pages_helpers.py

import trafilatura
import requests
import lzma
import os
import re
import time
from datetime import datetime
import json
from pprint import pprint
import subprocess

import config

from utils import *

from text_utils import *

from llm import *

from mode_llm import llm_html_to_md, md_to_text, get_html_body_with_soup

from crawl4ai import WebCrawler # pip install "crawl4ai @ git+https://github.com/unclecode/crawl4ai.git"

# Create an instance of WebCrawler
crawler = WebCrawler()

# Warm up the crawler (load necessary models)
crawler.warmup()

## Cách lấy cookies và headers sử dụng https://curlconverter.com
cookies = {
    'ASP.NET_SessionId': '42i3ivvgk14yd2tnxmddybvq',
    'Culture': 'vi',
    'Cookie_VB': 'close',
    'ruirophaply-covi19': '24',
    'SLG_G_WPT_TO': 'vi',
    'G_ENABLED_IDPS': 'google',
    'SLG_GWPT_Show_Hide_tmp': '1',
    'SLG_wptGlobTipTmp': '1',
    '__zlcmid': '1NOmxyopHgawxjN',
    '45C5EF': '96780c17-dee3-49b2-9bf7-6335c4348d4f',
    'vqc': '0',
}

headers = {
    'accept': 'text/html,application/xhtml+xml,application/xml;q=0.9,image/avif,image/webp,image/apng,*/*;q=0.8,application/signed-exchange;v=b3;q=0.7',
    'accept-language': 'en-US,en;q=0.9',
    'cache-control': 'max-age=0',
    'priority': 'u=0, i',
    'sec-ch-ua': '"Opera GX";v="111", "Chromium";v="125", "Not.A/Brand";v="24"',
    'sec-ch-ua-mobile': '?0',
    'sec-ch-ua-platform': '"Windows"',
    'sec-fetch-dest': 'document',
    'sec-fetch-mode': 'navigate',
    'sec-fetch-site': 'none',
    'sec-fetch-user': '?1',
    'sec-gpc': '1',
    'upgrade-insecure-requests': '1',
    'user-agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/125.0.0.0 Safari/537.36 OPR/111.0.0.0',
    # 'User-Agent': "Mozilla/5.0 (Linux; Android 6.0; Nexus 5 Build/MRA58N) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/113.0.0.0 Mobile Safari/537.36",
}


def norm_url_and_gen_filename(url):
    url = url.strip() # loại bỏ ký tự trống ở đầu và cuối
    if url[-1] == "/":  url = url[:-1] # loại bỏ "/" ở cuối

    # init filename và tạo sub folder nếu chưa có
    filename = f'.cache/{url}'
    return url, filename


def reset_content(url):
    _, filename = norm_url_and_gen_filename(url)

    text_filename = filename + ".txt"
    json_filename = filename + ".json"
    html_filename = filename + ".html"
    xz_filename   = filename + ".xz"

    for filename in [
        text_filename,
        json_filename,
        html_filename,
        xz_filename,
    ]:
        cmd = f"rm -rf '{filename}'"; print(cmd)
        subprocess.run(cmd, shell = True)


# from functools import lru_cache
# @lru_cache(maxsize=128)
def url_content(url, update_text=None):
    url, filename = norm_url_and_gen_filename(url)
    parts = filename.split("/")

    for i in range(1, len(parts)):
        path = "/".join(parts[:i])
        # print(path) # DEBUG
        os.makedirs(path, exist_ok=True)

    
    text_filename = filename + ".txt"
    json_filename = filename + ".json"
    html_filename = filename + ".html"
    xz_filename   = filename + ".xz"


    # Nếu được yêu cầu update nội dung text của url thì update
    if update_text is not None:

        print("\nUPDATE TEXT", filename)

        text, noise_ratio, max_noise = normalize_text(update_text, get_noise_info = True)

        # update meta, Lưu lại text và meta
        meta = json.load(open(json_filename))

        meta["updated_at"] = str(datetime.now())
        meta["updates_count"] += 1

        meta["noise_ratio"] = noise_ratio
        meta["max_noise"] = max_noise

        # Cắt ngắn text nếu đầu vào quá dài
        text = cut_if_too_long(text, meta)

        with open(json_filename, "wt") as f:
            f.write(json.dumps(meta, ensure_ascii = False))
    
        with open(text_filename, "wt") as f:
            f.write(text)

        # gọi gen_clear_view.py (chạy chậm) ở process riêng
        get_clear_view(filename)

        # Re-gen llm contents
        get_llm_gen_contents(url, use_cache = False)

        print(CYAN,end=""); print(meta); print(RESET,end="", flush=True)


    # Nếu tồn tại text_filename thì trả về,
    # đây có thể là nội dung text đã được chỉnh sửa nên ưu tiên trả về trước
    if os.path.exists(text_filename):
        print("\nGOT TEXT", filename, flush=True)

        norm_text = open(text_filename, "rt").read()
        return norm_text


    html = None
    # Thử đọc nội dung html đã được cache
    if os.path.exists(xz_filename):
        try:
            html = lzma.open(xz_filename,"rt").read()
            print("\nGOT HTML", filename, flush=True)
        except:
            pass



    blacklist = """

Your connection is not private

-----BEGIN CERTIFICATE-----

""".strip().split("\n")


    ## Thử các cách lấy HTML: requests vs crawl4ai vs reader
    get_html_method = "requests" 
    if html is None:
        # Thử lần 1 bằng requests
        print("\nGET HTML", filename, flush=True)

        try:
            html = requests.get(url, cookies=cookies, headers=headers)
            html = str(html.text)

            # Lưu lại nội dung html vào xz_filename
            with lzma.open(xz_filename, "wt") as f: f.write(html)

        except Exception as e:
            print(f"!!! REQUESTS Error {e} !!!")

        if isinstance(html, str):
            for x in blacklist:
                if x in html:
                    print(f"--{x}--")
                    i = html.find(x)
                    print(f"{RED}!!! REQUESTS đọc lỗi {html[i-30:i+200]} !!!{RESET}")
                    html = None
                    break


    meta = None
    if html is None or len(html) < 500:
        # Thử lần 2 bằng CRAWL4AI
        print("GET HTML CRAWL4AI", filename, flush=True)
        get_html_method = "crawl4ai" 

        try:
            result = crawler.run(url=url)
            html = result.html

            # Lưu lại nội dung html vào xz_filename
            with lzma.open(xz_filename, "wt") as f: f.write(html)

            # {'title': 'Ngập úng và thiệt hại trên 202.000 ha lúa | baotintuc.vn', 'description': 'Thống kê từ Bộ Nông nghiệp và Phát triển nông thôn, tính đến sáng 13/9, có trên 202.000 ha lúa, gần 39.300 ha hoa màu bị ngập úng, thiệt hại do ảnh hưởng của bão số 3.', 'keywords': None, 'author': 'baotintuc.vn', 'og:type': 'article', 'og:url': 'https://baotintuc.vn/xa-hoi/ngap-ung-va-thiet-hai-tren-202000-ha-lua-20240913095621343.htm', 'og:image': 'https://cdnmedia.baotintuc.vn/Upload/EqV5H9rWgvy9oNikwkHLXA/files/13092024-bao-1.jpg', 'og:image:url': 'https://cdnmedia.baotintuc.vn/Upload/EqV5H9rWgvy9oNikwkHLXA/files/13092024-bao-1.jpg', 'og:image:secure_url': 'https://cdnmedia.baotintuc.vn/Upload/EqV5H9rWgvy9oNikwkHLXA/files/13092024-bao-1.jpg', 'og:image:width': '460', 'og:image:height': '345', 'og:title': 'Ngập úng và thiệt hại trên 202.000 ha lúa', 'og:description': 'Thống kê từ Bộ Nông nghiệp và Phát triển nông thôn, đến sáng 13/9, có trên 202.000 ha lúa, gần 39.300 ha hoa màu bị ngập úng, thiệt hại do ảnh hưởng của bão số 3.', 'twitter: card': 'summary_large_image', 'twitter: image': 'https://cdnmedia.baotintuc.vn/Upload/EqV5H9rWgvy9oNikwkHLXA/files/13092024-bao-1.jpg'}
            meta = dict(result.metadata)

            for key in result.metadata.keys():
                if "og:" in key or "twitter:" in key:
                    meta.pop(key)

        except Exception as e:
            print(f"!!! CRAWL4AI Error {e} !!!")

        if isinstance(html, str):
            for x in blacklist:
                 if x in html:
                    i = html.find(x)
                    print(f"{RED}!!! CRAWL4AI đọc lỗi {html[i-30:i+200]} !!!{RESET}")
                    html = None
                    meta = {}
                    break


    if html is None or len(html) < 500:
        # Thử lần 3 bằng reader api
        print("GET HTML READER", filename, flush=True)
        get_html_method = "reader" 

        try:
            reader_url = "https://r.jina.ai/" + url
            # Use below header make https://jina.ai/reader return text not markdown by default
            html = requests.get(reader_url, headers = { 'X-Return-Format': 'html', }).text

            # Lưu lại nội dung html vào xz_filename
            with lzma.open(xz_filename, "wt") as f: f.write(html)

        except Exception as e:
            print(f"!!! READER Error {e} !!!")


        if isinstance(html, str):
            for x in blacklist:
                 if x in html:
                    i = html.find(x)
                    print(f"{RED}!!! READER đọc lỗi {html[i-30:i+200]} !!!{RESET}")
                    html = None
                    break


    ## Thử các cách extract text: trafilatura vs llm vs reader
    extract_method = "trafilatura" 
    # https://trafilatura.readthedocs.io/en/latest/corefunctions.html#extract
    try: 
        text = trafilatura.extract(html, 
            # favor_recall = True,
            include_tables = True,
            include_comments = False,
            with_metadata = False,
        )
    except:
        text = ""

    if meta is None: # Có thể meta đã đc lấy ở crawl4ai
        try:
            meta = trafilatura.extract(html, only_with_metadata = True)
            if meta and len(meta) > 0:
                # print(meta); input() # DEBUG
                meta = meta.split("---")[1]
                splits = re.findall(r'\S+: [^\n]+', meta)
                meta = { x.split(": ", 1)[0].strip() : x.split(": ", 1)[1].strip() for x in splits }
            else:
                meta = {}
        except:
            meta = {}


    # Chuẩn hóa text
    if text is None: text = ""
    text, noise_ratio, max_noise = normalize_text(text, get_noise_info = True)
    print(f">>> {RED}noise_ratio {pretty_num(noise_ratio)}, max_noise {max_noise}{RESET}")


    MEANINGFUL = 500
    MAX_NOISE_RATIO = 0.3

    too_short = ( len(text)  < MEANINGFUL )
    too_noise = ( noise_ratio > MAX_NOISE_RATIO or max_noise > MEANINGFUL )

    # ko lấy đc text hoặc text quá ngắn (cào trượt), hoặc text quá noise
    if text is None or too_short or too_noise: 
        # Lấy text thông qua phương pháp khác
        print("!!! Đoạn text dưới do trafilatura triết xuất có vấn đề?")
        print("too short", too_short)
        print("too noise", too_noise)
        print("- - - "*6)
        print(f"{YELLOW}{text}{RESET}")
        print("- - - "*6)

        print("!!! Dùng Jina Reader ...")
        reader_url = "https://r.jina.ai/" + url
        # Use below header make https://jina.ai/reader return text not markdown by default
        reader_text = requests.get(reader_url, headers = { 'X-Return-Format': 'text', }).text
        reader_text, reader_noise_ratio, reader_max_noise = normalize_text(reader_text, get_noise_info = True)

       # Chuẩn hóa text
        reader_text, reader_noise_ratio, reader_max_noise = normalize_text(reader_text, get_noise_info = True)

        reader_too_noise = ( reader_noise_ratio > MAX_NOISE_RATIO or reader_max_noise > MEANINGFUL )

        print(f">>> {RED}reader_noise_ratio {pretty_num(reader_noise_ratio)}, reader_max_noise {reader_max_noise}{RESET}")
        print(f">>> {RED}reader_too_noise {reader_too_noise}{RESET}")

        signal = int( len(text) * (1 - noise_ratio) ) + 1
        reader_signal = int( len(reader_text) * (1 - reader_noise_ratio) ) + 1

        samesame = ( abs(signal - reader_signal) / reader_signal ) < 0.2
        print(f">>> {RED}samesame {samesame}, original signal {pretty_num(signal)}, reader_signal {pretty_num(reader_signal)}{RESET}")

        # Nếu bản gốc quá ngắn nhưng bản reader quá noise thì thà chọn ngắn còn hơn
        original_too_shot_but_reader_too_noise = ( 
            too_short and (samesame or reader_noise_ratio >= 0.5 )
        )

        original_too_noise_but_reader_even_more_noise = (
            too_noise and noise_ratio < reader_noise_ratio and max_noise < reader_max_noise
        )

        if original_too_shot_but_reader_too_noise:
            print("!!! reader quá noise, chọn bản trafilatura too_short còn hơn.")

        if original_too_noise_but_reader_even_more_noise:
            print("!!! reader còn noise hơn bản trafilatura, bỏ qua.")


        if not original_too_shot_but_reader_too_noise and \
            not original_too_noise_but_reader_even_more_noise:

            choose_original_text = False

            if reader_too_noise: # vẫn còn noisy lắm, thử dùng readability.js

                if html is not None and len(html) > 200:

                    html_filename = filename + ".html"
                    with open(html_filename, "wt") as f:
                        f.write(html)

                    abi_text = subprocess.run(
                        f"node node_readability.js '{html_filename}' '{url}'", 
                        shell=True, 
                        capture_output=True,
                    ).stdout.decode('utf-8')

                    abi_text, abi_noise_ratio, abi_max_noise = \
                        normalize_text(abi_text, get_noise_info = True)

                    if abi_max_noise < reader_max_noise:
                        print(GREEN, ">>>", abi_text, "<<<", RESET)

                        if len(abi_text) < len(reader_text) and len(text) < len(reader_text): # chuộng text ngắn
                            choose_original_text = True


            if not choose_original_text:

                extract_method = "reader"
                text = reader_text
                noise_ratio = reader_noise_ratio
                max_noise = reader_max_noise


    # update meta, Lưu lại text và meta
    meta["url"] = url
    meta["get_html_method"] = get_html_method
    meta["extract_method"] = extract_method
    meta["created_at"] = str(datetime.now())
    meta["updates_count"] = 0
    meta["noise_ratio"] = noise_ratio
    meta["max_noise"] = max_noise
    meta["text_origin_len"] = len(text)

    if "hostname" in meta: meta.pop("hostname")
    if "sitename" in meta: meta.pop("sitename")

    # Thêm title và description vào text (nếu có)
    norm_text = normalize_text(text)
    text = add_title_desc_to_text(norm_text, meta)

    # Cắt ngắn text nếu đầu vào quá dài
    text = cut_if_too_long(text, meta)

    print(CYAN,end=""); print(meta); print(RESET,end="")

    with open(json_filename, "wt") as f:
        f.write(json.dumps(meta, ensure_ascii = False))

    with open(text_filename, "wt") as f:
        f.write(text)

    get_clear_view(filename)
    get_llm_gen_contents(url, use_cache = False)

    return text


def get_clear_view(filename):
    # gọi gen_clear_view.py (chạy chậm) ở process riêng
    subprocess.run(f"nohup python3 gen_clear_view.py '{filename}' &", shell = True)
    import time; time.sleep(1) # chờ 1 giây


def cut_if_too_long(text, meta, max_words = config.text_max_words):
    words = text.split()

    if len(words) > max_words:

        words = words[ : max_words]
        threshold = len(" ".join(words))

        meta["text_cutoff"] = True
        meta["text_cutoff_len"] = threshold

        return text[ : threshold ]

    else:
        return text


def add_title_desc_to_text(text, meta):
    content = []
    title       = meta["title"]       if "title"       in meta else None
    description = meta["description"] if "description" in meta else None

    if title is not None and len(title) > 5:
        content.append(f"**title**: {title}")

    if description is not None and len(description) > 10:
        content.append(f"**description**: {description}")

    content.append(text)
    return "\n\n".join(content)



def normalize_text(text, get_noise_info = False):
    text = text.strip()
    chunks = re.split(r'\s*(?:\n\s*)+', text, flags = re.MULTILINE)
    text = "\n\n".join([ x for x in chunks if len(x) > 20 ])

    if get_noise_info:
        noise_len = 1
        total_len = 1

        max_noise = 0
        continuous_noise = 0
        for x in chunks:
            n = len(x)
            total_len += n

            if n < 80:
                noise_len += n
                continuous_noise += n
                if continuous_noise > max_noise:
                    max_noise = continuous_noise
            else:
                continuous_noise = 0

        noise_ratio = noise_len / total_len
        return text, noise_ratio, max_noise
    else:
        return text


def get_clean_view(url):
    url, filename = norm_url_and_gen_filename(url)
    clean_view_filename = filename + "__clean_view.txt"

    if os.path.exists(clean_view_filename):
        return open(clean_view_filename, "rt").read()
    else:
        return None


def get_meta(url):
    url, filename = norm_url_and_gen_filename(url)
    json_filename = filename + ".json"
    return json.load(open(json_filename))


TAGS = "keyphrases figures summary".split()
###
def get_llm_gen_contents(url, use_cache = True):
    url, filename = norm_url_and_gen_filename(url)
    json_filename = filename + ".json"
    text_filename = filename + ".txt"

    if os.path.exists(json_filename):
        meta = json.load(open(json_filename, "rt"))
        
        generated = ( "llm_generated" in meta )

        if not use_cache or not generated:

            text = open(text_filename, "rt").read()
            marked_text, chunks = add_chunk_markers(text, para = True)

            raw = extract_keyphrases_figures_summary(marked_text)
            result = extract_xmls(raw, TAGS)

            result["raw"] = raw
            meta["llm_generated"] = result

            with open(json_filename, "wt") as f:
                f.write(json.dumps(meta, ensure_ascii = False))

        return meta["llm_generated"]

    else:
        return {
            "summary": "Tóm tắt nội dung ... văn bản nói về ...",
            "keyphrases": ["keywords 1", "keywords 2", "keywords 3"]
        }


default_urls_input = """

https://thuvienphapluat.vn/phap-luat/ho-so-dien-tu-thuc-hien-thu-tuc-hanh-chinh-la-gi-huong-dan-chuan-bi-va-nop-ho-so-dien-tu-khi-thuc-h-155754-140107.html

https://video.vnexpress.net/bon-ngay-chong-choi-lu-ngap-gan-3-m-cua-nguoi-dan-thai-nguyen-4791440.html

http://danvan.vn/Home/Tin-hoat-dong/Ban-dan-van/18706/Ban-Dan-van-Trung-uong-va-Hoi-Chu-thap-do-Viet-Nam-tham-tang-qua-nhan-dan-bi-anh-huong-bao-so-3-tai-Thai-Nguyen

https://baodauthau.vn/thai-nguyen-144-ty-dong-nang-cap-duong-cach-mang-thang-8-tp-song-cong-post164486.html

https://baothainguyen.vn/chinh-tri/202409/chu-tich-quoc-hoi-tran-thanh-man-lam-viec-voi-tinh-thai-nguyen-ve-cong-tackhac-phuc-hau-qua-bao-so-3-3f9253f/

https://baothainguyen.vn/giao-duc/202409/dam-bao-dieu-kien-de-hoc-sinh-tro-lai-truong-cham-nhat-ngay-16-9-9742985/

https://baothainguyen.vn/tai-nguyen-moi-truong/202409/khu-khuan-dien-rong-nhung-vung-bi-ngap-lut-tai-tp-thai-nguyen-585273d/

https://baothainguyen.vn/thoi-su-thai-nguyen/202409/dien-luc-tp-thai-nguyen-no-luccap-dien-tro-lai-cho-tren-2000-hotrong-ngay-12-9-da21a20/

https://baothainguyen.vn/xa-hoi/202409/tao-sinh-ke-giam-ngheo-vung-dong-bao-dan-toc-thieu-so-b8f041c/

https://baotintuc.vn/xa-hoi/ngap-ung-va-thiet-hai-tren-202000-ha-lua-20240913095621343.htm

https://daidoanket.vn/thai-nguyen-hai-nguoi-tu-vong-thiet-hai-hon-600-ty-dong-do-bao-yagi-10290104.html

https://dangcongsan.vn/xay-dung-dang/thai-nguyen-cong-bo-cac-quyet-dinh-ve-cong-tac-can-bo-677747.html

https://danviet.vn/62-y-bac-si-cua-binh-dinh-den-thai-nguyen-yen-bai-quyet-tam-cung-dong-bao-vuot-qua-kho-khan-20240913101402511.htm

https://laodong.vn/thoi-su/chu-tich-quoc-hoi-kiem-tra-cong-tac-khac-phuc-hau-qua-mua-lu-o-thai-nguyen-1393445.ldo

https://nhandan.vn/anh-chu-tich-quoc-hoi-tran-thanh-man-kiem-tra-cong-tac-khac-phuc-hau-qua-bao-so-3-tai-tinh-thai-nguyen-post830447.html

https://nld.com.vn/toi-7-gio-13-9-336-nguoi-chet-va-mat-tich-hon-130-ngan-nguoi-dan-phai-di-doi-do-bao-lu-196240913101124546.htm

https://phunuvietnam.vn/thai-nguyen-hoi-vien-phu-nu-chung-tay-khac-phuc-hau-qua-ngap-lut-20240912154801867.htm

https://phunuvietnam.vn/thai-nguyen-trien-khai-cong-tac-phong-chong-dich-sau-thien-tai-20240912174641866.htm

https://thainguyen.dcs.vn/hoat-dong-cua-cac-dang-bo/dang-bo-tp-thai-nguyen/hoi-nghi-ban-thuong-vu-thanh-uy-thai-nguyen-lan-thu-102-857.html

https://thainguyen.dms.gov.vn/tin-chi-tiet/-/chi-tiet/thai-nguyen-%C4%91am-bao-nguon-hang-hoa-phuc-vu-nhan-dan-89820-1404.html

https://thuonghieucongluan.com.vn/thai-nguyen-tiep-nhan-5-tan-gao-ho-tro-nhan-dan-bi-anh-huong-ngap-lut-a235642.html

https://tienphong.vn/nam-thanh-nien-o-thai-nguyen-bi-lu-cuon-khi-di-bat-ca-post1672693.tpo

https://tienphong.vn/ngan-hang-dau-tien-cong-bo-giam-lai-suat-cho-vay-sau-bao-so-3-post1672728.tpo

https://tuoitre.vn/chu-tich-quoc-hoi-tran-thanh-man-trao-30-ti-dong-ho-tro-khac-phuc-bao-lu-tai-thai-nguyen-20240912191724375.htm

https://tuoitre.vn/sau-lu-nguoi-dan-thai-nguyen-noi-chua-bao-gio-bun-ngap-nhieu-den-vay-202409121653144.htm

https://vietnamnet.vn/muc-nuoc-song-cau-o-thai-nguyen-giam-dan-nguoi-dan-tat-bat-don-dep-sau-lu-2321461.html

https://vtcnews.vn/trieu-nu-cuoi-huong-ve-thai-nguyen-sau-con-bao-ar895714.html

""".strip()

default_urls_input = """

https://vnexpress.net/sam-altman-ai-thong-minh-hon-con-nguoi-trong-vai-nghin-ngay-toi-4796649.html

https://vnexpress.net/may-tram-chay-ai-gia-tram-trieu-dong-tai-viet-nam-4796490.html

https://www.vngcloud.vn/blog/what-are-large-language-models

https://arxiv.org/html/2408.16737v1

https://arxiv.org/html/2409.15700v1

https://arxiv.org/html/2409.09916v1

https://arxiv.org/html/2409.06903v1

https://arxiv.org/html/2409.12558v1

https://arxiv.org/html/2409.10516v2

https://rlhflow.github.io/posts/2024-05-29-multi-objective-reward-modeling

https://arxiv.org/html/2405.07863v2

https://arxiv.org/html/2406.12845

""".strip()