"""Loader that uses unstructured to load files."""
from __future__ import annotations
import logging
import os
from abc import ABC, abstractmethod
from pathlib import Path
from typing import IO, Any, Callable, Iterator, List, Optional, Sequence, Union
from langchain_core._api.deprecation import deprecated
from langchain_core.documents import Document
from typing_extensions import TypeAlias
from langchain_community.document_loaders.base import BaseLoader
Element: TypeAlias = Any
logger = logging.getLogger(__file__)
[docs]
def satisfies_min_unstructured_version(min_version: str) -> bool:
"""Check if the installed `Unstructured` version exceeds the minimum version
for the feature in question."""
from unstructured.__version__ import __version__ as __unstructured_version__
min_version_tuple = tuple([int(x) for x in min_version.split(".")])
# NOTE(MthwRobinson) - enables the loader to work when you're using pre-release
# versions of unstructured like 0.4.17-dev1
_unstructured_version = __unstructured_version__.split("-")[0]
unstructured_version_tuple = tuple(
[int(x) for x in _unstructured_version.split(".")]
)
return unstructured_version_tuple >= min_version_tuple
[docs]
def validate_unstructured_version(min_unstructured_version: str) -> None:
"""Raise an error if the `Unstructured` version does not exceed the
specified minimum."""
if not satisfies_min_unstructured_version(min_unstructured_version):
raise ValueError(
f"unstructured>={min_unstructured_version} is required in this loader."
)
[docs]
class UnstructuredBaseLoader(BaseLoader, ABC):
"""Base Loader that uses `Unstructured`."""
[docs]
def __init__(
self,
mode: str = "single", # deprecated
post_processors: Optional[List[Callable[[str], str]]] = None,
**unstructured_kwargs: Any,
):
"""Initialize with file path."""
try:
import unstructured # noqa:F401
except ImportError:
raise ImportError(
"unstructured package not found, please install it with "
"`pip install unstructured`"
)
# `single` - elements are combined into one (default)
# `elements` - maintain individual elements
# `paged` - elements are combined by page
_valid_modes = {"single", "elements", "paged"}
if mode not in _valid_modes:
raise ValueError(
f"Got {mode} for `mode`, but should be one of `{_valid_modes}`"
)
if not satisfies_min_unstructured_version("0.5.4"):
if "strategy" in unstructured_kwargs:
unstructured_kwargs.pop("strategy")
self._check_if_both_mode_and_chunking_strategy_are_by_page(
mode, unstructured_kwargs
)
self.mode = mode
self.unstructured_kwargs = unstructured_kwargs
self.post_processors = post_processors or []
@abstractmethod
def _get_elements(self) -> List[Element]:
"""Get elements."""
@abstractmethod
def _get_metadata(self) -> dict[str, Any]:
"""Get file_path metadata if available."""
def _post_process_elements(self, elements: List[Element]) -> List[Element]:
"""Apply post processing functions to extracted unstructured elements.
Post processing functions are str -> str callables passed
in using the post_processors kwarg when the loader is instantiated.
"""
for element in elements:
for post_processor in self.post_processors:
element.apply(post_processor)
return elements
[docs]
def lazy_load(self) -> Iterator[Document]:
"""Load file."""
elements = self._get_elements()
self._post_process_elements(elements)
if self.mode == "elements":
for element in elements:
metadata = self._get_metadata()
# NOTE(MthwRobinson) - the attribute check is for backward compatibility
# with unstructured<0.4.9. The metadata attributed was added in 0.4.9.
if hasattr(element, "metadata"):
metadata.update(element.metadata.to_dict())
if hasattr(element, "category"):
metadata["category"] = element.category
if element.to_dict().get("element_id"):
metadata["element_id"] = element.to_dict().get("element_id")
yield Document(page_content=str(element), metadata=metadata)
elif self.mode == "paged":
logger.warning(
"`mode='paged'` is deprecated in favor of the 'by_page' chunking"
" strategy. Learn more about chunking here:"
" https://docs.unstructured.io/open-source/core-functionality/chunking"
)
text_dict: dict[int, str] = {}
meta_dict: dict[int, dict[str, Any]] = {}
for element in elements:
metadata = self._get_metadata()
if hasattr(element, "metadata"):
metadata.update(element.metadata.to_dict())
page_number = metadata.get("page_number", 1)
# Check if this page_number already exists in text_dict
if page_number not in text_dict:
# If not, create new entry with initial text and metadata
text_dict[page_number] = str(element) + "\n\n"
meta_dict[page_number] = metadata
else:
# If exists, append to text and update the metadata
text_dict[page_number] += str(element) + "\n\n"
meta_dict[page_number].update(metadata)
# Convert the dict to a list of Document objects
for key in text_dict.keys():
yield Document(page_content=text_dict[key], metadata=meta_dict[key])
elif self.mode == "single":
metadata = self._get_metadata()
text = "\n\n".join([str(el) for el in elements])
yield Document(page_content=text, metadata=metadata)
else:
raise ValueError(f"mode of {self.mode} not supported.")
def _check_if_both_mode_and_chunking_strategy_are_by_page(
self, mode: str, unstructured_kwargs: dict[str, Any]
) -> None:
if (
mode == "paged"
and unstructured_kwargs.get("chunking_strategy") == "by_page"
):
raise ValueError(
"Only one of `chunking_strategy='by_page'` or `mode='paged'` may be"
" set. `chunking_strategy` is preferred."
)
[docs]
@deprecated(
since="0.2.8",
removal="1.0",
alternative_import="langchain_unstructured.UnstructuredLoader",
)
class UnstructuredFileLoader(UnstructuredBaseLoader):
"""Load files using `Unstructured`.
The file loader uses the unstructured partition function and will automatically
detect the file type. You can run the loader in different modes: "single",
"elements", and "paged". The default "single" mode will return a single langchain
Document object. If you use "elements" mode, the unstructured library will split
the document into elements such as Title and NarrativeText and return those as
individual langchain Document objects. In addition to these post-processing modes
(which are specific to the LangChain Loaders), Unstructured has its own "chunking"
parameters for post-processing elements into more useful chunks for uses cases such
as Retrieval Augmented Generation (RAG). You can pass in additional unstructured
kwargs to configure different unstructured settings.
Examples
--------
from langchain_community.document_loaders import UnstructuredFileLoader
loader = UnstructuredFileLoader(
"example.pdf", mode="elements", strategy="fast",
)
docs = loader.load()
References
----------
https://docs.unstructured.io/open-source/core-functionality/partitioning
https://docs.unstructured.io/open-source/core-functionality/chunking
"""
[docs]
def __init__(
self,
file_path: Union[str, List[str], Path, List[Path]],
*,
mode: str = "single",
**unstructured_kwargs: Any,
):
"""Initialize with file path."""
self.file_path = file_path
super().__init__(mode=mode, **unstructured_kwargs)
def _get_elements(self) -> List[Element]:
from unstructured.partition.auto import partition
if isinstance(self.file_path, list):
elements: List[Element] = []
for file in self.file_path:
if isinstance(file, Path):
file = str(file)
elements.extend(partition(filename=file, **self.unstructured_kwargs))
return elements
else:
if isinstance(self.file_path, Path):
self.file_path = str(self.file_path)
return partition(filename=self.file_path, **self.unstructured_kwargs)
def _get_metadata(self) -> dict[str, Any]:
return {"source": self.file_path}
[docs]
def get_elements_from_api(
file_path: Union[str, List[str], Path, List[Path], None] = None,
file: Union[IO[bytes], Sequence[IO[bytes]], None] = None,
api_url: str = "https://api.unstructuredapp.io/general/v0/general",
api_key: str = "",
**unstructured_kwargs: Any,
) -> List[Element]:
"""Retrieve a list of elements from the `Unstructured API`."""
if is_list := isinstance(file_path, list):
file_path = [str(path) for path in file_path]
if isinstance(file, Sequence) or is_list:
from unstructured.partition.api import partition_multiple_via_api
_doc_elements = partition_multiple_via_api(
filenames=file_path, # type: ignore
files=file, # type: ignore
api_key=api_key,
api_url=api_url,
**unstructured_kwargs,
)
elements = []
for _elements in _doc_elements:
elements.extend(_elements)
return elements
else:
from unstructured.partition.api import partition_via_api
return partition_via_api(
filename=str(file_path) if file_path is not None else None,
file=file,
api_key=api_key,
api_url=api_url,
**unstructured_kwargs,
)
[docs]
@deprecated(
since="0.2.8",
removal="1.0",
alternative_import="langchain_unstructured.UnstructuredLoader",
)
class UnstructuredAPIFileLoader(UnstructuredBaseLoader):
"""Load files using `Unstructured` API.
By default, the loader makes a call to the hosted Unstructured API. If you are
running the unstructured API locally, you can change the API rule by passing in the
url parameter when you initialize the loader. The hosted Unstructured API requires
an API key. See the links below to learn more about our API offerings and get an
API key.
You can run the loader in different modes: "single", "elements", and "paged". The
default "single" mode will return a single langchain Document object. If you use
"elements" mode, the unstructured library will split the document into elements such
as Title and NarrativeText and return those as individual langchain Document
objects. In addition to these post-processing modes (which are specific to the
LangChain Loaders), Unstructured has its own "chunking" parameters for
post-processing elements into more useful chunks for uses cases such as Retrieval
Augmented Generation (RAG). You can pass in additional unstructured kwargs to
configure different unstructured settings.
Examples
```python
from langchain_community.document_loaders import UnstructuredAPIFileLoader
loader = UnstructuredAPIFileLoader(
"example.pdf", mode="elements", strategy="fast", api_key="MY_API_KEY",
)
docs = loader.load()
References
----------
https://docs.unstructured.io/api-reference/api-services/sdk
https://docs.unstructured.io/api-reference/api-services/overview
https://docs.unstructured.io/open-source/core-functionality/partitioning
https://docs.unstructured.io/open-source/core-functionality/chunking
"""
[docs]
def __init__(
self,
file_path: Union[str, List[str]],
*,
mode: str = "single",
url: str = "https://api.unstructuredapp.io/general/v0/general",
api_key: str = "",
**unstructured_kwargs: Any,
):
"""Initialize with file path."""
validate_unstructured_version(min_unstructured_version="0.10.15")
self.file_path = file_path
self.url = url
self.api_key = os.getenv("UNSTRUCTURED_API_KEY") or api_key
super().__init__(mode=mode, **unstructured_kwargs)
def _get_metadata(self) -> dict[str, Any]:
return {"source": self.file_path}
def _get_elements(self) -> List[Element]:
return get_elements_from_api(
file_path=self.file_path,
api_key=self.api_key,
api_url=self.url,
**self.unstructured_kwargs,
)
def _post_process_elements(self, elements: List[Element]) -> List[Element]:
"""Apply post processing functions to extracted unstructured elements.
Post processing functions are str -> str callables passed
in using the post_processors kwarg when the loader is instantiated.
"""
for element in elements:
for post_processor in self.post_processors:
element.apply(post_processor)
return elements
[docs]
@deprecated(
since="0.2.8",
removal="1.0",
alternative_import="langchain_unstructured.UnstructuredLoader",
)
class UnstructuredFileIOLoader(UnstructuredBaseLoader):
"""Load file-like objects opened in read mode using `Unstructured`.
The file loader uses the unstructured partition function and will automatically
detect the file type. You can run the loader in different modes: "single",
"elements", and "paged". The default "single" mode will return a single langchain
Document object. If you use "elements" mode, the unstructured library will split
the document into elements such as Title and NarrativeText and return those as
individual langchain Document objects. In addition to these post-processing modes
(which are specific to the LangChain Loaders), Unstructured has its own "chunking"
parameters for post-processing elements into more useful chunks for uses cases
such as Retrieval Augmented Generation (RAG). You can pass in additional
unstructured kwargs to configure different unstructured settings.
Examples
--------
from langchain_community.document_loaders import UnstructuredFileIOLoader
with open("example.pdf", "rb") as f:
loader = UnstructuredFileIOLoader(
f, mode="elements", strategy="fast",
)
docs = loader.load()
References
----------
https://docs.unstructured.io/open-source/core-functionality/partitioning
https://docs.unstructured.io/open-source/core-functionality/chunking
"""
[docs]
def __init__(
self,
file: IO[bytes],
*,
mode: str = "single",
**unstructured_kwargs: Any,
):
"""Initialize with file path."""
self.file = file
super().__init__(mode=mode, **unstructured_kwargs)
def _get_elements(self) -> List[Element]:
from unstructured.partition.auto import partition
return partition(file=self.file, **self.unstructured_kwargs)
def _get_metadata(self) -> dict[str, Any]:
return {}
def _post_process_elements(self, elements: List[Element]) -> List[Element]:
"""Apply post processing functions to extracted unstructured elements.
Post processing functions are str -> str callables passed
in using the post_processors kwarg when the loader is instantiated.
"""
for element in elements:
for post_processor in self.post_processors:
element.apply(post_processor)
return elements
[docs]
@deprecated(
since="0.2.8",
removal="1.0",
alternative_import="langchain_unstructured.UnstructuredLoader",
)
class UnstructuredAPIFileIOLoader(UnstructuredBaseLoader):
"""Send file-like objects with `unstructured-client` sdk to the Unstructured API.
By default, the loader makes a call to the hosted Unstructured API. If you are
running the unstructured API locally, you can change the API rule by passing in the
url parameter when you initialize the loader. The hosted Unstructured API requires
an API key. See the links below to learn more about our API offerings and get an
API key.
You can run the loader in different modes: "single", "elements", and "paged". The
default "single" mode will return a single langchain Document object. If you use
"elements" mode, the unstructured library will split the document into elements
such as Title and NarrativeText and return those as individual langchain Document
objects. In addition to these post-processing modes (which are specific to the
LangChain Loaders), Unstructured has its own "chunking" parameters for
post-processing elements into more useful chunks for uses cases such as Retrieval
Augmented Generation (RAG). You can pass in additional unstructured kwargs to
configure different unstructured settings.
Examples
--------
from langchain_community.document_loaders import UnstructuredAPIFileLoader
with open("example.pdf", "rb") as f:
loader = UnstructuredAPIFileIOLoader(
f, mode="elements", strategy="fast", api_key="MY_API_KEY",
)
docs = loader.load()
References
----------
https://docs.unstructured.io/api-reference/api-services/sdk
https://docs.unstructured.io/api-reference/api-services/overview
https://docs.unstructured.io/open-source/core-functionality/partitioning
https://docs.unstructured.io/open-source/core-functionality/chunking
"""
[docs]
def __init__(
self,
file: Union[IO[bytes], Sequence[IO[bytes]]],
*,
mode: str = "single",
url: str = "https://api.unstructuredapp.io/general/v0/general",
api_key: str = "",
**unstructured_kwargs: Any,
):
"""Initialize with file path."""
if isinstance(file, Sequence):
validate_unstructured_version(min_unstructured_version="0.6.3")
validate_unstructured_version(min_unstructured_version="0.6.2")
self.file = file
self.url = url
self.api_key = os.getenv("UNSTRUCTURED_API_KEY") or api_key
super().__init__(mode=mode, **unstructured_kwargs)
def _get_elements(self) -> List[Element]:
if self.unstructured_kwargs.get("metadata_filename"):
return get_elements_from_api(
file=self.file,
file_path=self.unstructured_kwargs.pop("metadata_filename"),
api_key=self.api_key,
api_url=self.url,
**self.unstructured_kwargs,
)
else:
raise ValueError(
"If partitioning a file via api,"
" metadata_filename must be specified as well.",
)
def _get_metadata(self) -> dict[str, Any]:
return {}
def _post_process_elements(self, elements: List[Element]) -> List[Element]:
"""Apply post processing functions to extracted unstructured elements.
Post processing functions are str -> str callables passed
in using the post_processors kwarg when the loader is instantiated.
"""
for element in elements:
for post_processor in self.post_processors:
element.apply(post_processor)
return elements
