非结构化

本笔记本介绍了如何使用 Unstructured 文档加载器 来加载多种类型的文件。Unstructured 目前支持加载文本文件、PowerPoint 演示文稿、HTML、PDF、图像等。

请参阅本指南，了解在本地设置 Unstructured 的更多说明，包括设置所需的系统依赖项。

概览

集成详情

类	包	本地	可序列化的	JS 支持
UnstructuredLoader	langchain_unstructured	✅	❌	✅

加载器功能

源	文档延迟加载	原生异步支持
UnstructuredLoader	✅	❌

设置

凭据

默认情况下，langchain-unstructured 会安装一个较小的版本，需要将分区逻辑卸载到 Unstructured API，这需要 API 密钥。如果您使用本地安装，则不需要 API 密钥。要获取您的 API 密钥，请前往此网站获取 API 密钥，然后在下方的单元格中进行设置：

import getpass
import os

if "UNSTRUCTURED_API_KEY" not in os.environ:
    os.environ["UNSTRUCTURED_API_KEY"] = getpass.getpass(
        "Enter your Unstructured API key: "
    )

安装

常规安装

运行本笔记本的其余部分需要以下包。

# Install package, compatible with API partitioning
%pip install --upgrade --quiet langchain-unstructured unstructured-client unstructured "unstructured[pdf]" python-magic

本地安装

如果您希望在本地运行分区逻辑，则需要安装一系列系统依赖项，具体详见此处的 Unstructured 文档。

例如，在 Mac 上，您可以使用以下命令安装所需的依赖项：

# base dependencies
brew install libmagic poppler tesseract

# If parsing xml / html documents:
brew install libxml2 libxslt

您可以使用以下命令安装本地所需的 pip 依赖项：

pip install "langchain-unstructured[local]"

初始化

UnstructuredLoader 允许从多种不同的文件类型中加载。要了解有关 unstructured 包的全部信息，请参阅其文档/。在本示例中，我们展示了从文本文件和 PDF 文件中加载的内容。

from langchain_unstructured import UnstructuredLoader

file_paths = [
    "./example_data/layout-parser-paper.pdf",
    "./example_data/state_of_the_union.txt",
]


loader = UnstructuredLoader(file_paths)

API 参考：UnstructuredLoader

加载

docs = loader.load()

docs[0]

INFO: pikepdf C++ to Python logger bridge initialized

Document(metadata={'source': './example_data/layout-parser-paper.pdf', 'coordinates': {'points': ((16.34, 213.36), (16.34, 253.36), (36.34, 253.36), (36.34, 213.36)), 'system': 'PixelSpace', 'layout_width': 612, 'layout_height': 792}, 'file_directory': './example_data', 'filename': 'layout-parser-paper.pdf', 'languages': ['eng'], 'last_modified': '2024-02-27T15:49:27', 'page_number': 1, 'filetype': 'application/pdf', 'category': 'UncategorizedText', 'element_id': 'd3ce55f220dfb75891b4394a18bcb973'}, page_content='1 2 0 2')

print(docs[0].metadata)

{'source': './example_data/layout-parser-paper.pdf', 'coordinates': {'points': ((16.34, 213.36), (16.34, 253.36), (36.34, 253.36), (36.34, 213.36)), 'system': 'PixelSpace', 'layout_width': 612, 'layout_height': 792}, 'file_directory': './example_data', 'filename': 'layout-parser-paper.pdf', 'languages': ['eng'], 'last_modified': '2024-02-27T15:49:27', 'page_number': 1, 'filetype': 'application/pdf', 'category': 'UncategorizedText', 'element_id': 'd3ce55f220dfb75891b4394a18bcb973'}

懒加载

pages = []
for doc in loader.lazy_load():
    pages.append(doc)

pages[0]

Document(metadata={'source': './example_data/layout-parser-paper.pdf', 'coordinates': {'points': ((16.34, 213.36), (16.34, 253.36), (36.34, 253.36), (36.34, 213.36)), 'system': 'PixelSpace', 'layout_width': 612, 'layout_height': 792}, 'file_directory': './example_data', 'filename': 'layout-parser-paper.pdf', 'languages': ['eng'], 'last_modified': '2024-02-27T15:49:27', 'page_number': 1, 'filetype': 'application/pdf', 'category': 'UncategorizedText', 'element_id': 'd3ce55f220dfb75891b4394a18bcb973'}, page_content='1 2 0 2')

后处理

如果您需要在提取后对 unstructured 元素进行后处理，可以在实例化 UnstructuredLoader 时，将一系列 str -> str 函数传递给 post_processors 关键字参数。这也适用于其他 Unstructured 加载器。下面是一个示例。

from langchain_unstructured import UnstructuredLoader
from unstructured.cleaners.core import clean_extra_whitespace

loader = UnstructuredLoader(
    "./example_data/layout-parser-paper.pdf",
    post_processors=[clean_extra_whitespace],
)

docs = loader.load()

docs[5:10]

API 参考：UnstructuredLoader

[Document(metadata={'source': './example_data/layout-parser-paper.pdf', 'coordinates': {'points': ((16.34, 393.9), (16.34, 560.0), (36.34, 560.0), (36.34, 393.9)), 'system': 'PixelSpace', 'layout_width': 612, 'layout_height': 792}, 'file_directory': './example_data', 'filename': 'layout-parser-paper.pdf', 'languages': ['eng'], 'last_modified': '2024-02-27T15:49:27', 'page_number': 1, 'parent_id': '89565df026a24279aaea20dc08cedbec', 'filetype': 'application/pdf', 'category': 'UncategorizedText', 'element_id': 'e9fa370aef7ee5c05744eb7bb7d9981b'}, page_content='2 v 8 4 3 5 1 . 3 0 1 2 : v i X r a'),
 Document(metadata={'source': './example_data/layout-parser-paper.pdf', 'coordinates': {'points': ((157.62199999999999, 114.23496279999995), (157.62199999999999, 146.5141628), (457.7358962799999, 146.5141628), (457.7358962799999, 114.23496279999995)), 'system': 'PixelSpace', 'layout_width': 612, 'layout_height': 792}, 'file_directory': './example_data', 'filename': 'layout-parser-paper.pdf', 'languages': ['eng'], 'last_modified': '2024-02-27T15:49:27', 'page_number': 1, 'filetype': 'application/pdf', 'category': 'Title', 'element_id': 'bde0b230a1aa488e3ce837d33015181b'}, page_content='LayoutParser: A Unified Toolkit for Deep Learning Based Document Image Analysis'),
 Document(metadata={'source': './example_data/layout-parser-paper.pdf', 'coordinates': {'points': ((134.809, 168.64029940800003), (134.809, 192.2517444), (480.5464199080001, 192.2517444), (480.5464199080001, 168.64029940800003)), 'system': 'PixelSpace', 'layout_width': 612, 'layout_height': 792}, 'file_directory': './example_data', 'filename': 'layout-parser-paper.pdf', 'languages': ['eng'], 'last_modified': '2024-02-27T15:49:27', 'page_number': 1, 'parent_id': 'bde0b230a1aa488e3ce837d33015181b', 'filetype': 'application/pdf', 'category': 'UncategorizedText', 'element_id': '54700f902899f0c8c90488fa8d825bce'}, page_content='Zejiang Shen1 ((cid:0)), Ruochen Zhang2, Melissa Dell3, Benjamin Charles Germain Lee4, Jacob Carlson3, and Weining Li5'),
 Document(metadata={'source': './example_data/layout-parser-paper.pdf', 'coordinates': {'points': ((207.23000000000002, 202.57205439999996), (207.23000000000002, 311.8195408), (408.12676, 311.8195408), (408.12676, 202.57205439999996)), 'system': 'PixelSpace', 'layout_width': 612, 'layout_height': 792}, 'file_directory': './example_data', 'filename': 'layout-parser-paper.pdf', 'languages': ['eng'], 'last_modified': '2024-02-27T15:49:27', 'page_number': 1, 'parent_id': 'bde0b230a1aa488e3ce837d33015181b', 'filetype': 'application/pdf', 'category': 'UncategorizedText', 'element_id': 'b650f5867bad9bb4e30384282c79bcfe'}, page_content='1 Allen Institute for AI shannons@allenai.org 2 Brown University ruochen zhang@brown.edu 3 Harvard University {melissadell,jacob carlson}@fas.harvard.edu 4 University of Washington bcgl@cs.washington.edu 5 University of Waterloo w422li@uwaterloo.ca'),
 Document(metadata={'source': './example_data/layout-parser-paper.pdf', 'coordinates': {'points': ((162.779, 338.45008160000003), (162.779, 566.8455408), (454.0372021523199, 566.8455408), (454.0372021523199, 338.45008160000003)), 'system': 'PixelSpace', 'layout_width': 612, 'layout_height': 792}, 'file_directory': './example_data', 'filename': 'layout-parser-paper.pdf', 'languages': ['eng'], 'last_modified': '2024-02-27T15:49:27', 'links': [{'text': ':// layout - parser . github . io', 'url': 'https://layout-parser.github.io', 'start_index': 1477}], 'page_number': 1, 'parent_id': 'bde0b230a1aa488e3ce837d33015181b', 'filetype': 'application/pdf', 'category': 'NarrativeText', 'element_id': 'cfc957c94fe63c8fd7c7f4bcb56e75a7'}, page_content='Abstract. Recent advances in document image analysis (DIA) have been primarily driven by the application of neural networks. Ideally, research outcomes could be easily deployed in production and extended for further investigation. However, various factors like loosely organized codebases and sophisticated model configurations complicate the easy reuse of im- portant innovations by a wide audience. Though there have been on-going efforts to improve reusability and simplify deep learning (DL) model development in disciplines like natural language processing and computer vision, none of them are optimized for challenges in the domain of DIA. This represents a major gap in the existing toolkit, as DIA is central to academic research across a wide range of disciplines in the social sciences and humanities. This paper introduces LayoutParser, an open-source library for streamlining the usage of DL in DIA research and applica- tions. The core LayoutParser library comes with a set of simple and intuitive interfaces for applying and customizing DL models for layout de- tection, character recognition, and many other document processing tasks. To promote extensibility, LayoutParser also incorporates a community platform for sharing both pre-trained models and full document digiti- zation pipelines. We demonstrate that LayoutParser is helpful for both lightweight and large-scale digitization pipelines in real-word use cases. The library is publicly available at https://layout-parser.github.io.')]

非结构化 API

如果您想使用更小的包快速上手并获取最新的分区功能，您可以 pip install unstructured-client 和 pip install langchain-unstructured。有关 UnstructuredLoader 的更多信息，请参阅 Unstructured 提供商页面。

当您传入api_key并设置partition_via_api=True时，加载器将使用托管的无服务器 Unstructured API 处理您的文档。您可以在此处生成免费的 Unstructured API 密钥。

查看说明 此处 如果您想自托管 Unstructured API 或在本地运行它。

from langchain_unstructured import UnstructuredLoader

loader = UnstructuredLoader(
    file_path="example_data/fake.docx",
    api_key=os.getenv("UNSTRUCTURED_API_KEY"),
    partition_via_api=True,
)

docs = loader.load()
docs[0]

API 参考：UnstructuredLoader

INFO: Preparing to split document for partition.
INFO: Given file doesn't have '.pdf' extension, so splitting is not enabled.
INFO: Partitioning without split.
INFO: Successfully partitioned the document.

Document(metadata={'source': 'example_data/fake.docx', 'category_depth': 0, 'filename': 'fake.docx', 'languages': ['por', 'cat'], 'filetype': 'application/vnd.openxmlformats-officedocument.wordprocessingml.document', 'category': 'Title', 'element_id': '56d531394823d81787d77a04462ed096'}, page_content='Lorem ipsum dolor sit amet.')

您还可以通过 UnstructuredLoader 在单个 API 中通过 Unstructured API 批量处理多个文件。

loader = UnstructuredLoader(
    file_path=["example_data/fake.docx", "example_data/fake-email.eml"],
    api_key=os.getenv("UNSTRUCTURED_API_KEY"),
    partition_via_api=True,
)

docs = loader.load()

print(docs[0].metadata["filename"], ": ", docs[0].page_content[:100])
print(docs[-1].metadata["filename"], ": ", docs[-1].page_content[:100])

INFO: Preparing to split document for partition.
INFO: Given file doesn't have '.pdf' extension, so splitting is not enabled.
INFO: Partitioning without split.
INFO: Successfully partitioned the document.
INFO: Preparing to split document for partition.
INFO: Given file doesn't have '.pdf' extension, so splitting is not enabled.
INFO: Partitioning without split.
INFO: Successfully partitioned the document.
``````output
fake.docx :  Lorem ipsum dolor sit amet.
fake-email.eml :  Violets are blue

非结构化 SDK 客户端

使用 Unstructured API 进行分区依赖于 Unstructured SDK 客户端。

如果您想自定义客户端，则必须将 UnstructuredClient 实例传递给 UnstructuredLoader。以下示例展示了如何自定义客户端的功能，例如使用您自己的 requests.Session()、传递替代的 server_url，以及自定义 RetryConfig 对象。有关自定义客户端或 SDK 客户端接受的其他参数的更多信息，请参阅 Unstructured Python SDK 文档和 API 参数 文档的客户端部分。请注意，所有 API 参数都应传递给 UnstructuredLoader。

警告： 以下示例可能未使用最新版本的 UnstructuredClient，未来版本中可能存在破坏性变更。有关最新示例，请参阅 Unstructured Python SDK 文档。

import requests
from langchain_unstructured import UnstructuredLoader
from unstructured_client import UnstructuredClient
from unstructured_client.utils import BackoffStrategy, RetryConfig

client = UnstructuredClient(
    api_key_auth=os.getenv(
        "UNSTRUCTURED_API_KEY"
    ),  # Note: the client API param is "api_key_auth" instead of "api_key"
    client=requests.Session(),  # Define your own requests session
    server_url="https://api.unstructuredapp.io/general/v0/general",  # Define your own api url
    retry_config=RetryConfig(
        strategy="backoff",
        retry_connection_errors=True,
        backoff=BackoffStrategy(
            initial_interval=500,
            max_interval=60000,
            exponent=1.5,
            max_elapsed_time=900000,
        ),
    ),  # Define your own retry config
)

loader = UnstructuredLoader(
    "./example_data/layout-parser-paper.pdf",
    partition_via_api=True,
    client=client,
    split_pdf_page=True,
    split_pdf_page_range=[1, 10],
)

docs = loader.load()

print(docs[0].metadata["filename"], ": ", docs[0].page_content[:100])

API 参考：UnstructuredLoader

INFO: Preparing to split document for partition.
INFO: Concurrency level set to 5
INFO: Splitting pages 1 to 10 (10 total)
INFO: Determined optimal split size of 2 pages.
INFO: Partitioning 5 files with 2 page(s) each.
INFO: Partitioning set #1 (pages 1-2).
INFO: Partitioning set #2 (pages 3-4).
INFO: Partitioning set #3 (pages 5-6).
INFO: Partitioning set #4 (pages 7-8).
INFO: Partitioning set #5 (pages 9-10).
INFO: HTTP Request: POST https://api.unstructuredapp.io/general/v0/general "HTTP/1.1 200 OK"
INFO: HTTP Request: POST https://api.unstructuredapp.io/general/v0/general "HTTP/1.1 200 OK"
INFO: HTTP Request: POST https://api.unstructuredapp.io/general/v0/general "HTTP/1.1 200 OK"
INFO: HTTP Request: POST https://api.unstructuredapp.io/general/v0/general "HTTP/1.1 200 OK"
INFO: Successfully partitioned set #1, elements added to the final result.
INFO: Successfully partitioned set #2, elements added to the final result.
INFO: Successfully partitioned set #3, elements added to the final result.
INFO: Successfully partitioned set #4, elements added to the final result.
INFO: Successfully partitioned set #5, elements added to the final result.
INFO: Successfully partitioned the document.
``````output
layout-parser-paper.pdf :  LayoutParser: A Unified Toolkit for Deep Learning Based Document Image Analysis

分块

UnstructuredLoader 不像旧的加载器 UnstructuredFileLoader 等那样支持将 mode 作为分组文本的参数。它转而支持“分块”。非结构化数据中的分块机制不同于您可能熟悉的其他基于纯文本特征（如可能表示段落边界或列表项边界的字符序列"\n\n"或"\n"）来形成块的分块机制。相反，所有文档都利用针对每种文档格式的特定知识进行拆分，将文档划分为语义单元（文档元素），我们仅在单个元素超过所需的最大块大小时才需要进行文本拆分。通常，分块会合并连续的元素，以形成尽可能大但不超过最大块大小的块。分块会产生一个由 CompositeElement、Table 或 TableChunk 元素组成的序列。每个“块”都是这三种类型之一的实例。

查看此页面以获取更多关于分块选项的详细信息，但要重现与mode="single"相同的行为，您可以设置chunking_strategy="basic"、max_characters=<some-really-big-number>和include_orig_elements=False。

from langchain_unstructured import UnstructuredLoader

loader = UnstructuredLoader(
    "./example_data/layout-parser-paper.pdf",
    chunking_strategy="basic",
    max_characters=1000000,
    include_orig_elements=False,
)

docs = loader.load()

print("Number of LangChain documents:", len(docs))
print("Length of text in the document:", len(docs[0].page_content))

API 参考：UnstructuredLoader

Number of LangChain documents: 1
Length of text in the document: 42772

加载网页

UnstructuredLoader 在本地运行时接受一个 web_url 关键字参数，该参数会填充底层 Unstructured partition 的 url 参数。这允许解析远程托管的文档，例如 HTML 网页。

使用示例：

from langchain_unstructured import UnstructuredLoader

loader = UnstructuredLoader(web_url="https://www.example.com")
docs = loader.load()

for doc in docs:
    print(f"{doc}\n")

API 参考：UnstructuredLoader

page_content='Example Domain' metadata={'category_depth': 0, 'languages': ['eng'], 'filetype': 'text/html', 'url': 'https://www.example.com', 'category': 'Title', 'element_id': 'fdaa78d856f9d143aeeed85bf23f58f8'}

page_content='This domain is for use in illustrative examples in documents. You may use this domain in literature without prior coordination or asking for permission.' metadata={'languages': ['eng'], 'parent_id': 'fdaa78d856f9d143aeeed85bf23f58f8', 'filetype': 'text/html', 'url': 'https://www.example.com', 'category': 'NarrativeText', 'element_id': '3652b8458b0688639f973fe36253c992'}

page_content='More information...' metadata={'category_depth': 0, 'link_texts': ['More information...'], 'link_urls': ['https://www.iana.org/domains/example'], 'languages': ['eng'], 'filetype': 'text/html', 'url': 'https://www.example.com', 'category': 'Title', 'element_id': '793ab98565d6f6d6f3a6d614e3ace2a9'}

API 参考

有关所有 UnstructuredLoader 功能和配置的详细文档，请参阅 API 参考：https://python.langchain.com/api_reference/unstructured/document_loaders/langchain_unstructured.document_loaders.UnstructuredLoader.html

相关

• 文档加载器 概念指南
• 文档加载器 操操作指南