import os import io import base64 import gc from huggingface_hub.utils import HfHubHTTPError from langchain_core.prompts import PromptTemplate from langchain_huggingface import HuggingFaceEndpoint import io, base64 from PIL import Image import torch import gradio as gr import spaces import numpy as np import pandas as pd import pymupdf from PIL import Image from pypdf import PdfReader from dotenv import load_dotenv from welcome_text import WELCOME_INTRO from doctr.io import DocumentFile from doctr.models import ocr_predictor from transformers import LlavaNextProcessor, LlavaNextForConditionalGeneration import chromadb from chromadb.utils import embedding_functions from chromadb.utils.data_loaders import ImageLoader from langchain_core.prompts import PromptTemplate from langchain.text_splitter import RecursiveCharacterTextSplitter from langchain_huggingface import HuggingFaceEndpoint from utils import extract_pdfs, extract_images, clean_text, image_to_bytes from utils import * # ───────────────────────────────────────────────────────────────────────────── # Load .env load_dotenv() HF_TOKEN = os.getenv("HUGGINGFACE_TOKEN") processor = None vision_model = None # OCR + multimodal image description setup ocr_model = ocr_predictor( "db_resnet50", "crnn_mobilenet_v3_large", pretrained=True, assume_straight_pages=True ) 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 ).to("cuda") # Add at the top of your module, alongside your other globals CURRENT_VDB = None @spaces.GPU() def get_image_description(image: Image.Image) -> str: """ Lazy-loads the Llava processor + model inside the GPU worker, runs captioning, and returns a one-sentence description. """ global processor, vision_model # On first call, instantiate + move to CUDA if processor is None or vision_model is None: 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 ).to("cuda") torch.cuda.empty_cache() gc.collect() prompt = "[INST] \nDescribe the image in a sentence [/INST]" inputs = processor(prompt, image, return_tensors="pt").to("cuda") output = vision_model.generate(**inputs, max_new_tokens=100) return processor.decode(output[0], skip_special_tokens=True) # Vector DB setup # at top of file, alongside your other imports from chromadb.utils import embedding_functions from chromadb.utils.data_loaders import ImageLoader import chromadb from langchain.text_splitter import RecursiveCharacterTextSplitter from utils import image_to_bytes # your helper # 1) Create one shared embedding function (defaulting to All-MiniLM-L6-v2, 384-dim) SHARED_EMB_FN = embedding_functions.SentenceTransformerEmbeddingFunction( model_name="all-MiniLM-L6-v2" ) def get_vectordb(text: str, images: list[Image.Image], img_names: list[str]): """ Build an in-memory ChromaDB instance with two collections: • text_db (chunks of the PDF text) • image_db (image descriptions + raw image bytes) Returns the Chroma client for later querying. """ # ——— 1) Init & wipe old ———————————————— client = chromadb.EphemeralClient() for col in ("text_db", "image_db"): if col in [c.name for c in client.list_collections()]: client.delete_collection(col) # ——— 2) Create fresh collections ————————— text_col = client.get_or_create_collection( name="text_db", embedding_function=SHARED_EMB_FN, data_loader=ImageLoader(), # loader only matters for images, benign here ) img_col = client.get_or_create_collection( name="image_db", embedding_function=SHARED_EMB_FN, metadata={"hnsw:space": "cosine"}, data_loader=ImageLoader(), ) # ——— 3) Add images if any ——————————————— if images: descs = [] metas = [] for idx, img in enumerate(images): # build one-line caption (or fallback) try: caption = get_image_description(img) except Exception: caption = "⚠️ could not describe image" descs.append(f"{img_names[idx]}: {caption}") metas.append({"image": image_to_bytes(img)}) img_col.add( ids=[str(i) for i in range(len(images))], documents=descs, metadatas=metas, ) # ——— 4) Chunk & add text ——————————————— splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50) docs = splitter.create_documents([text]) text_col.add( ids=[str(i) for i in range(len(docs))], documents=[d.page_content for d in docs], ) return client # Text extraction def result_to_text(result, as_text=False): pages = [] for pg in result.pages: txt = " ".join(w.value for block in pg.blocks for line in block.lines for w in line.words) pages.append(clean_text(txt)) return "\n\n".join(pages) if as_text else pages OCR_CHOICES = { "db_resnet50 + crnn_mobilenet_v3_large": ("db_resnet50", "crnn_mobilenet_v3_large"), "db_resnet50 + crnn_resnet31": ("db_resnet50", "crnn_resnet31"), } @spaces.GPU() def extract_data_from_pdfs( docs: list[str], session: dict, include_images: str, # "Include Images" or "Exclude Images" do_ocr: str, # "Get Text With OCR" or "Get Available Text Only" ocr_choice: str, # key into OCR_CHOICES vlm_choice: str, # HF repo ID for LlavaNext progress=gr.Progress() ): """ 1) (Optional) OCR setup 2) Vision+Lang model setup & monkey-patch get_image_description 3) Extract text & images 4) Build and stash vector DB in CURRENT_VDB """ if not docs: raise gr.Error("No documents to process") # 1) OCR pipeline if requested if do_ocr == "Get Text With OCR": db_m, crnn_m = OCR_CHOICES[ocr_choice] local_ocr = ocr_predictor(db_m, crnn_m, pretrained=True, assume_straight_pages=True) else: local_ocr = None # 2) Vision–language model proc = LlavaNextProcessor.from_pretrained(vlm_choice) vis = ( LlavaNextForConditionalGeneration .from_pretrained(vlm_choice, torch_dtype=torch.float16, low_cpu_mem_usage=True) .to("cuda") ) # Monkey-patch our pipeline for image captions def describe(img: Image.Image) -> str: torch.cuda.empty_cache() gc.collect() prompt = "[INST] \nDescribe the image in a sentence [/INST]" inputs = proc(prompt, img, return_tensors="pt").to("cuda") output = vis.generate(**inputs, max_new_tokens=100) return proc.decode(output[0], skip_special_tokens=True) global get_image_description, CURRENT_VDB get_image_description = describe # 3) Extract text + images progress(0.2, "Extracting text and images…") all_text = "" images, names = [], [] for path in docs: if local_ocr: pdf = DocumentFile.from_pdf(path) res = local_ocr(pdf) all_text += result_to_text(res, as_text=True) + "\n\n" else: txt = PdfReader(path).pages[0].extract_text() or "" all_text += txt + "\n\n" if include_images == "Include Images": imgs = extract_images([path]) images.extend(imgs) names.extend([os.path.basename(path)] * len(imgs)) # 4) Build + store the vector DB progress(0.6, "Indexing in vector DB…") CURRENT_VDB = get_vectordb(all_text, images, names) session["processed"] = True sample_imgs = images[:4] if include_images == "Include Images" else [] # ─── return *exactly four* picklable outputs ─── return ( session, # gr.State: so UI knows we're ready all_text[:2000] + "...", # preview text sample_imgs, # preview images "

Done!

" # Done message ) # Chat function def conversation( session: dict, question: str, num_ctx: int, img_ctx: int, history: list, temp: float, max_tok: int, model_id: str ): """ Uses the global CURRENT_VDB (set by extract_data_from_pdfs) to answer. """ global CURRENT_VDB if not session.get("processed") or CURRENT_VDB is None: raise gr.Error("Please extract data first") llm = HuggingFaceEndpoint( repo_id=model_id, temperature=temp, max_new_tokens=max_tok, huggingfacehub_api_token=HF_TOKEN ) # 1) Text retrieval text_col = CURRENT_VDB.get_collection("text_db") docs = text_col.query( query_texts=[question], n_results=int(num_ctx), include=["documents"] )["documents"][0] # 2) Image retrieval img_col = CURRENT_VDB.get_collection("image_db") img_q = img_col.query( query_texts=[question], n_results=int(img_ctx), include=["metadatas", "documents"] ) img_descs = img_q["documents"][0] or ["No images found"] images = [] for meta in img_q["metadatas"][0]: b64 = meta.get("image", "") try: images.append(Image.open(io.BytesIO(base64.b64decode(b64)))) except: pass img_desc = "\n".join(img_descs) # 3) Build prompt & call LLM prompt = PromptTemplate( template=""" Context: {text} Included Images: {img_desc} Question: {q} Answer: """, input_variables=["text", "img_desc", "q"], ) user_input = prompt.format( text="\n\n".join(docs), img_desc=img_desc, q=question ) try: answer = llm.invoke(user_input) except HfHubHTTPError as e: if e.response.status_code == 404: answer = f"❌ Model `{model_id}` not hosted on HF Inference API." else: answer = f"⚠️ HF API error: {e}" except Exception as e: answer = f"⚠️ Unexpected error: {e}" new_history = history + [ {"role": "user", "content": question}, {"role": "assistant", "content": answer} ] return new_history, docs, images # ───────────────────────────────────────────────────────────────────────────── # Gradio UI CSS = """ footer {visibility:hidden;} """ MODEL_OPTIONS = [ "HuggingFaceH4/zephyr-7b-beta", "mistralai/Mistral-7B-Instruct-v0.2", "openchat/openchat-3.5-0106", "google/gemma-7b-it", "deepseek-ai/deepseek-llm-7b-chat", "microsoft/Phi-3-mini-4k-instruct", "TinyLlama/TinyLlama-1.1B-Chat-v1.0", "Qwen/Qwen1.5-7B-Chat", "tiiuae/falcon-7b-instruct", # Falcon 7B Instruct "bigscience/bloomz-7b1", # BLOOMZ 7B "facebook/opt-2.7b", ] with gr.Blocks(css=CSS, theme=gr.themes.Soft()) as demo: # We only need a single State to track whether extraction has happened session_state = gr.State({}) # ─── Welcome Screen ───────────────────────────────────────────── with gr.Column(visible=True) as welcome_col: gr.Markdown( f"
\n{WELCOME_INTRO}\n
", elem_id="welcome_md" ) start_btn = gr.Button("🚀 Start") # ─── Main App (hidden until Start is clicked) ─────────────────── with gr.Column(visible=False) as app_col: gr.Markdown("## 📚 Multimodal Chat-PDF Playground") with gr.Tabs(): # Tab 1: Upload & Extract with gr.TabItem("1. Upload & Extract"): docs = gr.File( file_count="multiple", file_types=[".pdf"], label="Upload PDFs" ) include_dd = gr.Radio( ["Include Images", "Exclude Images"], value="Exclude Images", label="Images" ) ocr_radio = gr.Radio( ["Get Text With OCR", "Get Available Text Only"], value="Get Available Text Only", label="OCR" ) ocr_dd = gr.Dropdown( choices=[ "db_resnet50 + crnn_mobilenet_v3_large", "db_resnet50 + crnn_resnet31" ], value="db_resnet50 + crnn_mobilenet_v3_large", label="OCR Model" ) vlm_dd = gr.Dropdown( choices=[ "llava-hf/llava-v1.6-mistral-7b-hf", "llava-hf/llava-v1.5-mistral-7b" ], value="llava-hf/llava-v1.6-mistral-7b-hf", label="Vision-Language Model" ) extract_btn = gr.Button("Extract") preview_text = gr.Textbox( lines=10, label="Sample Text", interactive=False ) preview_img = gr.Gallery( label="Sample Images", rows=2, value=[] ) preview_html = gr.HTML() extract_btn.click( fn=extract_data_from_pdfs, inputs=[ docs, session_state, include_dd, ocr_radio, ocr_dd, vlm_dd ], outputs=[ session_state, # updates processed flag preview_text, # shows sample text preview_img, # shows sample images preview_html # shows “Done!” message ] ) # Tab 2: Chat with gr.TabItem("2. Chat"): with gr.Row(): with gr.Column(scale=3): chat = gr.Chatbot(type="messages", label="Chat") msg = gr.Textbox( placeholder="Ask about your PDF...", label="Your question" ) send = gr.Button("Send") with gr.Column(scale=1): model_dd = gr.Dropdown( MODEL_OPTIONS, value=MODEL_OPTIONS[0], label="Choose Chat Model" ) num_ctx = gr.Slider(1, 20, value=3, label="Text Contexts") img_ctx = gr.Slider(1, 10, value=2, label="Image Contexts") temp = gr.Slider(0.1, 1.0, step=0.1, value=0.4, label="Temperature") max_tok = gr.Slider(10, 1000, step=10, value=200, label="Max Tokens") send.click( fn=conversation, inputs=[ session_state, # drives conversation msg, num_ctx, img_ctx, chat, temp, max_tok, model_dd ], outputs=[ chat, gr.Dataframe(), # returns the retrieved docs gr.Gallery(label="Relevant Images", rows=2, value=[]) ] ) gr.HTML("
Made with ❤️ by Zamal
") # ─── Wire the Start button ─────────────────────────────────────── start_btn.click( fn=lambda: (gr.update(visible=False), gr.update(visible=True)), inputs=[], outputs=[welcome_col, app_col] ) if __name__ == "__main__": demo.launch()