated_text']

def question_answer(chat_history, url, file, question):
    try:
        if url.strip() == '' and file is None:
            return '[ERROR]: Both URL and PDF is empty. Provide at least one.'
        if url.strip() != '' and file is not None:
            return '[ERROR]: Both URL and PDF is provided. Please provide only one (either URL or PDF).'
        if url.strip() != '':
            glob_url = url
            download_pdf(glob_url, 'corpus.pdf')
            load_recommender('corpus.pdf')
        else:
            old_file_name = file.name
            file_name = file.name
            file_name = file_name[:-12] + file_name[-4:]
            os.rename(old_file_name, file_name)
            load_recommender(file_name)
        if question.strip() == '':
            return '[ERROR]: Question field is empty'
        topn_chunks = recommender(question)
        prompt = ""
        prompt += 'search results:\n\n'
        for c in topn_chunks:
            prompt += c + '\n\n'
        
        prompt += "Instructions: Compose a comprehensive reply to the query using the search results given. "\
                  "Cite each reference using [ Page Number] notation (every result has this number at the beginning). "\
                  "Citation should be done at the end of each sentence. If the search results mention multiple subjects "\
                  "with the same name, create separate answers for each. Only include information found in the results and "\
                  "don't add any additional information. Make sure the answer is correct and don't output false content. "\
                  "If the text does not relate to the query, simply state 'Found Nothing'. Ignore outlier "\
                  "search results which has nothing to do with the question. Only answer what is asked. The "\
                  "answer should be short and concise. \n\nQuery: {question}\nAnswer: "
    
        prompt += f"Query: {question}\nAnswer:"
        
        sequences = pipeline(
           prompt,
            max_length=200,
            do_sample=True,
            top_k=10,
            num_return_sequences=1,
            eos_token_id=tokenizer.eos_token_id,
        )
        answer = sequences[0]['generated_text']
        chat_history.append([question, answer])
        return chat_history
    except Exception as e:
        return f'[ERROR]: {str(e)}'

questions = [
    "What did the study investigate?",
    "Can you provide a summary of this document?",
    "What are the methodologies used in this study?",
    "What are the data intervals used in this study? Give me the start dates and end dates.",
    "What are the main limitations of this study?",
    "What are the main shortcomings of this study?",
    "What are the main findings of the study?",
    "What are the main results of the study?",
    "What are the main contributions of this study?",
    "What is the conclusion of this paper?",
    "What are the input features used in this study?",
    "What is the dependent variable in this study?",
]


title = 'PDF GPT Turbo'
description = """ PDF GPT Turbo allows you to chat with your PDF file using Universal Sentence Encoder and Falcon. It gives hallucination free response than other tools as the embeddings are better than OpenAI. The returned response can even cite the page number in square brackets([]) where the information is located, adding credibility to the responses and helping to locate pertinent information quickly."""

with gr.Blocks(css="""#chatbot { font-size: 14px; min-height: 800px; }""") as demo:

    gr.Markdown(f'<center><h1>{title}</h1></center>')
    gr.Markdown(description)

    with gr.Row():
        
        with gr.Group():
            url = gr.Textbox(label='Enter PDF URL here   (Example: https://arxiv.org/pdf/1706.03762.pdf )')
            gr.Markdown("<center><h4>OR<h4></center>")
            file = gr.File(label='Upload your PDF/ Research Paper / Book here', file_types=['.pdf'])
            question = gr.Textbox(label='Enter your question here')
            gr.Examples(
                [[q] for q in questions],
                inputs=[question],
                label="PRE-DEFINED QUESTIONS: Click on a question to auto-fill the input box, then press Enter!",
            )
            btn = gr.Button(value='Submit')
            

            btn.style(full_width=True)

        with gr.Group():
            chatbot = gr.Chatbot(placeholder="Chat History", label="Chat History", lines=20, elem_id="chatbot")
            

    # Bind the click event of the button to the question_answer function
    btn.click(
        question_answer,
        inputs=[chatbot, url, file, question],
        outputs=[chatbot],
    )

demo.launch()