app.py

import tqdm
from PIL import Image
import hashlib
import torch
import torch.nn.functional as F
import fitz
import threading
import gradio as gr
import spaces
import os
from transformers import AutoModel
from transformers import AutoTokenizer
from PIL import Image
import torch
import os
import numpy as np
import json

cache_dir = '/data/KB'
os.makedirs(cache_dir, exist_ok=True)

@spaces.GPU(duration=100)
def weighted_mean_pooling(hidden, attention_mask):
    attention_mask_ = attention_mask * attention_mask.cumsum(dim=1)
    s = torch.sum(hidden * attention_mask_.unsqueeze(-1).float(), dim=1)
    d = attention_mask_.sum(dim=1, keepdim=True).float()
    reps = s / d
    return reps

@spaces.GPU(duration=100)
@torch.no_grad()
def encode(text_or_image_list):
    global model, tokenizer
    if (isinstance(text_or_image_list[0], str)):
        inputs = {
            "text": text_or_image_list,
            'image': [None] * len(text_or_image_list),
            'tokenizer': tokenizer
        }
    else:
        inputs = {
            "text": [''] * len(text_or_image_list),
            'image': text_or_image_list,
            'tokenizer': tokenizer
        }
    outputs = model(**inputs)
    attention_mask = outputs.attention_mask
    hidden = outputs.last_hidden_state

    reps = weighted_mean_pooling(hidden, attention_mask)   
    embeddings = F.normalize(reps, p=2, dim=1).detach().cpu().numpy()
    return embeddings

def get_image_md5(img: Image.Image):
    img_byte_array = img.tobytes()
    hash_md5 = hashlib.md5()
    hash_md5.update(img_byte_array)
    hex_digest = hash_md5.hexdigest()
    return hex_digest

def calculate_md5_from_binary(binary_data):
    hash_md5 = hashlib.md5()
    hash_md5.update(binary_data)
    return hash_md5.hexdigest()

@spaces.GPU(duration=100)
def add_pdf_gradio(pdf_file_binary, progress=gr.Progress()):
    global model, tokenizer
    model.eval()
    
    knowledge_base_name = calculate_md5_from_binary(pdf_file_binary)
    
    this_cache_dir = os.path.join(cache_dir, knowledge_base_name)
    os.makedirs(this_cache_dir, exist_ok=True)

    with open(os.path.join(this_cache_dir, f"src.pdf"), 'wb') as file:
        file.write(pdf_file_binary)

    dpi = 200
    doc = fitz.open("pdf", pdf_file_binary)
    
    reps_list = []
    images = []
    image_md5s = []

    for page in progress.tqdm(doc):
        # with self.lock: # because we hope one 16G gpu only process one image at the same time
        pix = page.get_pixmap(dpi=dpi)
        image = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
        image_md5 = get_image_md5(image)
        image_md5s.append(image_md5)
        with torch.no_grad():
            reps = encode([image])
        reps_list.append(reps)
        images.append(image)

    for idx in range(len(images)):
        image = images[idx]
        image_md5 = image_md5s[idx]
        cache_image_path = os.path.join(this_cache_dir, f"{image_md5}.png")
        image.save(cache_image_path)

    np.save(os.path.join(this_cache_dir, f"reps.npy"), reps_list)

    with open(os.path.join(this_cache_dir, f"md5s.txt"), 'w') as f:
        for item in image_md5s:
            f.write(item+'\n')
    
    return knowledge_base_name

@spaces.GPU(duration=100)
def retrieve_gradio(knowledge_base: str, query: str, topk: int):
    global model, tokenizer

    model.eval()

    target_cache_dir = os.path.join(cache_dir, knowledge_base)

    if not os.path.exists(target_cache_dir):
        return None
    
    md5s = []
    with open(os.path.join(target_cache_dir, f"md5s.txt"), 'r') as f:
        for line in f:
            md5s.append(line.rstrip('\n'))
    
    doc_reps = np.load(os.path.join(target_cache_dir, f"reps.npy"))

    query_with_instruction = "Represent this query for retrieving relavant document: " + query
    with torch.no_grad():
        query_rep = torch.Tensor(encode([query_with_instruction]))

    query_md5 = hashlib.md5(query.encode()).hexdigest()

    doc_reps_cat = torch.cat([torch.Tensor(i) for i in doc_reps], dim=0)
    print(f"doc_reps_shape: {doc_reps.shape}")
    print(f"query_rep_shape: {query_rep.shape}, doc_reps_cat_shape: {doc_reps_cat.shape}")
    similarities = torch.matmul(query_rep, doc_reps_cat.T)

    topk_values, topk_doc_ids = torch.topk(similarities, k=topk)

    topk_values_np = topk_values.cpu().numpy()

    topk_doc_ids_np = topk_doc_ids.squeeze().cpu().numpy()

    similarities_np = similarities.cpu().numpy()
    print(f"topk_doc_ids_np: {topk_doc_ids_np}, topk_values_np: {topk_values_np}")
    images_topk = [Image.open(os.path.join(target_cache_dir, f"{md5s[idx]}.png")) for idx in topk_doc_ids_np]

    with open(os.path.join(target_cache_dir, f"q-{query_md5}.json"), 'w') as f:
        f.write(json.dumps(
            {
                "knowledge_base": knowledge_base,
                "query": query,
                "retrived_docs": [os.path.join(target_cache_dir, f"{md5s[idx]}.png") for idx in topk_doc_ids_np]
            }, indent=4, ensure_ascii=False
        ))

    return images_topk


def upvote(knowledge_base, query):
    global model, tokenizer

    target_cache_dir = os.path.join(cache_dir, knowledge_base)

    query_md5 = hashlib.md5(query.encode()).hexdigest()

    with open(os.path.join(target_cache_dir, f"q-{query_md5}.json"), 'r') as f:
        data = json.loads(f.read())

    data["user_preference"] = "upvote"

    with open(os.path.join(target_cache_dir, f"q-{query_md5}-withpref.json"), 'w') as f:
        f.write(json.dumps(data, indent=4, ensure_ascii=False))

    print("up", os.path.join(target_cache_dir, f"q-{query_md5}-withpref.json"))

    gr.Info('Received, babe! Thank you!')

    return


def downvote(knowledge_base, query):
    global model, tokenizer

    target_cache_dir = os.path.join(cache_dir, knowledge_base)

    query_md5 = hashlib.md5(query.encode()).hexdigest()

    with open(os.path.join(target_cache_dir, f"q-{query_md5}.json"), 'r') as f:
        data = json.loads(f.read())

    data["user_preference"] = "downvote"

    with open(os.path.join(target_cache_dir, f"q-{query_md5}-withpref.json"), 'w') as f:
        f.write(json.dumps(data, indent=4, ensure_ascii=False))

    print("down", os.path.join(target_cache_dir, f"q-{query_md5}-withpref.json"))

    gr.Info('Received, babe! Thank you!')

    return



device = 'cuda'

print("emb model load begin...")
model_path = 'openbmb/VisRAG-Ret' # replace with your local model path
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModel.from_pretrained(model_path, trust_remote_code=True)
model.eval()
model.to(device)
print("emb model load success!")

print("gen model load begin...")
gen_model_path = 'openbmb/MiniCPM-V-2_6'
gen_tokenizer = AutoTokenizer.from_pretrained(gen_model_path, attn_implementation='sdpa', trust_remote_code=True)
gen_model = AutoModel.from_pretrained(gen_model_path, trust_remote_code=True,
    attn_implementation='sdpa', torch_dtype=torch.bfloat16)
gen_model.eval()
gen_model.to(device)
print("gen model load success!")


@spaces.GPU(duration=50)
def answer_question(images, question):
    global gen_model, gen_tokenizer
    # here each element of images is a tuple of (image_path, None).
    images_ = [Image.open(image[0]).convert('RGB') for image in images]
    msgs = [{'role': 'user', 'content': [question, *images_]}]
    answer = gen_model.chat(
        image=None,
        msgs=msgs,
        tokenizer=gen_tokenizer
    )
    print(answer)
    return answer


with gr.Blocks() as app:
    gr.Markdown("# MiniCPMV-RAG-PDFQA: Two Vision Language Models Enable End-to-End RAG")
    
    gr.Markdown("""
- A Vision Language Model Dense Retriever ([minicpm-visual-embedding-v0](https://huggingface.co/RhapsodyAI/minicpm-visual-embedding-v0)) **directly reads** your PDFs **without need of OCR**, produce **multimodal dense representations** and build your personal library. 

- **Ask a question**, it retrieve most relavant pages, then [MiniCPM-V-2.6](https://huggingface.co/spaces/openbmb/MiniCPM-V-2_6) will answer your question based on pages recalled, with strong multi-image understanding capability. 

    - It helps you read a long **visually-intensive** or **text-oriented** PDF document and find the pages that answer your question.

    - It helps you build a personal library and retireve book pages from a large collection of books.

    - It works like a human: read, store, retrieve, and answer with full vision.
""")
    
    gr.Markdown("- Currently online demo support PDF document with less than 50 pages due to GPU time limit. Deploy on your own machine for longer PDFs and books.")
    
    with gr.Row():
        file_input = gr.File(type="binary", label="Step 1: Upload PDF")
        file_result = gr.Text(label="Knowledge Base ID (remember it, it is re-usable!)")
        process_button = gr.Button("Process PDF (Don't click until PDF upload success)")
    
    process_button.click(add_pdf_gradio, inputs=[file_input], outputs=file_result)

    with gr.Row():
        kb_id_input = gr.Text(label="Your Knowledge Base ID (paste your Knowledge Base ID here, it is re-usable:)")
        query_input = gr.Text(label="Your Queston")
        topk_input = inputs=gr.Number(value=5, minimum=1, maximum=10, step=1, label="Number of pages to retrieve")
        retrieve_button = gr.Button("Step2: Retrieve Pages")
    
    with gr.Row():
        images_output = gr.Gallery(label="Retrieved Pages")
    
    retrieve_button.click(retrieve_gradio, inputs=[kb_id_input, query_input, topk_input], outputs=images_output)

    with gr.Row():
        button = gr.Button("Step 3: Answer Question with Retrieved Pages")

        gen_model_response = gr.Textbox(label="MiniCPM-V-2.6's Answer")

        button.click(fn=answer_question, inputs=[images_output, query_input], outputs=gen_model_response)
    
    with gr.Row():
        downvote_button = gr.Button("🤣Downvote")
        upvote_button = gr.Button("🤗Upvote")
        
    upvote_button.click(upvote, inputs=[kb_id_input, query_input], outputs=None)
    downvote_button.click(downvote, inputs=[kb_id_input, query_input], outputs=None)

    gr.Markdown("By using this demo, you agree to share your use data with us for research purpose, to help improve user experience.")


app.launch()