139 lines
4.4 KiB
Python

from typing import Callable, Dict, List, Literal, Optional, Any, cast
from llama_index.readers.base import BaseReader
from llama_index.readers.schema.base import Document
from llama_index.langchain_helpers.text_splitter import TextSplitter
import unicodedata
from pathlib import Path
path = Path(__file__).parent / "Readability.js"
readabilityjs = ""
with open(path, "r") as f:
readabilityjs = f.read()
inject_readability = f"""
(function(){{
{readabilityjs}
function executor() {{
return new Readability({{}}, document).parse();
}}
return executor();
}}())
"""
def nfkc_normalize(text: str) -> str:
return unicodedata.normalize('NFKC', text)
class ReadabilityWebPageReader(BaseReader):
"""Readability Webpage Loader
Extracting relevant information from a fully rendered web page.
During the processing, it is always assumed that web pages used as data sources contain textual content.
1. Load the page and wait for it rendered. (playwright)
2. Inject Readability.js to extract the main content.
Args:
proxy (Optional[str], optional): Proxy server. Defaults to None.
wait_until (Optional[Literal["commit", "domcontentloaded", "load", "networkidle"]], optional): Wait until the page is loaded. Defaults to "domcontentloaded".
text_splitter (TextSplitter, optional): Text splitter. Defaults to None.
normalizer (Optional[Callable[[str], str]], optional): Text normalizer. Defaults to nfkc_normalize.
"""
def __init__(self, proxy: Optional[str] = None, wait_until: Optional[
Literal["commit", "domcontentloaded", "load", "networkidle"]
] = "domcontentloaded",
text_splitter: Optional[TextSplitter] = None,
normalize: Optional[Callable[[str], str]] = nfkc_normalize
) -> None:
self._launch_options = {
"headless": True,
}
self._wait_until = wait_until
if proxy:
self._launch_options["proxy"] = {
"server": proxy,
}
self._text_splitter = text_splitter
self._normalize = normalize
def load_data(self, url: str) -> List[Document]:
"""render and load data content from url.
Args:
url (str): URL to scrape.
Returns:
List[Document]: List of documents.
"""
from playwright.sync_api import sync_playwright
with sync_playwright() as p:
browser = p.chromium.launch(**self._launch_options)
article = self.scrape_page(
browser,
url,
)
extra_info = {key: article[key] for key in [
"title",
"length",
"excerpt",
"byline",
"dir",
"lang",
"siteName",
]}
if self._normalize is not None:
article["textContent"] = self._normalize(article["textContent"])
texts = []
if self._text_splitter is not None:
texts = self._text_splitter.split_text(article["textContent"])
else:
texts = [article["textContent"]]
browser.close()
return [Document(x, extra_info=extra_info) for x in texts]
def scrape_page(
self,
browser: Any,
url: str,
) -> Dict[str, str]:
"""Scrape a single article url.
Args:
browser (Any): a Playwright Chromium browser.
url (str): URL of the article to scrape.
Returns:
Ref: https://github.com/mozilla/readability
title: article title;
content: HTML string of processed article content;
textContent: text content of the article, with all the HTML tags removed;
length: length of an article, in characters;
excerpt: article description, or short excerpt from the content;
byline: author metadata;
dir: content direction;
siteName: name of the site.
lang: content language
"""
from playwright.sync_api._generated import Browser
browser = cast(Browser, browser)
page = browser.new_page(ignore_https_errors=True)
page.set_default_timeout(60000)
page.goto(url, wait_until=self._wait_until)
r = page.evaluate(inject_readability)
page.close()
print("scraped:", url)
return r