from __future__ import annotations import asyncio import inspect import logging import re from typing import ( Callable, Iterator, List, Optional, Sequence, Set, Union, cast, ) import aiohttp import requests from langchain_core.documents import Document from langchain_core.utils.html import extract_sub_links from langchain_community.document_loaders.base import BaseLoader logger = logging.getLogger(__name__) def _metadata_extractor( raw_html: str, url: str, response: Union[requests.Response, aiohttp.ClientResponse] ) -> dict: """Extract metadata from raw html using BeautifulSoup.""" content_type = getattr(response, "headers").get("Content-Type", "") metadata = {"source": url, "content_type": content_type} try: from bs4 import BeautifulSoup except ImportError: logger.warning( "The bs4 package is required for default metadata extraction. " "Please install it with `pip install -U beautifulsoup4`." ) return metadata soup = BeautifulSoup(raw_html, "html.parser") if title := soup.find("title"): metadata["title"] = title.get_text() if description := soup.find("meta", attrs={"name": "description"}): metadata["description"] = description.get("content", None) if html := soup.find("html"): metadata["language"] = html.get("lang", None) return metadata class RecursiveUrlLoader(BaseLoader): """Recursively load all child links from a root URL. **Security Note**: This loader is a crawler that will start crawling at a given URL and then expand to crawl child links recursively. Web crawlers should generally NOT be deployed with network access to any internal servers. Control access to who can submit crawling requests and what network access the crawler has. While crawling, the crawler may encounter malicious URLs that would lead to a server-side request forgery (SSRF) attack. To mitigate risks, the crawler by default will only load URLs from the same domain as the start URL (controlled via prevent_outside named argument). This will mitigate the risk of SSRF attacks, but will not eliminate it. For example, if crawling a host which hosts several sites: https://some_host/alice_site/ https://some_host/bob_site/ A malicious URL on Alice's site could cause the crawler to make a malicious GET request to an endpoint on Bob's site. Both sites are hosted on the same host, so such a request would not be prevented by default. See https://python.langchain.com/v0.2/docs/security/ Setup: This class has no required additional dependencies. You can optionally install ``beautifulsoup4`` for richer default metadata extraction: .. code-block:: bash pip install -U beautifulsoup4 Instantiate: .. code-block:: python from langchain_community.document_loaders import RecursiveUrlLoader loader = RecursiveUrlLoader( "https://docs.python.org/3.9/", # max_depth=2, # use_async=False, # extractor=None, # metadata_extractor=None, # exclude_dirs=(), # timeout=10, # check_response_status=True, # continue_on_failure=True, # prevent_outside=True, # base_url=None, # ... ) Lazy load: .. code-block:: python docs = [] docs_lazy = loader.lazy_load() # async variant: # docs_lazy = await loader.alazy_load() for doc in docs_lazy: docs.append(doc) print(docs[0].page_content[:100]) print(docs[0].metadata) .. code-block:: python < {'source': 'https://docs.python.org/3.9/', 'content_type': 'text/html', 'title': '3.9.19 Documentation', 'language': None} Async load: .. code-block:: python docs = await loader.aload() print(docs[0].page_content[:100]) print(docs[0].metadata) .. code-block:: python < {'source': 'https://docs.python.org/3.9/', 'content_type': 'text/html', 'title': '3.9.19 Documentation', 'language': None} Content parsing / extraction: By default the loader sets the raw HTML from each link as the Document page content. To parse this HTML into a more human/LLM-friendly format you can pass in a custom ``extractor`` method: .. code-block:: python # This example uses `beautifulsoup4` and `lxml` import re from bs4 import BeautifulSoup def bs4_extractor(html: str) -> str: soup = BeautifulSoup(html, "lxml") return re.sub(r"\n\n+", "\n\n", soup.text).strip() loader = RecursiveUrlLoader( "https://docs.python.org/3.9/", extractor=bs4_extractor, ) print(loader.load()[0].page_content[:200]) .. code-block:: python 3.9.19 Documentation Download Download these documents Docs by version Python 3.13 (in development) Python 3.12 (stable) Python 3.11 (security-fixes) Python 3.10 (security-fixes) Python 3.9 (securit Metadata extraction: Similarly to content extraction, you can specify a metadata extraction function to customize how Document metadata is extracted from the HTTP response. .. code-block:: python import aiohttp import requests from typing import Union def simple_metadata_extractor( raw_html: str, url: str, response: Union[requests.Response, aiohttp.ClientResponse] ) -> dict: content_type = getattr(response, "headers").get("Content-Type", "") return {"source": url, "content_type": content_type} loader = RecursiveUrlLoader( "https://docs.python.org/3.9/", metadata_extractor=simple_metadata_extractor, ) loader.load()[0].metadata .. code-block:: python {'source': 'https://docs.python.org/3.9/', 'content_type': 'text/html'} Filtering URLs: You may not always want to pull every URL from a website. There are four parameters that allow us to control what URLs we pull recursively. First, we can set the ``prevent_outside`` parameter to prevent URLs outside of the ``base_url`` from being pulled. Note that the ``base_url`` does not need to be the same as the URL we pass in, as shown below. We can also use ``link_regex`` and ``exclude_dirs`` to be more specific with the URLs that we select. In this example, we only pull websites from the python docs, which contain the string "index" somewhere and are not located in the FAQ section of the website. .. code-block:: python loader = RecursiveUrlLoader( "https://docs.python.org/3.9/", prevent_outside=True, base_url="https://docs.python.org", link_regex=r']*?\s+)?href="([^"]*(?=index)[^"]*)"', exclude_dirs=['https://docs.python.org/3.9/faq'] ) docs = loader.load() .. code-block:: python ['https://docs.python.org/3.9/', 'https://docs.python.org/3.9/py-modindex.html', 'https://docs.python.org/3.9/genindex.html', 'https://docs.python.org/3.9/tutorial/index.html', 'https://docs.python.org/3.9/using/index.html', 'https://docs.python.org/3.9/extending/index.html', 'https://docs.python.org/3.9/installing/index.html', 'https://docs.python.org/3.9/library/index.html', 'https://docs.python.org/3.9/c-api/index.html', 'https://docs.python.org/3.9/howto/index.html', 'https://docs.python.org/3.9/distributing/index.html', 'https://docs.python.org/3.9/reference/index.html', 'https://docs.python.org/3.9/whatsnew/index.html'] """ # noqa: E501 def __init__( self, url: str, max_depth: Optional[int] = 2, use_async: Optional[bool] = None, extractor: Optional[Callable[[str], str]] = None, metadata_extractor: Optional[_MetadataExtractorType] = None, exclude_dirs: Optional[Sequence[str]] = (), timeout: Optional[int] = 10, prevent_outside: bool = True, link_regex: Union[str, re.Pattern, None] = None, headers: Optional[dict] = None, check_response_status: bool = False, continue_on_failure: bool = True, *, base_url: Optional[str] = None, autoset_encoding: bool = True, encoding: Optional[str] = None, ) -> None: """Initialize with URL to crawl and any subdirectories to exclude. Args: url: The URL to crawl. max_depth: The max depth of the recursive loading. use_async: Whether to use asynchronous loading. If True, lazy_load function will not be lazy, but it will still work in the expected way, just not lazy. extractor: A function to extract document contents from raw HTML. When extract function returns an empty string, the document is ignored. Default returns the raw HTML. metadata_extractor: A function to extract metadata from args: raw HTML, the source url, and the requests.Response/aiohttp.ClientResponse object (args in that order). Default extractor will attempt to use BeautifulSoup4 to extract the title, description and language of the page. ..code-block:: python import requests import aiohttp def simple_metadata_extractor( raw_html: str, url: str, response: Union[requests.Response, aiohttp.ClientResponse] ) -> dict: content_type = getattr(response, "headers").get("Content-Type", "") return {"source": url, "content_type": content_type} exclude_dirs: A list of subdirectories to exclude. timeout: The timeout for the requests, in the unit of seconds. If None then connection will not timeout. prevent_outside: If True, prevent loading from urls which are not children of the root url. link_regex: Regex for extracting sub-links from the raw html of a web page. headers: Default request headers to use for all requests. check_response_status: If True, check HTTP response status and skip URLs with error responses (400-599). continue_on_failure: If True, continue if getting or parsing a link raises an exception. Otherwise, raise the exception. base_url: The base url to check for outside links against. autoset_encoding: Whether to automatically set the encoding of the response. If True, the encoding of the response will be set to the apparent encoding, unless the `encoding` argument has already been explicitly set. encoding: The encoding of the response. If manually set, the encoding will be set to given value, regardless of the `autoset_encoding` argument. """ # noqa: E501 self.url = url self.max_depth = max_depth if max_depth is not None else 2 self.use_async = use_async if use_async is not None else False self.extractor = extractor if extractor is not None else lambda x: x metadata_extractor = ( metadata_extractor if metadata_extractor is not None else _metadata_extractor ) self.autoset_encoding = autoset_encoding self.encoding = encoding self.metadata_extractor = _wrap_metadata_extractor(metadata_extractor) self.exclude_dirs = exclude_dirs if exclude_dirs is not None else () if any(url.startswith(exclude_dir) for exclude_dir in self.exclude_dirs): raise ValueError( f"Base url is included in exclude_dirs. Received base_url: {url} and " f"exclude_dirs: {self.exclude_dirs}" ) self.timeout = timeout self.prevent_outside = prevent_outside if prevent_outside is not None else True self.link_regex = link_regex self.headers = headers self.check_response_status = check_response_status self.continue_on_failure = continue_on_failure self.base_url = base_url if base_url is not None else url def _get_child_links_recursive( self, url: str, visited: Set[str], *, depth: int = 0 ) -> Iterator[Document]: """Recursively get all child links starting with the path of the input URL. Args: url: The URL to crawl. visited: A set of visited URLs. depth: Current depth of recursion. Stop when depth >= max_depth. """ if depth >= self.max_depth: return # Get all links that can be accessed from the current URL visited.add(url) try: response = requests.get(url, timeout=self.timeout, headers=self.headers) if self.encoding is not None: response.encoding = self.encoding elif self.autoset_encoding: response.encoding = response.apparent_encoding if self.check_response_status and 400 <= response.status_code <= 599: raise ValueError(f"Received HTTP status {response.status_code}") except Exception as e: if self.continue_on_failure: logger.warning( f"Unable to load from {url}. Received error {e} of type " f"{e.__class__.__name__}" ) return else: raise e content = self.extractor(response.text) if content: yield Document( page_content=content, metadata=self.metadata_extractor(response.text, url, response), ) # Store the visited links and recursively visit the children sub_links = extract_sub_links( response.text, url, base_url=self.base_url, pattern=self.link_regex, prevent_outside=self.prevent_outside, exclude_prefixes=self.exclude_dirs, continue_on_failure=self.continue_on_failure, ) for link in sub_links: # Check all unvisited links if link not in visited: yield from self._get_child_links_recursive( link, visited, depth=depth + 1 ) async def _async_get_child_links_recursive( self, url: str, visited: Set[str], *, session: Optional[aiohttp.ClientSession] = None, depth: int = 0, ) -> List[Document]: """Recursively get all child links starting with the path of the input URL. Args: url: The URL to crawl. visited: A set of visited URLs. depth: To reach the current url, how many pages have been visited. """ if not self.use_async: raise ValueError( "Async functions forbidden when not initialized with `use_async`" ) if depth >= self.max_depth: return [] # Disable SSL verification because websites may have invalid SSL certificates, # but won't cause any security issues for us. close_session = session is None session = ( session if session is not None else aiohttp.ClientSession( connector=aiohttp.TCPConnector(ssl=False), timeout=aiohttp.ClientTimeout(total=self.timeout), headers=self.headers, ) ) visited.add(url) try: async with session.get(url) as response: text = await response.text() if self.check_response_status and 400 <= response.status <= 599: raise ValueError(f"Received HTTP status {response.status}") except (aiohttp.client_exceptions.InvalidURL, Exception) as e: if close_session: await session.close() if self.continue_on_failure: logger.warning( f"Unable to load {url}. Received error {e} of type " f"{e.__class__.__name__}" ) return [] else: raise e results = [] content = self.extractor(text) if content: results.append( Document( page_content=content, metadata=self.metadata_extractor(text, url, response), ) ) if depth < self.max_depth - 1: sub_links = extract_sub_links( text, url, base_url=self.base_url, pattern=self.link_regex, prevent_outside=self.prevent_outside, exclude_prefixes=self.exclude_dirs, continue_on_failure=self.continue_on_failure, ) # Recursively call the function to get the children of the children sub_tasks = [] to_visit = set(sub_links).difference(visited) for link in to_visit: sub_tasks.append( self._async_get_child_links_recursive( link, visited, session=session, depth=depth + 1 ) ) next_results = await asyncio.gather(*sub_tasks) for sub_result in next_results: if isinstance(sub_result, Exception) or sub_result is None: # We don't want to stop the whole process, so just ignore it # Not standard html format or invalid url or 404 may cause this. continue # locking not fully working, temporary hack to ensure deduplication results += [r for r in sub_result if r not in results] if close_session: await session.close() return results def lazy_load(self) -> Iterator[Document]: """Lazy load web pages. When use_async is True, this function will not be lazy, but it will still work in the expected way, just not lazy.""" visited: Set[str] = set() if self.use_async: results = asyncio.run( self._async_get_child_links_recursive(self.url, visited) ) return iter(results or []) else: return self._get_child_links_recursive(self.url, visited) _MetadataExtractorType1 = Callable[[str, str], dict] _MetadataExtractorType2 = Callable[ [str, str, Union[requests.Response, aiohttp.ClientResponse]], dict ] _MetadataExtractorType = Union[_MetadataExtractorType1, _MetadataExtractorType2] def _wrap_metadata_extractor( metadata_extractor: _MetadataExtractorType, ) -> _MetadataExtractorType2: if len(inspect.signature(metadata_extractor).parameters) == 3: return cast(_MetadataExtractorType2, metadata_extractor) else: def _metadata_extractor_wrapper( raw_html: str, url: str, response: Union[requests.Response, aiohttp.ClientResponse], ) -> dict: return cast(_MetadataExtractorType1, metadata_extractor)(raw_html, url) return _metadata_extractor_wrapper