如何处理长文本进行提取

当处理像 PDF 这样的文件时，你可能会遇到文本长度超出语言模型上下文窗口的情况。要处理这类文本，可以考虑以下策略：

更换 LLM 选择一个支持更大上下文窗口的不同 LLM。
暴力破解 将文档分块，并从每个块中提取内容。
RAG 将文档分块，对分块进行索引，并且只从看起来“相关”的块子集中提取内容。

请记住，这些策略有不同的权衡取舍，最佳策略可能取决于你设计的应用程序！

本指南演示了如何实现策略 2 和 3。

设置

首先，我们将安装本指南所需的依赖项：

%pip install -qU langchain-community lxml faiss-cpu langchain-openai

Note: you may need to restart the kernel to use updated packages.

现在我们需要一些示例数据！让我们从 Wikipedia 下载一篇关于汽车的文章，并将其加载为 LangChain 的 Document。

import re

import requests
from langchain_community.document_loaders import BSHTMLLoader

# Download the content
response = requests.get("https://en.wikipedia.org/wiki/Car")
# Write it to a file
with open("car.html", "w", encoding="utf-8") as f:
    f.write(response.text)
# Load it with an HTML parser
loader = BSHTMLLoader("car.html")
document = loader.load()[0]
# Clean up code
# Replace consecutive new lines with a single new line
document.page_content = re.sub("\n\n+", "\n", document.page_content)

API Reference:BSHTMLLoader

print(len(document.page_content))

定义模式

遵循提取教程，我们将使用 Pydantic 来定义我们希望提取的信息的模式。在这种情况下，我们将提取一系列“关键进展”（例如重要的历史事件），其中包含年份和描述。

请注意，我们还包含了一个 evidence 键，并指示模型 verbatim 提供文章中的相关句子。这使我们能够将提取结果与（模型重建的）原始文档中的文本进行比较。

from typing import List, Optional

from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from pydantic import BaseModel, Field


class KeyDevelopment(BaseModel):
    """Information about a development in the history of cars."""

    year: int = Field(
        ..., description="The year when there was an important historic development."
    )
    description: str = Field(
        ..., description="What happened in this year? What was the development?"
    )
    evidence: str = Field(
        ...,
        description="Repeat in verbatim the sentence(s) from which the year and description information were extracted",
    )


class ExtractionData(BaseModel):
    """Extracted information about key developments in the history of cars."""

    key_developments: List[KeyDevelopment]


# Define a custom prompt to provide instructions and any additional context.
# 1) You can add examples into the prompt template to improve extraction quality
# 2) Introduce additional parameters to take context into account (e.g., include metadata
#    about the document from which the text was extracted.)
prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            "You are an expert at identifying key historic development in text. "
            "Only extract important historic developments. Extract nothing if no important information can be found in the text.",
        ),
        ("human", "{text}"),
    ]
)

API Reference:ChatPromptTemplate | MessagesPlaceholder

创建提取器

让我们选择一个 LLM。由于我们正在使用工具调用，因此需要一个支持工具调用功能的模型。请参阅此表了解可用的 LLM。

Select chat model:

pip install -qU "langchain[google-genai]"

import getpass
import os

if not os.environ.get("GOOGLE_API_KEY"):
  os.environ["GOOGLE_API_KEY"] = getpass.getpass("Enter API key for Google Gemini: ")

from langchain.chat_models import init_chat_model

llm = init_chat_model("gemini-2.0-flash", model_provider="google_genai")

extractor = prompt | llm.with_structured_output(
    schema=ExtractionData,
    include_raw=False,
)

暴力破解方法

将文档分割成块，使每个块都能放入 LLM 的上下文窗口。

from langchain_text_splitters import TokenTextSplitter

text_splitter = TokenTextSplitter(
    # Controls the size of each chunk
    chunk_size=2000,
    # Controls overlap between chunks
    chunk_overlap=20,
)

texts = text_splitter.split_text(document.page_content)

API Reference:TokenTextSplitter

使用 batch 功能在并行模式下运行跨每个块的提取！

tip

您通常可以使用 .batch() 来并行化提取！.batch 在底层使用线程池来帮助您并行化工作负载。

如果您的模型是通过 API 公开的，这很可能会加快您的提取流程！

# Limit just to the first 3 chunks
# so the code can be re-run quickly
first_few = texts[:3]

extractions = extractor.batch(
    [{"text": text} for text in first_few],
    {"max_concurrency": 5},  # limit the concurrency by passing max concurrency!
)

合并结果

在从所有分块中提取数据之后，我们需要将这些提取结果合并在一起。

key_developments = []

for extraction in extractions:
    key_developments.extend(extraction.key_developments)

key_developments[:10]

[KeyDevelopment(year=1769, description='Nicolas-Joseph Cugnot built the first steam-powered road vehicle.', evidence='The French inventor Nicolas-Joseph Cugnot built the first steam-powered road vehicle in 1769, while the Swiss inventor François Isaac de Rivaz designed and constructed the first internal combustion-powered automobile in 1808.'),
 KeyDevelopment(year=1808, description='François Isaac de Rivaz designed and constructed the first internal combustion-powered automobile.', evidence='The French inventor Nicolas-Joseph Cugnot built the first steam-powered road vehicle in 1769, while the Swiss inventor François Isaac de Rivaz designed and constructed the first internal combustion-powered automobile in 1808.'),
 KeyDevelopment(year=1886, description='Carl Benz invented the modern car, a practical, marketable automobile for everyday use, and patented his Benz Patent-Motorwagen.', evidence='The modern car—a practical, marketable automobile for everyday use—was invented in 1886, when the German inventor Carl Benz patented his Benz Patent-Motorwagen.'),
 KeyDevelopment(year=1901, description='The Oldsmobile Curved Dash became the first mass-produced car.', evidence='The 1901 Oldsmobile Curved Dash and the 1908 Ford Model T, both American cars, are widely considered the first mass-produced[3][4] and mass-affordable[5][6][7] cars, respectively.'),
 KeyDevelopment(year=1908, description='The Ford Model T became the first mass-affordable car.', evidence='The 1901 Oldsmobile Curved Dash and the 1908 Ford Model T, both American cars, are widely considered the first mass-produced[3][4] and mass-affordable[5][6][7] cars, respectively.'),
 KeyDevelopment(year=1885, description='Carl Benz built the original Benz Patent-Motorwagen, the first modern car.', evidence='The original Benz Patent-Motorwagen, the first modern car, built in 1885 and awarded the patent for the concept'),
 KeyDevelopment(year=1881, description='Gustave Trouvé demonstrated a three-wheeled car powered by electricity.', evidence='In November 1881, French inventor Gustave Trouvé demonstrated a three-wheeled car powered by electricity at the International Exposition of Electricity.'),
 KeyDevelopment(year=1888, description="Bertha Benz undertook the first road trip by car to prove the road-worthiness of her husband's invention.", evidence="In August 1888, Bertha Benz, the wife and business partner of Carl Benz, undertook the first road trip by car, to prove the road-worthiness of her husband's invention."),
 KeyDevelopment(year=1896, description='Benz designed and patented the first internal-combustion flat engine, called boxermotor.', evidence='In 1896, Benz designed and patented the first internal-combustion flat engine, called boxermotor.'),
 KeyDevelopment(year=1897, description='The first motor car in central Europe and one of the first factory-made cars in the world was produced by Czech company Nesselsdorfer Wagenbau (later renamed to Tatra), the Präsident automobil.', evidence='The first motor car in central Europe and one of the first factory-made cars in the world, was produced by Czech company Nesselsdorfer Wagenbau (later renamed to Tatra) in 1897, the Präsident automobil.')]

基于 RAG 的方法

另一个简单的想法是将文本分块，但不是从每个块中提取信息，而是只关注最相关的块。

caution

确定哪些块是相关的可能很困难。

例如，在我们这里使用的 car 文章中，文章的大部分都包含关键的开发信息。因此，使用 RAG，我们可能会舍弃大量相关信息。

我们建议根据你的用例进行实验，并确定此方法是否有效。

要实现基于 RAG 的方法：

将你的文档分块并建立索引（例如，在向量数据库中）；
在 extractor 链前面添加一个使用向量数据库的检索步骤。

下面是一个依赖于 FAISS 向量数据库的简单示例。

from langchain_community.vectorstores import FAISS
from langchain_core.documents import Document
from langchain_core.runnables import RunnableLambda
from langchain_openai import OpenAIEmbeddings
from langchain_text_splitters import CharacterTextSplitter

texts = text_splitter.split_text(document.page_content)
vectorstore = FAISS.from_texts(texts, embedding=OpenAIEmbeddings())

retriever = vectorstore.as_retriever(
    search_kwargs={"k": 1}
)  # Only extract from first document

API Reference:FAISS | Document | RunnableLambda | OpenAIEmbeddings | CharacterTextSplitter

在这种情况下，RAG 提取器只查看顶层文档。

rag_extractor = {
    "text": retriever | (lambda docs: docs[0].page_content)  # fetch content of top doc
} | extractor

results = rag_extractor.invoke("Key developments associated with cars")

for key_development in results.key_developments:
    print(key_development)

year=2006 description='Car-sharing services in the US experienced double-digit growth in revenue and membership.' evidence='in the US, some car-sharing services have experienced double-digit growth in revenue and membership growth between 2006 and 2007.'
year=2020 description='56 million cars were manufactured worldwide, with China producing the most.' evidence='In 2020, there were 56 million cars manufactured worldwide, down from 67 million the previous year. The automotive industry in China produces by far the most (20 million in 2020).'

常见问题

不同的方法在成本、速度和准确性方面都有各自的优缺点。

请注意以下问题：

分块处理内容意味着，如果信息分散在多个块中，大型语言模型（LLM）可能无法提取信息。
较大的块重叠可能会导致同一信息被提取两次，因此请做好去重的准备！
大型语言模型（LLM）可能会编造数据。如果寻找大段文本中的单个事实并使用暴力方法，您可能会得到更多编造的数据。

设置​

定义模式​

创建提取器​

暴力破解方法​

合并结果​

基于 RAG 的方法​

常见问题​

设置