> ## Documentation Index > Fetch the complete documentation index at: https://langchain-docs-zh.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # Gemini Enterprise Agent Platform reranker 集成 > 使用 LangChain Python 集成 Gemini Enterprise Agent 平台重新排序文档转换器。 > [排名API](https://cloud.google.com/generative-ai-app-builder/docs/ranking)是[App Builder in Agent Studio on Gemini Enterprise Agent Platform](https://cloud.google.com/generative-ai-app-builder/docs/builder-apis)中的独立API之一。它通过评估文档与查询的相关性对文档列表进行重新排序。相较于仅基于语义相似度的嵌入技术，排名API能提供文档回答具体查询的精确得分。该API可用于提升检索初始候选文档后的搜索结果质量。 > 排名API是无状态的，因此在调用API之前无需对文档进行索引。您只需传入查询和文档即可。这使得该API特别适合从任何文档检索器中重排序文档。 > 如需更多信息，请参考[对文档进行评分和重排序](https://cloud.google.com/generative-ai-app-builder/docs/ranking)。 ```python theme={null} pip install -qU langchain langchain-google-community langchain-google-community[vertexaisearch] langchain-google-vertexai langchain-chroma langchain-text-splitters beautifulsoup4 ``` ### 配置 ```python theme={null} PROJECT_ID = "" REGION = "" RANKING_LOCATION_ID = "global" # @param {type:"string"} # Initialize GCP project for Gemini Enterprise Agent Platform from google.cloud import aiplatform aiplatform.init(project=PROJECT_ID, location=REGION) ``` ### 加载数据并准备 本示例将使用[Google Wiki页面](https://en.wikipedia.org/wiki/Google)演示Vertex排名API的工作原理。 我们采用标准的`load -> split -> embed data`流水线。 嵌入通过[Vertex Embeddings API](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/embeddings/get-text-embeddings#supported_models)模型 - `textembedding-gecko@003`生成 > [!WARNING] > > `langchain-community`包已不再维护。从`langchain_community`导入的示例可能过时或无法使用。请谨慎使用。 ```python theme={null} from langchain_chroma import Chroma from langchain_community.document_loaders import WebBaseLoader from langchain_google_vertexai import VertexAIEmbeddings from langchain_text_splitters import RecursiveCharacterTextSplitter vectordb = None # Load wiki page loader = WebBaseLoader("https://en.wikipedia.org/wiki/Google") data = loader.load() # Split doc into chunks text_splitter = RecursiveCharacterTextSplitter(chunk_size=800, chunk_overlap=5) splits = text_splitter.split_documents(data) print(f"Your {len(data)} documents have been split into {len(splits)} chunks") if vectordb is not None: # delete existing vectordb if it already exists vectordb.delete_collection() embedding = VertexAIEmbeddings(model_name="textembedding-gecko@003") vectordb = Chroma.from_documents(documents=splits, embedding=embedding) ``` ```text theme={null} Your 1 documents have been split into 266 chunks ``` ```python theme={null} import pandas as pd from langchain_classic.retrievers.contextual_compression import ContextualCompressionRetriever from langchain_google_community.vertex_rank import VertexAIRank # Instantiate the VertexAIReranker with the SDK manager reranker = VertexAIRank( project_id=PROJECT_ID, location_id=RANKING_LOCATION_ID, ranking_config="default_ranking_config", title_field="source", top_n=5, ) basic_retriever = vectordb.as_retriever(search_kwargs={"k": 5}) # fetch top 5 documents # Create the ContextualCompressionRetriever with the VertexAIRanker as a Reranker retriever_with_reranker = ContextualCompressionRetriever( base_compressor=reranker, base_retriever=basic_retriever ) ``` ### 测试Vertex排名API 让我们用相同的查询对`basic_retriever`和`retriever_with_reranker`进行检索，并比较结果。 排名API从`basic_retriever`获取输入并传递给排名API。 排名API用于提升排序质量，并为每个记录确定与查询的相关性得分。 您可以看到未排序文档与排序后文档的差异。排名API将语义相关性最高的文档移至大语言模型上下文窗口顶部，从而帮助其生成更优的推理答案。 ```python theme={null} import pandas as pd # Use the basic_retriever and the retriever_with_reranker to get relevant documents query = "how did the name google originate?" retrieved_docs = basic_retriever.invoke(query) reranked_docs = retriever_with_reranker.invoke(query) # Create two lists of results for unranked and ranked docs unranked_docs_content = [docs.page_content for docs in retrieved_docs] ranked_docs_content = [docs.page_content for docs in reranked_docs] # Create a comparison DataFrame using the padded lists comparison_df = pd.DataFrame( { "Unranked Documents": unranked_docs_content, "Ranked Documents": ranked_docs_content, } ) comparison_df ``` ```html theme={null}
Unranked Documents Ranked Documents
0 ^ a b Brin, Sergey; Page, Lawrence (1998). "The anatomy of a large-scale hypertextual Web search engine" (PDF). Computer Networks and ISDN Systems. 30 (1–7): 107–117. CiteSeerX 10.1.1.115.5930. doi:10.1016/S0169-7552(98)00110-X. ISSN 0169-7552. S2CID 7587743. Archived (PDF) from the original on September 27, 2015. Retrieved April 7, 2019.\n\n^ "About: RankDex". Archived from the original on January 20, 2012. Retrieved September 29, 2010., RankDex\n\n^ "Method for node ranking in a linked database". Google Patents. Archived from the original on October 15, 2015. Retrieved October 19, 2015.\n\n^ Koller, David (January 2004). "Origin of the name "Google"". Stanford University. Archived from the original on June 27, 2012. The name "Google" originated from a misspelling of "googol",[211]\[212] which refers to the number represented by a 1 followed by one-hundred zeros. Page and Brin write in their original paper on PageRank:[33] "We chose our system name, Google, because it is a common spelling of googol, or 10100[,] and fits well with our goal of building very large-scale search engines." Having found its way increasingly into everyday language, the verb "google" was added to the Merriam Webster Collegiate Dictionary and the Oxford English Dictionary in 2006, meaning "to use the Google search engine to obtain information on the Internet."[213]\[214] Google's mission statement, from the outset, was "to organize the world's information and make it universally accessible and useful",[215] and its unofficial
1 Eventually, they changed the name to Google; the name of the search engine was a misspelling of the word googol,[21]\[36]\[37] a very large number written 10100 (1 followed by 100 zeros), picked to signify that the search engine was intended to provide large quantities of information.[38] Eventually, they changed the name to Google; the name of the search engine was a misspelling of the word googol,[21]\[36]\[37] a very large number written 10100 (1 followed by 100 zeros), picked to signify that the search engine was intended to provide large quantities of information.[38]
2 The name "Google" originated from a misspelling of "googol",[211]\[212] which refers to the number represented by a 1 followed by one-hundred zeros. Page and Brin write in their original paper on PageRank:[33] "We chose our system name, Google, because it is a common spelling of googol, or 10100[,] and fits well with our goal of building very large-scale search engines." Having found its way increasingly into everyday language, the verb "google" was added to the Merriam Webster Collegiate Dictionary and the Oxford English Dictionary in 2006, meaning "to use the Google search engine to obtain information on the Internet."[213]\[214] Google's mission statement, from the outset, was "to organize the world's information and make it universally accessible and useful",[215] and its unofficial ^ Meijer, Bart (January 3, 2019). "Google shifted $23 billion to tax haven Bermuda in 2017: filing". Reuters. Archived from the original on January 3, 2019. Retrieved January 3, 2019. Google moved 19.9 billion euros ($22.7 billion) through a Dutch shell company to Bermuda in 2017, as part of an arrangement that allows it to reduce its foreign tax bill\n\n^ Hamburger, Tom; Gold, Matea (April 13, 2014). "Google, once disdainful of lobbying, now a master of Washington influence". The Washington Post. Archived from the original on October 27, 2017. Retrieved August 22, 2017.\n\n^ Koller, David (January 2004). "Origin of the name, "Google."". Stanford University. Archived from the original on June 27, 2012. Retrieved May 28, 2006.
3 ^ Meijer, Bart (January 3, 2019). "Google shifted $23 billion to tax haven Bermuda in 2017: filing". Reuters. Archived from the original on January 3, 2019. Retrieved January 3, 2019. Google moved 19.9 billion euros ($22.7 billion) through a Dutch shell company to Bermuda in 2017, as part of an arrangement that allows it to reduce its foreign tax bill\n\n^ Hamburger, Tom; Gold, Matea (April 13, 2014). "Google, once disdainful of lobbying, now a master of Washington influence". The Washington Post. Archived from the original on October 27, 2017. Retrieved August 22, 2017.\n\n^ Koller, David (January 2004). "Origin of the name, "Google."". Stanford University. Archived from the original on June 27, 2012. Retrieved May 28, 2006. ^ a b Brin, Sergey; Page, Lawrence (1998). "The anatomy of a large-scale hypertextual Web search engine" (PDF). Computer Networks and ISDN Systems. 30 (1–7): 107–117. CiteSeerX 10.1.1.115.5930. doi:10.1016/S0169-7552(98)00110-X. ISSN 0169-7552. S2CID 7587743. Archived (PDF) from the original on September 27, 2015. Retrieved April 7, 2019.\n\n^ "About: RankDex". Archived from the original on January 20, 2012. Retrieved September 29, 2010., RankDex\n\n^ "Method for node ranking in a linked database". Google Patents. Archived from the original on October 15, 2015. Retrieved October 19, 2015.\n\n^ Koller, David (January 2004). "Origin of the name "Google"". Stanford University. Archived from the original on June 27, 2012.
4 ^ Swant, Marty. "The World's Valuable Brands". Forbes. Archived from the original on October 18, 2020. Retrieved January 19, 2022.\n\n^ "Best Global Brands". Interbrand. Archived from the original on February 1, 2022. Retrieved March 7, 2011.\n\n^ a b c d "How we started and where we are today – Google". about.google. Archived from the original on April 22, 2020. Retrieved April 24, 2021.\n\n^ Brezina, Corona (2013). Sergey Brin, Larry Page, Eric Schmidt, and Google (1st ed.). New York: Rosen Publishing Group. p. 18. ISBN 978-1-4488-6911-4. LCCN 2011039480.\n\n^ a b c "Our history in depth". Google Company. Archived from the original on April 1, 2012. Retrieved July 15, 2017. ^ Swant, Marty. "The World's Valuable Brands". Forbes. Archived from the original on October 18, 2020. Retrieved January 19, 2022.\n\n^ "Best Global Brands". Interbrand. Archived from the original on February 1, 2022. Retrieved March 7, 2011.\n\n^ a b c d "How we started and where we are today – Google". about.google. Archived from the original on April 22, 2020. Retrieved April 24, 2021.\n\n^ Brezina, Corona (2013). Sergey Brin, Larry Page, Eric Schmidt, and Google (1st ed.). New York: Rosen Publishing Group. p. 18. ISBN 978-1-4488-6911-4. LCCN 2011039480.\n\n^ a b c "Our history in depth". Google Company. Archived from the original on April 1, 2012. Retrieved July 15, 2017.
``` 让我们检查部分重排序后的文档。我们观察到检索器仍返回与LangChain类型相关的[文档](https://reference.langchain.com/python/langchain-core/documents/base/Document)，但作为元数据字段，我们还接收了排名API的`relevance_score`。 ```python theme={null} for i in range(2): print(f"Document {i}") print(reranked_docs[i]) print("----------------------------------------------------------\n") ``` ```html theme={null} ``` ```python theme={null} Document 0 page_content='The name "Google" originated from a misspelling of "googol",[211]\[212] which refers to the number represented by a 1 followed by one-hundred zeros. Page and Brin write in their original paper on PageRank:[33] "We chose our system name, Google, because it is a common spelling of googol, or 10100[,] and fits well with our goal of building very large-scale search engines." Having found its way increasingly into everyday language, the verb "google" was added to the Merriam Webster Collegiate Dictionary and the Oxford English Dictionary in 2006, meaning "to use the Google search engine to obtain information on the Internet."[213]\[214] Google\'s mission statement, from the outset, was "to organize the world\'s information and make it universally accessible and useful",[215] and its unofficial' metadata={'id': '2', 'relevance_score': 0.9800000190734863, 'source': 'https://en.wikipedia.org/wiki/Google'} ---------------------------------------------------------- Document 1 page_content='Eventually, they changed the name to Google; the name of the search engine was a misspelling of the word googol,[21]\[36]\[37] a very large number written 10100 (1 followed by 100 zeros), picked to signify that the search engine was intended to provide large quantities of information.[38]' metadata={'id': '1', 'relevance_score': 0.75, 'source': 'https://en.wikipedia.org/wiki/Google'} ---------------------------------------------------------- ``` ### 整合应用 本示例展示了一个完整的RAG链，包含使用Vertex排名API进行重排序的简单提示词模板。 ```python theme={null} from langchain_classic.chains import LLMChain from langchain_core.documents import Document from langchain_core.prompts import PromptTemplate from langchain_core.runnables import RunnableParallel, RunnablePassthrough from langchain_google_vertexai import VertexAI llm = VertexAI(model_name="gemini-2.5-pro") # Instantiate the VertexAIReranker with the SDK manager reranker = VertexAIRank( project_id=PROJECT_ID, location_id=RANKING_LOCATION_ID, ranking_config="default_ranking_config", title_field="source", # metadata field key from your existing documents top_n=5, ) # value of k can be set to a higher value as well for tweaking performance # eg: # of docs: basic_retriever(100) -> reranker(5) basic_retriever = vectordb.as_retriever(search_kwargs={"k": 5}) # fetch top 5 documents # Create the ContextualCompressionRetriever with the VertexAIRanker as a Reranker retriever_with_reranker = ContextualCompressionRetriever( base_compressor=reranker, base_retriever=basic_retriever ) template = """ {context} Question: {query} Don't give information outside the context or repeat your findings. Answer: """ prompt = PromptTemplate.from_template(template) reranker_setup_and_retrieval = RunnableParallel( {"context": retriever_with_reranker, "query": RunnablePassthrough()} ) chain = reranker_setup_and_retrieval | prompt | llm ``` ```html theme={null} ``` ```python theme={null} query = "how did the name google originate?" ``` ```html theme={null} ``` ```python theme={null} chain.invoke(query) ``` ```html theme={null} ``` ```text theme={null} 'The name "Google" originated as a misspelling of the word "googol," a mathematical term for the number 1 followed by 100 zeros. Larry Page and Sergey Brin, the founders of Google, chose the name because it reflected their goal of building a search engine that could handle massive amounts of information. \n' ``` *** > 官方原文：[src/oss/python/integrations/document\_transformers/google\_cloud\_vertexai\_rerank.mdx](https://github.com/langchain-ai/docs/blob/main/src/oss/python/integrations/document_transformers/google_cloud_vertexai_rerank.mdx) · 非官方翻译，译文由 AI 生成，以英文原文为准。 This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.