app.py

"""

app.py



Document Analysis Gradio app — updated to support PDF, DOCX (Word), and PPTX (PowerPoint).

- robust file reading

- streaming PDF extraction

- docx/pptx extraction to pages_texts (one element per paragraph/slide)

- token-aware truncation, chunked summarization, sampled Q&A

- multi-file upload UI (processes first supported file)

"""

import os
# disable noisy HF symlink warning on Windows
os.environ.setdefault("HF_HUB_DISABLE_SYMLINKS_WARNING", "1")

import re
import io
import math
import tempfile
import threading
from pathlib import Path
from typing import List, Tuple, Optional

import gradio as gr
import pdfplumber
import nltk
from nltk.tokenize import sent_tokenize
from transformers import pipeline, AutoTokenizer
import torch
import pandas as pd
from tqdm.auto import tqdm

# Try to import docx and pptx; if not present, we'll handle gracefully at runtime.
try:
    from docx import Document as DocxDocument  # python-docx
except Exception:
    DocxDocument = None

try:
    from pptx import Presentation as PptxPresentation  # python-pptx
except Exception:
    PptxPresentation = None

# -------------------------
# NLTK: ensure punkt available, fallback later
# -------------------------
try:
    nltk.download("punkt", quiet=True)
    try:
        nltk.download("punkt_tab", quiet=True)
    except Exception:
        pass
except Exception:
    pass

# -------------------------
# Device detection
# -------------------------
DEVICE = 0 if torch.cuda.is_available() else -1
print("Device set to use", "cuda" if DEVICE >= 0 else "cpu")

# -------------------------
# Cached pipelines and tokenizers
# -------------------------
_models = {}
_tokenizers = {}
_models_lock = threading.Lock()


def get_pipeline(name: str, task: str):
    """Return a cached HF pipeline for given task and model name."""
    key = f"{task}__{name}"
    with _models_lock:
        if key in _models:
            return _models[key]
        print(f"Loading pipeline: task={task}, model={name} ... (this may take a while on first run)")
        p = pipeline(task, model=name, device=DEVICE)
        _models[key] = p
        try:
            _tokenizers[name] = p.tokenizer
        except Exception:
            pass
        return p


def get_tokenizer(name: str):
    """Return a cached tokenizer (fallback to AutoTokenizer if not present)."""
    if name in _tokenizers:
        return _tokenizers[name]
    try:
        tok = AutoTokenizer.from_pretrained(name)
        _tokenizers[name] = tok
        return tok
    except Exception:
        return None


# -------------------------
# Default models (adjust if you want smaller/faster ones)
# -------------------------
# -------------------------
# Default models (adjusted for speed)
# -------------------------
SUMMARIZER_MODEL = "t5-small"  # This is already fast
QG_MODEL = "valhalla/t5-small-qg-hl" # Use the 'small' version
QA_MODEL = "distilbert-base-cased-distilled-squad" # Much faster than RoBERTa


# -------------------------
# Helpers: filenames / types
# -------------------------
SUPPORTED_EXTS = [".pdf", ".docx", ".pptx", ".txt", ".md", ".rtf", ".png", ".jpg", ".jpeg", ".tiff"]

def ext_of_name(name: str) -> str:
    return Path(name).suffix.lower()


def read_uploaded_file_to_bytes(file_obj):
    """

    Accept many shapes of Gradio file objects and return bytes.

    Supports list/tuple (returns first readable file), dict-like, file-likes, paths, bytes.

    """
    # If a list/tuple of uploaded files, try each candidate
    if isinstance(file_obj, (list, tuple)):
        last_err = None
        for elem in file_obj:
            try:
                return read_uploaded_file_to_bytes(elem)
            except Exception as e:
                last_err = e
                continue
        raise ValueError(f"No readable file in list. Last error: {last_err}")

    if file_obj is None:
        raise ValueError("No file provided")

    # if it's already bytes
    if isinstance(file_obj, (bytes, bytearray)):
        return bytes(file_obj)

    # dict-like
    if isinstance(file_obj, dict):
        for key in ("file", "tmp_path", "name", "data", "path"):
            val = file_obj.get(key)
            if isinstance(val, (bytes, bytearray)):
                return bytes(val)
            if isinstance(val, str) and Path(val).exists():
                return Path(val).read_bytes()
        maybe = file_obj.get("file")
        if hasattr(maybe, "read"):
            data = maybe.read()
            if isinstance(data, str):
                return data.encode("utf-8")
            return data

    # string path
    if isinstance(file_obj, str) and Path(file_obj).exists():
        return Path(file_obj).read_bytes()

    # has a .name attribute that points to a file
    if hasattr(file_obj, "name") and isinstance(getattr(file_obj, "name"), str) and Path(file_obj.name).exists():
        try:
            return Path(file_obj.name).read_bytes()
        except Exception:
            pass

    # file-like with .read()
    if hasattr(file_obj, "read"):
        try:
            data = file_obj.read()
            if isinstance(data, str):
                return data.encode("utf-8")
            return data
        except Exception:
            pass

    # last resort: string representation -> path
    try:
        s = str(file_obj)
        if Path(s).exists():
            return Path(s).read_bytes()
    except Exception:
        pass

    raise ValueError(f"Unsupported uploaded file object type: {type(file_obj)}")


def get_uploaded_filenames(file_obj) -> List[str]:
    """

    Return a list of human-friendly filenames from uploaded file object(s).

    """
    names = []
    if file_obj is None:
        return names
    if isinstance(file_obj, (list, tuple)):
        for elem in file_obj:
            names.extend(get_uploaded_filenames(elem))
        return names
    if isinstance(file_obj, dict):
        for key in ("name", "filename", "file", "tmp_path", "path"):
            if key in file_obj:
                val = file_obj.get(key)
                if isinstance(val, str):
                    names.append(Path(val).name)
                elif hasattr(val, "name"):
                    names.append(Path(val.name).name)
        maybe = file_obj.get("file")
        if maybe is not None:
            if hasattr(maybe, "name"):
                names.append(Path(maybe.name).name)
        return names
    if isinstance(file_obj, str):
        return [Path(file_obj).name]
    if hasattr(file_obj, "name"):
        return [Path(getattr(file_obj, "name")).name]
    return [str(file_obj)]


def find_first_supported_file(files) -> Tuple[Optional[object], Optional[str]]:
    """

    From the uploaded list or single file-like object, find the first file with a supported extension.

    Returns (file_obj, filename) or (None, None) if none supported.

    """
    if not files:
        return None, None
    candidates = []
    if isinstance(files, (list, tuple)):
        for f in files:
            names = get_uploaded_filenames(f)
            for n in names:
                candidates.append((f, n))
    else:
        names = get_uploaded_filenames(files)
        for n in names:
            candidates.append((files, n))
    for fobj, name in candidates:
        ext = ext_of_name(name)
        if ext in SUPPORTED_EXTS:
            return fobj, name
    # fallback: if none matched, return first uploaded
    if candidates:
        return candidates[0]
    return None, None


# -------------------------
# PDF extraction (streaming)
# -------------------------
def extract_text_from_pdf_streaming(file_bytes: bytes, do_ocr: bool = False, extracted_txt_path: str = None):
    """Write PDF bytes to temp file and extract each page's text; returns extracted_txt_path, pages_texts list"""
    with tempfile.NamedTemporaryFile(delete=False, suffix=".pdf") as tmp_pdf:
        tmp_pdf.write(file_bytes)
        tmp_pdf_path = tmp_pdf.name

    if extracted_txt_path is None:
        tmp_txt = tempfile.NamedTemporaryFile(delete=False, suffix=".txt", mode="w+", encoding="utf-8")
        extracted_txt_path = tmp_txt.name
        tmp_txt.close()

    pages_texts = []
    try:
        with pdfplumber.open(tmp_pdf_path) as pdf:
            for page in pdf.pages:
                page_text = page.extract_text()
                if (not page_text or page_text.strip() == "") and do_ocr:
                    try:
                        from pdf2image import convert_from_path
                        import pytesseract
                        images = convert_from_path(tmp_pdf_path, first_page=page.page_number, last_page=page.page_number)
                        if images:
                            page_text = pytesseract.image_to_string(images[0])
                    except Exception:
                        page_text = ""
                if page_text is None:
                    page_text = ""
                with open(extracted_txt_path, "a", encoding="utf-8") as fout:
                    fout.write(page_text)
                    fout.write("\n\n---PAGE_BREAK---\n\n")
                pages_texts.append(page_text)
    finally:
        try:
            os.remove(tmp_pdf_path)
        except Exception:
            pass

    return extracted_txt_path, pages_texts


# -------------------------
# DOCX extraction
# -------------------------
def extract_text_from_docx_bytes(file_bytes: bytes) -> Tuple[str, List[str]]:
    """

    Returns (tmp_text_path, pages_texts) where pages_texts is a list of paragraph groups.

    We'll treat each paragraph as a small 'page' or group paragraphs into ~200-word chunks.

    """
    if DocxDocument is None:
        raise RuntimeError("python-docx not installed. Install with `pip install python-docx`")

    with tempfile.NamedTemporaryFile(delete=False, suffix=".docx") as tmp:
        tmp.write(file_bytes)
        tmp_path = tmp.name

    pages_texts = []
    try:
        doc = DocxDocument(tmp_path)
        # Collect non-empty paragraphs
        paras = [p.text.strip() for p in doc.paragraphs if p.text and p.text.strip()]
        # Group paragraphs into passages of ~200 words
        current = []
        cur_words = 0
        for p in paras:
            w = len(p.split())
            if cur_words + w <= 200:
                current.append(p)
                cur_words += w
            else:
                pages_texts.append(" ".join(current).strip())
                current = [p]
                cur_words = w
        if current:
            pages_texts.append(" ".join(current).strip())
        # write full extracted text to tmp file for download/debug if needed
        tmp_txt = tempfile.NamedTemporaryFile(delete=False, suffix=".txt", mode="w+", encoding="utf-8")
        with open(tmp_txt.name, "w", encoding="utf-8") as f:
            f.write("\n\n".join(pages_texts))
        txt_path = tmp_txt.name
    finally:
        try:
            os.remove(tmp_path)
        except Exception:
            pass

    return txt_path, pages_texts


# -------------------------
# PPTX extraction
# -------------------------
def extract_text_from_pptx_bytes(file_bytes: bytes) -> Tuple[str, List[str]]:
    """

    Returns (tmp_text_path, pages_texts) where each slide's text is one element.

    """
    if PptxPresentation is None:
        raise RuntimeError("python-pptx not installed. Install with `pip install python-pptx`")

    with tempfile.NamedTemporaryFile(delete=False, suffix=".pptx") as tmp:
        tmp.write(file_bytes)
        tmp_path = tmp.name

    pages_texts = []
    try:
        prs = PptxPresentation(tmp_path)
        for slide in prs.slides:
            texts = []
            for shape in slide.shapes:
                try:
                    if hasattr(shape, "text") and shape.text:
                        texts.append(shape.text.strip())
                except Exception:
                    continue
            slide_text = "\n".join([t for t in texts if t])
            pages_texts.append(slide_text)
        tmp_txt = tempfile.NamedTemporaryFile(delete=False, suffix=".txt", mode="w+", encoding="utf-8")
        with open(tmp_txt.name, "w", encoding="utf-8") as f:
            f.write("\n\n".join(pages_texts))
        txt_path = tmp_txt.name
    finally:
        try:
            os.remove(tmp_path)
        except Exception:
            pass

    return txt_path, pages_texts


# -------------------------
# TXT/MD extraction
# -------------------------
def extract_text_from_txt_bytes(file_bytes: bytes) -> Tuple[str, List[str]]:
    s = file_bytes.decode("utf-8", errors="ignore")
    # split into passages by blank lines or ~200 words
    paras = [p.strip() for p in re.split(r"\n\s*\n", s) if p.strip()]
    pages_texts = []
    cur = []
    cur_words = 0
    for p in paras:
        w = len(p.split())
        if cur_words + w <= 200:
            cur.append(p)
            cur_words += w
        else:
            pages_texts.append(" ".join(cur).strip())
            cur = [p]
            cur_words = w
    if cur:
        pages_texts.append(" ".join(cur).strip())
    tmp_txt = tempfile.NamedTemporaryFile(delete=False, suffix=".txt", mode="w+", encoding="utf-8")
    with open(tmp_txt.name, "w", encoding="utf-8") as f:
        f.write("\n\n".join(pages_texts))
    return tmp_txt.name, pages_texts


# -------------------------
# Token-aware truncation helpers
# -------------------------
def truncate_by_tokens(text: str, tokenizer, reserve: int = 64) -> str:
    if not text:
        return text
    if tokenizer is None:
        return text if len(text) <= 3000 else text[:3000]
    try:
        ids = tokenizer.encode(text, add_special_tokens=False)
        max_len = getattr(tokenizer, "model_max_length", 512)
        allowed = max(1, max_len - reserve)
        if len(ids) > allowed:
            ids = ids[:allowed]
            return tokenizer.decode(ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
        return text
    except Exception:
        return text if len(text) <= 3000 else text[:3000]


# -------------------------
# Sentence tokenization fallback
# -------------------------
def safe_sentence_tokenize(text: str) -> List[str]:
    try:
        sents = sent_tokenize(text)
        if isinstance(sents, list) and len(sents) > 0:
            return sents
    except Exception:
        pass
    pieces = re.split(r"(?<=[.!?])\s+", text.strip())
    return [p.strip() for p in pieces if p.strip()]


# -------------------------
# Summarization (chunked, token-truncated)
# -------------------------
def summarize_text_chunked(summarizer, pages_texts: List[str], pages_per_chunk: int = 8) -> str:
    if not pages_texts:
        return "(no text)"
    summaries = []
    tokenizer = getattr(summarizer, "tokenizer", None) or get_tokenizer(SUMMARIZER_MODEL)
    num_pages = len(pages_texts)
    for i in range(0, num_pages, pages_per_chunk):
        chunk_pages = pages_texts[i : i + pages_per_chunk]
        chunk_text = "\n\n".join([p for p in chunk_pages if p.strip()])
        if not chunk_text.strip():
            continue
        safe_chunk = truncate_by_tokens(chunk_text, tokenizer, reserve=64)
        try:
            out = summarizer(safe_chunk, max_length=150, min_length=30, do_sample=False, truncation=True)
            summaries.append(out[0]["summary_text"])
        except Exception:
            summaries.append(safe_chunk[:800])
    return "\n\n".join(summaries) if summaries else "(no summary produced)"


# -------------------------
# Passage splitting / QG / QA with token truncation
# -------------------------
def split_into_passages_from_pages(pages_texts: List[str], max_words: int = 200) -> List[str]:
    all_passages = []
    for page_text in pages_texts:
        if not page_text or not page_text.strip():
            continue
        sents = safe_sentence_tokenize(page_text)
        cur = []
        cur_len = 0
        for s in sents:
            w = len(s.split())
            if cur_len + w <= max_words:
                cur.append(s)
                cur_len += w
            else:
                if cur:
                    all_passages.append(" ".join(cur).strip())
                cur = [s]
                cur_len = w
        if cur:
            all_passages.append(" ".join(cur).strip())
    return all_passages


def generate_questions_from_passage(qg_pipeline, passage: str, min_questions: int = 3) -> List[str]:
    tok = getattr(qg_pipeline, "tokenizer", None) or get_tokenizer(QG_MODEL)
    safe_passage = truncate_by_tokens(passage, tok, reserve=32)
    prompt = f"generate questions: {safe_passage}"
    try:
        out = qg_pipeline(prompt, max_length=256, do_sample=False, truncation=True)
        gen_text = out[0].get("generated_text") or out[0].get("text") or ""
    except Exception:
        gen_text = ""
    candidates = []
    if "<sep>" in gen_text:
        candidates = gen_text.split("<sep>")
    elif "\n" in gen_text:
        candidates = [line.strip() for line in gen_text.splitlines() if line.strip()]
    else:
        parts = [p.strip() for p in gen_text.split("?") if p.strip()]
        candidates = [p + "?" for p in parts]
    questions = [q.strip() for q in candidates if q.strip()]
    if len(questions) < min_questions:
        sentences = safe_sentence_tokenize(safe_passage)
        for i in range(len(sentences)):
            if len(questions) >= min_questions:
                break
            small = ". ".join(sentences[i : i + 2])
            try:
                out2 = qg_pipeline(f"generate questions: {small}", max_length=128, do_sample=False, truncation=True)
                txt = out2[0].get("generated_text") or out2[0].get("text") or ""
            except Exception:
                txt = ""
            if "<sep>" in txt:
                more = txt.split("<sep>")
            else:
                more = [l.strip() for l in txt.splitlines() if l.strip()]
            for m in more:
                if len(questions) >= min_questions:
                    break
                maybe = m.strip()
                if maybe and maybe not in questions:
                    questions.append(maybe)
    return questions[:max(min_questions, len(questions))]


def answer_questions_for_passage(qa_pipeline, passage: str, questions: List[str]) -> List[Tuple[str, str, float]]:
    results = []
    tok = getattr(qa_pipeline, "tokenizer", None) or get_tokenizer(QA_MODEL)
    for q in questions:
        try:
            safe_ctx = truncate_by_tokens(passage, tok, reserve=64)
            res = qa_pipeline(question=q, context=safe_ctx)
            answer = res.get("answer", "")
            score = float(res.get("score", 0.0))
        except Exception:
            answer = ""
            score = 0.0
        results.append((q, answer, score))
    return results


# -------------------------
# Unified extract_text_for_file: dispatch by extension
# -------------------------
def extract_text_for_file(file_bytes: bytes, filename: str, do_ocr: bool = False) -> Tuple[str, List[str]]:
    """

    Given raw bytes and a filename, return (extracted_txt_path, pages_texts).

    Supports PDF (streamed), DOCX, PPTX, TXT/MD. For images, attempt OCR if do_ocr True.

    """
    ext = ext_of_name(filename)
    if ext == ".pdf":
        return extract_text_from_pdf_streaming(file_bytes, do_ocr=do_ocr)
    if ext == ".docx":
        return extract_text_from_docx_bytes(file_bytes)
    if ext == ".pptx":
        return extract_text_from_pptx_bytes(file_bytes)
    if ext in (".txt", ".md", ".rtf"):
        return extract_text_from_txt_bytes(file_bytes)
    # images: try OCR if requested
    if ext in (".png", ".jpg", ".jpeg", ".tiff", ".bmp") and do_ocr:
        # write image bytes to temp file and run OCR with pytesseract if available
        try:
            from PIL import Image
            import pytesseract
            with tempfile.NamedTemporaryFile(delete=False, suffix=ext) as tmp:
                tmp.write(file_bytes)
                tmp_path = tmp.name
            img = Image.open(tmp_path)
            ocr_txt = pytesseract.image_to_string(img)
            pages = [p.strip() for p in re.split(r"\n\s*\n", ocr_txt) if p.strip()]
            tmp_txt = tempfile.NamedTemporaryFile(delete=False, suffix=".txt", mode="w+", encoding="utf-8")
            with open(tmp_txt.name, "w", encoding="utf-8") as f:
                f.write("\n\n".join(pages))
            try:
                os.remove(tmp_path)
            except Exception:
                pass
            return tmp_txt.name, pages
        except Exception:
            # fallback: treat as empty
            tmp_txt = tempfile.NamedTemporaryFile(delete=False, suffix=".txt", mode="w+", encoding="utf-8")
            with open(tmp_txt.name, "w", encoding="utf-8") as f:
                f.write("")
            return tmp_txt.name, [""]
    # unsupported extension: write bytes to txt and return raw decode
    tmp_txt = tempfile.NamedTemporaryFile(delete=False, suffix=".txt", mode="w+", encoding="utf-8")
    try:
        s = file_bytes.decode("utf-8", errors="ignore")
    except Exception:
        s = ""
    with open(tmp_txt.name, "w", encoding="utf-8") as f:
        f.write(s)
    pages = [p.strip() for p in re.split(r"\n\s*\n", s) if p.strip()]
    return tmp_txt.name, pages if pages else [s]


# -------------------------
# Main analyze function
# -------------------------
def analyze_document(file_obj, filename: str, do_ocr: bool, max_passages_for_qa: int = 5):
    """

    file_obj: uploaded object

    filename: filename string (to detect extension)

    """
    try:
        file_bytes = read_uploaded_file_to_bytes(file_obj)
    except Exception as e:
        return f"(error reading file: {e})", "", []

    extracted_txt_path, pages_texts = extract_text_for_file(file_bytes, filename, do_ocr=do_ocr)

    # preview
    preview_chars = 20000
    extracted_preview = ""
    try:
        with open(extracted_txt_path, "r", encoding="utf-8", errors="ignore") as f:
            extracted_preview = f.read(preview_chars)
            if len(extracted_preview) >= preview_chars:
                extracted_preview += "\n\n... (preview truncated) ..."
    except Exception:
        extracted_preview = "(could not read extracted text preview)"

    summarizer = get_pipeline(SUMMARIZER_MODEL, "summarization")
    combined_summary = summarize_text_chunked(summarizer, pages_texts, pages_per_chunk=8)

    all_passages = split_into_passages_from_pages(pages_texts, max_words=200)
    total = len(all_passages)
    if total == 0:
        return extracted_preview, combined_summary, []

    if total <= max_passages_for_qa:
        chosen_passages = list(enumerate(all_passages))
    else:
        step = max(1, math.floor(total / max_passages_for_qa))
        chosen_passages = [(i, all_passages[i]) for i in range(0, total, step)][:max_passages_for_qa]

    qg = get_pipeline(QG_MODEL, "text2text-generation")
    qa = get_pipeline(QA_MODEL, "question-answering")

    answered = []
    answered_set = set()
    for (p_idx, passage) in chosen_passages:
        if not passage.strip():
            continue
        questions = generate_questions_from_passage(qg, passage, min_questions=3)
        unique_questions = [q for q in questions if q not in answered_set]
        if not unique_questions:
            continue
        answers = answer_questions_for_passage(qa, passage, unique_questions)
        for q, a, score in answers:
            answered.append({"passage_idx": int(p_idx), "question": q, "answer": a, "score": float(score)})
            answered_set.add(q)

    return extracted_preview, combined_summary, answered


# -------------------------
# Gradio UI
# -------------------------
def build_demo():
    with gr.Blocks(title="Document Analysis (LLMs)") as demo:
        gr.Markdown("# Document Analysis using LLMs\nUpload a supported file (PDF, DOCX, PPTX, TXT) and get summary + Q&A.")
        with gr.Row():
            with gr.Column(scale=1):
                files_in = gr.File(label="Upload files (PDF, DOCX, PPTX, TXT, images...)", file_count="multiple")
                do_ocr = gr.Checkbox(label="Try OCR for images/PDF pages (requires OCR libs)", value=False)
                max_pass = gr.Slider(label="Max passages to run Q&A on (lower = faster)", minimum=1, maximum=20, step=1, value=5)
                run_btn = gr.Button("Analyze Document")

            with gr.Column(scale=2):
                tabs = gr.Tabs()
                with tabs:
                    with gr.TabItem("Uploaded files"):
                        uploaded_list = gr.Textbox(label="Uploaded filenames", lines=4)
                    with gr.TabItem("Extracted Text"):
                        extracted_out = gr.Textbox(label="Extracted text (preview)", lines=15)
                    with gr.TabItem("Summary"):
                        summary_out = gr.Textbox(label="Summary", lines=8)
                    with gr.TabItem("Q&A"):
                        qa_out = gr.Dataframe(headers=["passage_idx", "question", "answer", "score"],
                                              datatype=["number", "text", "text", "number"])

        def _run(files, do_ocr_val, max_pass_val):
            names = get_uploaded_filenames(files)
            uploaded_str = "\n".join(names) if names else "(no files uploaded)"

            fobj, fname = find_first_supported_file(files)
            if fobj is None or fname is None:
                return uploaded_str, "(no supported file found)", "", pd.DataFrame(columns=["passage_idx", "question", "answer", "score"])

            text, summary, qa = analyze_document(fobj, fname, do_ocr=do_ocr_val, max_passages_for_qa=int(max_pass_val))
            if not qa:
                qa_df = pd.DataFrame(columns=["passage_idx", "question", "answer", "score"])
            else:
                qa_df = pd.DataFrame(qa)
                qa_df = qa_df.loc[:, ["passage_idx", "question", "answer", "score"]]
                qa_df["passage_idx"] = qa_df["passage_idx"].astype(int)
                qa_df["question"] = qa_df["question"].astype(str)
                qa_df["answer"] = qa_df["answer"].astype(str)
                qa_df["score"] = qa_df["score"].astype(float)

            return uploaded_str, text or "(no text extracted)", summary or "(no summary)", qa_df

        run_btn.click(_run, inputs=[files_in, do_ocr, max_pass], outputs=[uploaded_list, extracted_out, summary_out, qa_out])

    return demo


if __name__ == "__main__":
    demo = build_demo()
    demo.launch()
    # demo.launch(server_name="0.0.0.0")