微信扫码
与创始人交个朋友
我要投稿
本篇文章你将了解到图数据库的基础知识,以及它们为何在基于 RAG 的应用场景中具有相关性
前言
在 LangGraph 入门(1)-介绍 文章中,我探讨了大语言模型(LLM)加持下的应用中,图形技术的运用。我们深入了解了这些数据结构如何在多代理框架中被利用。更具体地说,我们介绍了一个名为 LangGraph 的新库,它在2024年1月推出,以图形(一个数学概念)作为代理应用的框架基础。
LangGraph 旨在克服传统 LangChain 链条的主要限制——即它们在运行时缺少循环。这个问题可以通过引入图形结构来轻松解决,因为图形结构可以轻松地在链中引入循环,而这些链本质上是有向无环图(DAGs)。
此外,在组织检索增强生成(RAG)场景中的知识库时,图形工具也显示出其强大的能力。具体来说,它们能够增强“检索”阶段,从而实现更有意义的上下文检索,最终提高生成响应的准确性。为了达到这一目标,我们的策略是将知识库存储在基于图的数据库(例如Neo4j)中,借助LLM的语义能力来正确识别和映射实体和关系。
为了现这一目标,LangChain 开发了一个名为 LLMGraphTransformer 的强大库,其目标是将非结构化文本数据转换为基于图的表示形式。
为了更好地理解这个库的工作原理,我们首先了解一下图的工作原理及相关术语。
图是一种用来建模对象之间成对关系的数学结构。它主要由两个元素组成:节点和关系。
节点:节点可以被看作是传统数据库中的记录。每个节点代表一个对象或实体,比如一个人或一个地方。节点通过标签进行分类,这有助于根据它们的角色(比如“顾客”或“产品”)对它们进行分类和查询。
关系:这些是节点之间的连接,定义了不同实体之间的交互或关系。例如,一个人可能通过“EMPLOYED_BY”关系与公司相连,或通过“LIVES_IN”关系与某地相连。为了以这种结构存储数据,后面便有了图数据库的出现。图数据库被设计成将数据之间的关系视为与数据本身同等重要,它们优化了处理互联数据和复杂查询的效率。其中最知名且广泛使用的是 Neo4j,它采用了一种灵活的图结构,除了节点和关系外,还包括属性、标签和路径特征来表示和存储数据。
属性:节点和关系都可以包含属性,这些属性以键值对的形式存储。这些属性提供了关于实体的具体细节,比如一个人的姓名或年龄,或一段关系的长度。
标签:标签是分配给节点的标记,用于将它们分类为不同类型。一个节点可以有多个标签,这有助于更灵活和动态地查询图。
路径:路径描述了连接它们的节点和关系的序列。它们表示图中的路线,显示了不同节点如何相互连接。在查询中,路径用于发现节点之间的关系,如在社交网络中发现从一个人到另一个人的所有可能路径。这使得Neo4j特别适合于社交网络、推荐系统和欺诈检测等应用,其中关系和动态查询至关重要。
检索增强生成(RAG)是一种在LLM驱动应用场景中的强大技术,它解决了这样一个问题:“如果我想要向我的LLM询问一些不在其训练集中的信息怎么办?”。RAG 背后的理念是将 LLM 与我们想要浏览的知识库分离,后者被适当地向量化或嵌入并存储到向量数据库中。
RAG 分为三个阶段:
检索:根据用户的查询及其对应的向量,检索最相似的文档片段(那些与用户查询向量更接近的向量对应的文档片段),作为LLM的基本上下文。
增强:通过额外的指令、规则、安全防护措施和类似的做法来丰富检索到的上下文,这些做法是提示工程技术的典型。
生成:基于增强的上下文,LLM生成对用户查询的响应。
如前所述,典型的 RAG 应用都假设有一个底层的 VectorDB,并存储了所有嵌入的知识库。然而,当涉及到 GraphRAG 时,这种方法稍有不同。
事实上,GraphRAG 在“检索”阶段发挥作用,利用图结构的灵活性来存储知识库,目标是检索到更相关的文档片段,然后将其增强为上下文(你可以在这里阅读微软关于GraphRAG的第一个实验(https://www.microsoft.com/en-us/research/blog/graphrag-unlocking-llm-discovery-on-narrative-private-data/))。你可以利用图数据库检索知识的两种主要方式:
你可以完全依赖图搜索(这是一种关键字搜索)来检索你的相关文档,然后将它们用作(增强的)上下文来生成最终响应。
你可以将图搜索与更高级的 LLM 相关搜索结合起来,例如通过嵌入进行向量搜索。
注意:Neo4j 也支持向量搜索,这使得它非常适合混合 GraphRAG 场景。
无论你采取哪种方法,过程中的关键步骤是将你的知识库组织到你的图中。LLMGraphTransformer 就是这么一个强大的库,他可以将非结构化知识映射到你的图数据库变得更容易,下文字中,我们会用 Neo4j 作为用例来实现。
LangChain 有一个特别活跃的生态系统,包括库、预建组件和连接器,使得将 LLM 集成到你的应用中变得更容易。它的最新发布之一确实是朝着 GraphRAG 的方向:LLMGraphTransformer。
LLMGraphTransformer 强大之处在于,它利用 LLM 来解析和分类文本中的实体及其关系。实际上,得益于 LLM 的自然语言能力,生成的图倾向于捕捉文档中最复杂的相互连接,与以前的方法相比,这使得它极其准确。
因此,你可以仅用几行代码就从你的非结构化文档中得到一个完全填充的图(如果你想查看其背后的逻辑,你可以在这里看到源代码(https://api.python.langchain.com/en/latest/_modules/langchain_experimental/graph_transformers/llm.html#LLMGraphTransformer))。
我们来看一个例子。首先,我将使用一个免费的 Neo4j Aura 数据库实例(你可以按照这个教程创建你自己的Neo4j 数据库(https://neo4j.com/cloud/platform/aura-graph-database/?utm_source=Google&utm_medium=PaidSearch&utm_campaign=UCGenAI&utm_content=AMS-Search-SEMCO-UCGenAI-None-SEM-SEM-NonABM&utm_term=neo4j+ai&utm_adgroup=genai-llm&gad_source=1&gclid=CjwKCAjwxLKxBhA7EiwAXO0R0E7DBMdsxo8LGQQU_oHNTuC5jfoeq0gf69CBYGxVSNr5ibUyvIxiyRoCOYYQAvD_BwE))和Azure OpenAI GPT-4模型。
一旦你的 AuraDB 实例被创建,你将能够在这里看到它运行(https://login.neo4j.com/u/signup/identifier?state=hKFo2SBPbDhNSW4zQnlpX1JuVkZTUExOTmJGNjd3WFhXU19Pb6Fur3VuaXZlcnNhbC1sb2dpbqN0aWTZIHVweGxoUE5nUXAyOC1Lbm5WQUpLS1hla0xrT3J3dlo4o2NpZNkgV1NMczYwNDdrT2pwVVNXODNnRFo0SnlZaElrNXpZVG8):
这些是将需要用来连接到实例的变量:
os.environ["NEO4J_URI"] = os.getenv("NEO4J_URI")os.environ["NEO4J_USERNAME"] = "neo4j"os.environ["NEO4J_PASSWORD"] = os.getenv("NEO4J_PASSWORD")api_key = os.getenv("AZURE_OPENAI_API_KEY")azure_endpoint = os.getenv("AZURE_OPENAI_ENDPOINT")api_version = "2023-07-01-preview"
我们初始化LLM:
llm = AzureChatOpenAI(model="gpt-4",azure_deployment="gpt-4",api_key=api_key,azure_endpoint=azure_endpoint,openai_api_version=api_version,)
我们选择使用《哈利波特与魔法石》的前几行作为示例文档(https://docenti.unimc.it/antonella.pascali/teaching/2018/19055/files/ultima-lezione/harry-potter-and-the-philosophers-stone)):
# 初始化 LLMTransformer 模型
llm_transformer = LLMGraphTransformer(llm=llm)
# 转换文件
from langchain_core.documents import Document
text = """
Mr. and Mrs. Dursley, of number four, Privet Drive, were proud to say
that they were perfectly normal, thank you very much. They were the last
people you'd expect to be involved in anything strange or mysterious,
because they just didn't hold with such nonsense.
Mr. Dursley was the director of a firm called Grunnings, which made
drills. He was a big, beefy man with hardly any neck, although he did
have a very large mustache. Mrs. Dursley was thin and blonde and had
nearly twice the usual amount of neck, which came in very useful as she
spent so much of her time craning over garden fences, spying on the
neighbors. The Dursleys had a small son called Dudley and in their
opinion there was no finer boy anywhere.
The Dursleys had everything they wanted, but they also had a secret, and
their greatest fear was that somebody would discover it. They didn't
think they could bear it if anyone found out about the Potters. Mrs.
Potter was Mrs. Dursley's sister, but they hadn't met for several years;
in fact, Mrs. Dursley pretended she didn't have a sister, because her
sister and her good-for-nothing husband were as unDursleyish as it was
possible to be. The Dursleys shuddered to think what the neighbors would
say if the Potters arrived in the street. The Dursleys knew that the
Potters had a small son, too, but they had never even seen him. This boy
was another good reason for keeping the Potters away; they didn't want
Dudley mixing with a child like that.
"""
documents = [Document(page_content=text)]
graph_documents = llm_transformer.convert_to_graph_documents(documents)
print(f"Nodes:{graph_documents[0].nodes}")
print(f"Relationships:{graph_documents[0].relationships}")
Nodes:[Node(id='Mr. Dursley', type='Person'), Node(id='Mrs. Dursley', type='Person'), Node(id='Dudley', type='Person'), Node(id='Privet Drive', type='Location'), Node(id='Grunnings', type='Organization'), Node(id='Mrs. Potter', type='Person'), Node(id='The Potters', type='Family')]Relationships:[Relationship(source=Node(id='Mr. Dursley', type='Person'), target=Node(id='Mrs. Dursley', type='Person'), type='MARRIED_TO'), Relationship(source=Node(id='Mr. Dursley', type='Person'), target=Node(id='Dudley', type='Person'), type='PARENT_OF'), Relationship(source=Node(id='Mrs. Dursley', type='Person'), target=Node(id='Dudley', type='Person'), type='PARENT_OF'), Relationship(source=Node(id='Mr. Dursley', type='Person'), target=Node(id='Grunnings', type='Organization'), type='WORKS_AT'), Relationship(source=Node(id='Mr. Dursley', type='Person'), target=Node(id='Privet Drive', type='Location'), type='LIVES_AT'), Relationship(source=Node(id='Mrs. Dursley', type='Person'), target=Node(id='Privet Drive', type='Location'), type='LIVES_AT'), Relationship(source=Node(id='Mrs. Dursley', type='Person'), target=Node(id='Mrs. Potter', type='Person'), type='SISTER_OF'), Relationship(source=Node(id='The Dursleys', type='Family'), target=Node(id='The Potters', type='Family'), type='WANTS_TO_AVOID')]
正如你所看到的,我们的 llm_transformer 能够从数据中捕捉到相关的实体和关系,无需我们指定任何内容。现在我们需要将这些节点和关系存储在我们的 AuraDB 中。
graph.add_graph_documents(graph_documents,baseEntityLabel=True,include_source=True)
我们现在有了一个填充的图数据库。我们现在可以在我们的在线 AuraDB 实例中检查文档是否已正确上传:
还可以使用以下 Python 函数绘制我们数据库的图形表示:
# 直接显示给定Cypher查询的结果图
default_cypher = "MATCH (s)-[r:!MENTIONS]->(t) RETURN s,r,t LIMIT 50"
def showGraph(cypher: str = default_cypher):
# 创建一个neo4j会话来运行查询
driver = GraphDatabase.driver(
uri = os.environ["NEO4J_URI"],
auth = (os.environ["NEO4J_USERNAME"],
os.environ["NEO4J_PASSWORD"]))
session = driver.session()
widget = GraphWidget(graph = session.run(cypher).graph())
widget.node_label_mapping = 'id'
#display(widget)
return widget
showGraph()
现在我们有了我们的 GraphDB,我们可以添加向量搜索功能来增强检索阶段。为此,我们将需要一个嵌入模型,我将使用 Azure OpenAI 的 text-embedding-ada-002 如下:
from langchain_openai import AzureOpenAIEmbeddings
embeddings = AzureOpenAIEmbeddings(
model="text-embedding-ada-002",
api_key=api_key,
azure_endpoint=azure_endpoint,
openai_api_version=api_version,
)
vector_index = Neo4jVector.from_existing_graph(embeddings,search_type="hybrid",node_label="Document",text_node_properties=["text"],embedding_node_property="embedding")
vector_index 已经可以使用向量相似性方法进行查询了:
query = "Who is Dudley?"
results = vector_index.similarity_search(query, k=1)
print(results[0].page_content)
Mr. and Mrs. Dursley, of number four, Privet Drive, were proud to saythat they were perfectly normal, thank you very much. They were the lastpeople you'd expect to be involved in anything strange or mysterious,because they just didn't hold with such nonsense.Mr. Dursley was the director of a firm called Grunnings, which madedrills. He was a big, beefy man with hardly any neck, although he didhave a very large mustache. Mrs. Dursley was thin and blonde and hadnearly twice the usual amount of neck, which came in very useful as shespent so much of her time craning over garden fences, spying on theneighbors. The Dursleys had a small son called Dudley and in theiropinion there was no finer boy anywhere.The Dursleys had everything they wanted, but they also had a secret, andtheir greatest fear was that somebody would discover it. They didn'tthink they could bear it if anyone found out about the Potters. Mrs.Potter was Mrs. Dursley's sister, but they hadn't met for several years;in fact, Mrs. Dursley pretended she didn't have a sister, because hersister and her good-for-nothing husband were as unDursleyish as it waspossible to be. The Dursleys shuddered to think what the neighbors wouldsay if the Potters arrived in the street. The Dursleys knew that thePotters had a small son, too, but they had never even seen him. This boywas another good reason for keeping the Potters away; they didn't wantDudley mixing with a child like that.
由于没有将文本分割成块,查询会返回整个文档。我们将在下一篇文章中处理较大的文档。
最后一步是从我们的模型中获取实际生成的答案。为此,我们可以利用两种不同的方法:
利用预构建的组件 CypherChain,它利用Neo4j 的 Cypher 查询语言.)与图形数据库进行交互。由于它与 Neo4j 原生集成,因此它与 AuraDB 图形功能进行交互,从而通过理解查询结果实现上下文感知响应。当精度、上下文感知和与 Neo4j 图形功能的直接交互至关重要时,建议使用它。
from langchain.chains import GraphCypherQAChain
chain = GraphCypherQAChain.from_llm(graph=graph, llm=llm, verbose=True)
response = chain.invoke({"query": "What is Mr. Dursley's job?"})
response
> Entering new GraphCypherQAChain chain...
Generated Cypher:
MATCH (p:Person {id: "Mr. Dursley"})-[:WORKS_AT]->(o:Organization) RETURN o.id
Full Context:
[{'o.id': 'Grunnings'}]
> Finished chain.
{'query': "What is Mr. Dursley's job?",
'result': 'Mr. Dursley works at Grunnings.'}
利用经典QA链并使用方法 vector_index.as_retriever(),它可以应用于 LangChain 中的数据存储(无论是vectordb还是graphdb)。
from langchain.chains import RetrievalQA
qa_chain = RetrievalQA.from_chain_type(
llm, retriever=vector_index.as_retriever()
)
result = qa_chain({"query": "What is Mr. Dursley's job?"})
result["result"]
'Mr. Dursley is the director of a firm called Grunnings, which makes drills.'
现在,先不考虑回答的准确性——在文档没有被分割情况下,没有进行基准测试的意义。我们将在下一篇文章中介绍些组件之间的差异,以及它们如何能够带来出色的 RAG 性能。
我们介绍了图数据库的基础知识,以及它们为何在基于 RAG 的应用场景中具有相关性。在下一篇文章中,我们将看到一个基于图的方法的实际实现,并且会利用本篇文章中介绍的所有组件。整个 GitHub 代码也将连下一篇文章一起提供。
https://en.wikipedia.org/wiki/Directed_acyclic_graph?ref=blog.langchain.dev
https://blog.langchain.dev/langgraph/
https://api.python.langchain.com/en/latest/_modules/langchain_experimental/graph_transformers/llm.html#LLMGraphTransformer
https://neo4j.com/docs/getting-started/cypher-intro/#:~:text=Cypher is Neo4j's graph query,how to go get it).
https://www.microsoft.com/en-us/research/blog/graphrag-unlocking-llm-discovery-on-narrative-private-data/
https://python.langchain.com/docs/use_cases/graph/quickstart
53AI,企业落地应用大模型首选服务商
产品:大模型应用平台+智能体定制开发+落地咨询服务
承诺:先做场景POC验证,看到效果再签署服务协议。零风险落地应用大模型,已交付160+中大型企业
2024-03-30
2024-04-26
2024-05-14
2024-04-12
2024-05-10
2024-05-28
2024-07-18
2024-04-25
2024-05-22
2024-04-26