app.py

import gradio as gr
from unstructured.partition.pdf import partition_pdf
import pymupdf
from PIL import Image
import numpy as np
import io
import pandas as pd
from langchain.text_splitter import RecursiveCharacterTextSplitter
import gc
import torch
import chromadb
from chromadb.utils.embedding_functions import OpenCLIPEmbeddingFunction
from chromadb.utils.data_loaders import ImageLoader
from sentence_transformers import SentenceTransformer
from chromadb.utils import embedding_functions
from transformers import LlavaNextProcessor, LlavaNextForConditionalGeneration
import base64
from langchain_community.llms import HuggingFaceEndpoint
from langchain import PromptTemplate
import spaces

if torch.cuda.is_available():
    processor = LlavaNextProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")
    vision_model = LlavaNextForConditionalGeneration.from_pretrained(
        "llava-hf/llava-v1.6-mistral-7b-hf",
        torch_dtype=torch.float16,
        low_cpu_mem_usage=True,
        load_in_4bit=True,
    )


def image_to_bytes(image):
    img_byte_arr = io.BytesIO()
    image.save(img_byte_arr, format="PNG")
    return base64.b64encode(img_byte_arr.getvalue()).decode("utf-8")


@spaces.GPU(duration=60*4)
def get_image_descriptions(images):
    torch.cuda.empty_cache()
    gc.collect()

    descriptions = []
    prompt = "[INST] <image>\nDescribe the image in a sentence [/INST]"

    for img in images:
        inputs = processor(prompt, img, return_tensors="pt").to("cuda:0")
        output = vision_model.generate(**inputs, max_new_tokens=100)
        descriptions.append(processor.decode(output[0], skip_special_tokens=True))
    return descriptions


CSS = """
#table_col {background-color: rgb(33, 41, 54);}
"""


def extract_pdfs(docs, doc_collection):
    if docs:
        doc_collection = []
        doc_collection.extend(docs)
    return (
        doc_collection,
        gr.Tabs(selected=1),
        pd.DataFrame([i.split("/")[-1] for i in list(docs)], columns=["Filename"]),
    )


def extract_images(docs):
    images = []
    for doc_path in docs:
        doc = pymupdf.open(doc_path)  # open a document

        for page_index in range(len(doc)):  # iterate over pdf pages
            page = doc[page_index]  # get the page
            image_list = page.get_images()

            for image_index, img in enumerate(
                image_list, start=1
            ):  # enumerate the image list
                xref = img[0]  # get the XREF of the image
                pix = pymupdf.Pixmap(doc, xref)  # create a Pixmap

                if pix.n - pix.alpha > 3:  # CMYK: convert to RGB first
                    pix = pymupdf.Pixmap(pymupdf.csRGB, pix)

                images.append(Image.open(io.BytesIO(pix.pil_tobytes("JPEG"))))
    return images


# def get_vectordb(text, images, tables):
def get_vectordb(text, images):
    client = chromadb.EphemeralClient()
    loader = ImageLoader()
    sentence_transformer_ef = embedding_functions.SentenceTransformerEmbeddingFunction(
        model_name="multi-qa-mpnet-base-dot-v1"
    )
    if "text_db" in [i.name for i in client.list_collections()]:
        client.delete_collection("text_db")
    if "image_db" in [i.name for i in client.list_collections()]:
        client.delete_collection("image_db")
    
    text_collection = client.get_or_create_collection(
        name="text_db",
        embedding_function=sentence_transformer_ef,
        data_loader=loader,
    )
    image_collection = client.get_or_create_collection(
        name="image_db",
        embedding_function=sentence_transformer_ef,
        data_loader=loader,
        metadata={"hnsw:space": "cosine"},
    )

    image_descriptions = get_image_descriptions(images)
    image_dict = [{"image": image_to_bytes(img) for img in images}]

    if len(images)>0:
        image_collection.add(
            ids=[str(i) for i in range(len(images))],
            documents=image_descriptions,
            metadatas=image_dict,
        )

    splitter = RecursiveCharacterTextSplitter(
        chunk_size=500,
        chunk_overlap=10,
    )

    if len(text)>0:
        docs = splitter.create_documents([text])
        doc_texts = [i.page_content for i in docs]
        text_collection.add(
            ids=[str(i) for i in list(range(len(doc_texts)))], documents=doc_texts
        )
    return client


def extract_data_from_pdfs(docs, session, progress=gr.Progress()):
    if len(docs) == 0:
        raise gr.Error("No documents to process")
    progress(0, "Extracting Images")

    images = extract_images(docs)

    progress(0.25, "Extracting Text")

    strategy = "hi_res"
    model_name = "yolox"
    all_elements = []

    for doc in docs:
        elements = partition_pdf(
            filename=doc,
            strategy=strategy,
            infer_table_structure=True,
            model_name=model_name,
        )

        all_elements.extend(elements)

    all_text = ""

    # tables = []

    prev = None
    for i in all_elements:
        meta = i.to_dict()
        if meta["type"].lower() not in ["table", "figurecaption"]:
            if meta["type"].lower() in ["listitem", "title"]:
                all_text += "\n\n" + meta["text"] + "\n"
            else:
                all_text += meta["text"]
        elif meta["type"] == "Table":
            continue
            # tables.append(meta["metadata"]["text_as_html"])

    # html = "<br>".join(tables)
    # display = "<h3>Sample Tables</h3>" + "<br>".join(tables[:2])
    # html = gr.HTML(html)
    # vectordb = get_vectordb(all_text, images, tables)

    progress(0.5, "Generating image descriptions")
    image_descriptions = "\n".join(get_image_descriptions(images))

    progress(0.75, "Inserting data into vector database")
    vectordb = get_vectordb(all_text, images)

    progress(1, "Completed")
    session["processed"] = True
    return (
        vectordb,
        session,
        gr.Row(visible=True),
        all_text[:2000] + "...",
        # display,
        images[:2],
        "<h1 style='text-align: center'>Completed<h1>",
        # image_descriptions
    )


sentence_transformer_ef = embedding_functions.SentenceTransformerEmbeddingFunction(
    model_name="multi-qa-mpnet-base-dot-v1"
)


def conversation(vectordb_client, msg, num_context, img_context, history):

    text_collection = vectordb_client.get_collection(
        "text_db", embedding_function=sentence_transformer_ef
    )
    image_collection = vectordb_client.get_collection(
        "image_db", embedding_function=sentence_transformer_ef
    )

    results = text_collection.query(
        query_texts=[msg], include=["documents"], n_results=num_context
    )["documents"][0]
    
    similar_images = image_collection.query(
        query_texts=[msg],
        include=["metadatas", "distances", "documents"],
        n_results=img_context,
    )
    img_links = [i["image"] for i in similar_images["metadatas"][0]]

    images_and_locs = [
        Image.open(io.BytesIO(base64.b64decode(i[1])))
        for i in zip(similar_images["distances"][0], img_links)
    ]
    img_desc = "\n".join(similar_images["documents"][0])
    if len(img_links) == 0:
        img_desc = "No Images Are Provided"
    template = """
    Context:
    {context}

    Included Images:
    {images}
    
    Question:
    {question}

    Answer:

    """
    prompt = PromptTemplate(template=template, input_variables=["context", "question"])
    context = "\n\n".join(results)
    # references = [gr.Textbox(i, visible=True, interactive=False) for i in results]
    response = llm(prompt.format(context=context, question=msg, images=img_desc))
    yield history + [(msg, response)], results, images_and_locs


def check_validity_and_llm(session_states):
    if session_states.get("processed", False) == True:
        return gr.Tabs(selected=2)
    raise gr.Error("Please extract data first")


def get_stats(vectordb):
    eles = vectordb.get()
    # words =
    text_data = [f"Chunks: {len(eles)}", "HIII"]
    return "\n".join(text_data), "", ""


llm = HuggingFaceEndpoint(
    repo_id="mistralai/Mixtral-8x7B-Instruct-v0.1",
    temperature=0.4,
    max_new_tokens=800,
)

with gr.Blocks(css=CSS) as demo:

    vectordb = gr.State()
    doc_collection = gr.State(value=[])
    session_states = gr.State(value={})
    references = gr.State(value=[])
    
    gr.Markdown(
        """<h2><center>Multimodal PDF Chatbot</center></h2>
    <h3><center><b>Interact With Your PDF Documents</b></center></h3>"""
    )
    gr.Markdown(
        """<center><h3><b>Note: </b> This application leverages advanced Retrieval-Augmented Generation (RAG) techniques to provide context-aware responses from your PDF documents</center><h3><br>
    <center>Utilizing multimodal capabilities, this chatbot can interpret and answer queries based on both textual and visual information within your PDFs.</center>"""
    )
    gr.Markdown(
        """
    <center><b>Warning: </b> Extracting text and images from your document and generating embeddings may take some time due to the use of OCR and multimodal LLMs for image description<center>
    """
    )
    with gr.Tabs() as tabs:
        with gr.TabItem("Upload PDFs", id=0) as pdf_tab:
            with gr.Row():
                with gr.Column():
                    documents = gr.File(
                        file_count="multiple",
                        file_types=["pdf"],
                        interactive=True,
                        label="Upload your PDF file/s",
                    )
                    pdf_btn = gr.Button(value="Next", elem_id="button1")

        with gr.TabItem("Extract Data", id=1) as preprocess:
            with gr.Row():
                with gr.Column():
                    back_p1 = gr.Button(value="Back")
                with gr.Column():
                    embed = gr.Button(value="Extract Data")
                with gr.Column():
                    next_p1 = gr.Button(value="Next")

            with gr.Row() as row:
                with gr.Column():
                    selected = gr.Dataframe(
                        interactive=False,
                        col_count=(1, "fixed"),
                        headers=["Selected Files"],
                    )
                with gr.Column(variant="panel"):
                    prog = gr.HTML(
                        value="<h1 style='text-align: center'>Click the 'Extract' button to extract data from PDFs<h1>"
                    )

            with gr.Accordion("See Parts of Extracted Data", open=False):
                with gr.Column(visible=True) as sample_data:
                    with gr.Row():
                        with gr.Column():
                            ext_text = gr.Textbox(
                                label="Sample Extracted Text", lines=15
                            )
                        with gr.Column():
                            images = gr.Gallery(
                                label="Sample Extracted Images", columns=1, rows=2
                            )



        with gr.TabItem("Chat", id=2) as chat_tab:
            with gr.Column():
                choice = gr.Radio(
                    ["chromaDB"],
                    value="chromaDB",
                    label="Vector Database",
                    interactive=True,
                )
                num_context = gr.Slider(
                    label="Number of text context elements",
                    minimum=1,
                    maximum=20,
                    step=1,
                    interactive=True,
                    value=3,
                )
                img_context = gr.Slider(
                    label="Number of image context elements",
                    minimum=1,
                    maximum=10,
                    step=1,
                    interactive=True,
                    value=2,
                )
            with gr.Row():
                with gr.Column():
                    ret_images = gr.Gallery("Similar Images", columns=1, rows=2)
                with gr.Column():
                    chatbot = gr.Chatbot(height=400)
            with gr.Accordion("Text References", open=False):
                # text_context = gr.Row()
                
                @gr.render(inputs=[references])
                def gen_refs(refs):
                    n = len(refs)
                    for i in range(n):
                        gr.Textbox(label=f"Ref-{i+1}", value=refs[i], lines=3)
                

            with gr.Row():
                msg = gr.Textbox(
                    placeholder="Type your question here (e.g. 'What is this document about?')",
                    interactive=True,
                    container=True,
                )
            with gr.Row():
                submit_btn = gr.Button("Submit message")
                clear_btn = gr.ClearButton([msg, chatbot], value="Clear conversation")

    pdf_btn.click(
        fn=extract_pdfs,
        inputs=[documents, doc_collection],
        outputs=[doc_collection, tabs, selected],
    )
    embed.click(
        extract_data_from_pdfs,
        inputs=[doc_collection, session_states],
        outputs=[
            vectordb,
            session_states,
            sample_data,
            ext_text,
            images,
            prog,
        ],
    )

    submit_btn.click(
        conversation,
        [vectordb, msg, num_context, img_context, chatbot],
        [chatbot, references, ret_images],
    )


    back_p1.click(lambda: gr.Tabs(selected=0), None, tabs)

    next_p1.click(check_validity_and_llm, session_states, tabs)
if __name__ == "__main__":
    demo.launch(share=True)