以编程方式使用文档

>Elasticsearch 是一个分布式、RESTful 风格的搜索和分析引擎,能够执行向量搜索和词汇搜索。它构建在 Apache Lucene 库之上。

本笔记本展示了如何使用与 Elasticsearch 向量存储

设置

要使用 Elasticsearch 向量搜索,您必须安装 langchain-elasticsearch package.

pip install -qU langchain-elasticsearch

凭证

有两种主要方式来设置用于:

  1. Elastic Cloud:Elastic Cloud 是一个托管的 Elasticsearch 服务。注册 免费试用.

要连接到不需要 登录凭证(启动启用了安全性的 docker 实例),请将 Elasticsearch URL 和索引名称连同 嵌入对象传递给构造函数。

  1. 本地安装 Elasticsearch:通过本地运行来开始使用 Elasticsearch。最简单的方法是使用官方 Elasticsearch Docker 镜像。请参阅 Elasticsearch Docker 文档 获取更多信息。

本地运行 Elasticsearch

本地运行 Elasticsearch 进行开发和测试的最简单方法是使用 start-local 脚本。此脚本使用 Docker 设置 Elasticsearch(以及可选的 Kibana),只需一条简单的命令。

curl -fsSL https://elastic.co/start-local | sh

这将创建一个 elastic-start-local 文件夹,其中包含配置文件和启动脚本。启动 Elasticsearch:

cd elastic-start-local
./start.sh

Elasticsearch 将在以下地址可用 http://localhost:9200elastic 用户和 API 密钥会自动生成并存储在 .env 文件位于 elastic-start-local folder.

如果只需要 Elasticsearch 而不需要 Kibana,可以使用 --esonly option:

curl -fsSL https://elastic.co/start-local | sh -s -- --esonly

使用身份验证运行

对于生产环境,我们建议启用安全性运行。要使用登录凭证连接,您可以使用参数 es_api_key or es_useres_password.

# | output: false
# | echo: false
from langchain_openai import OpenAIEmbeddings

embeddings = OpenAIEmbeddings(model="text-embedding-3-large")
from langchain_elasticsearch import ElasticsearchStore

elastic_vector_search = ElasticsearchStore(
    es_url="http://localhost:9200",
    index_name="langchain_index",
    embedding=embeddings,
    es_user="elastic",
    es_password="changeme",
)

如何获取默认 "elastic" 用户的密码?

要获取默认 "elastic" 用户的 Elastic Cloud 密码:

  1. 登录 Elastic Cloud 控制台,网址为 cloud.elastic.co
  2. 转到"安全性" > "用户"
  3. 找到 "elastic" 用户并点击"编辑"
  4. 点击"重置密码"
  5. 按照提示重置密码

如何获取 API 密钥?

要获取 API 密钥:

  1. 登录 Elastic Cloud 控制台,网址为 cloud.elastic.co
  2. 打开 Kibana 并转到堆栈管理 > API 密钥
  3. 点击"创建 API 密钥"
  4. 输入 API 密钥名称并点击"创建"
  5. 复制 API 密钥并粘贴到 api_key 参数中

Elastic Cloud

要连接到 Elastic Cloud 上的 Elasticsearch 实例,您可以使用 es_cloud_id 参数或 es_url.

elastic_vector_search = ElasticsearchStore(
    es_cloud_id="<cloud_id>",
    index_name="test_index",
    embedding=embeddings,
    es_user="elastic",
    es_password="changeme",
)

如果您想获得最佳的一流模型调用自动追踪功能,您还可以设置您的 LangSmith API 密钥,方法:取消下方注释:

os.environ["LANGSMITH_API_KEY"] = getpass.getpass("Enter your LangSmith API key: ")
os.environ["LANGSMITH_TRACING"] = "true"

初始化

Elasticsearch 在 localhost:9200 本地运行,使用 Docker。有关从 Elastic Cloud 连接到 Elasticsearch 的更多详细信息,请参阅 使用身份验证进行连接 above.

from langchain_elasticsearch import ElasticsearchStore

vector_store = ElasticsearchStore(
    "langchain-demo", embedding=embeddings, es_url="http://localhost:9201"
)

管理向量存储

向向量存储添加项目

from uuid import uuid4

from langchain_core.documents import Document

document_1 = Document(
    page_content="I had chocolate chip pancakes and scrambled eggs for breakfast this morning.",
    metadata={"source": "tweet"},
)

document_2 = Document(
    page_content="The weather forecast for tomorrow is cloudy and overcast, with a high of 62 degrees.",
    metadata={"source": "news"},
)

document_3 = Document(
    page_content="Building an exciting new project with LangChain - come check it out!",
    metadata={"source": "tweet"},
)

document_4 = Document(
    page_content="Robbers broke into the city bank and stole $1 million in cash.",
    metadata={"source": "news"},
)

document_5 = Document(
    page_content="Wow! That was an amazing movie. I can't wait to see it again.",
    metadata={"source": "tweet"},
)

document_6 = Document(
    page_content="Is the new iPhone worth the price? Read this review to find out.",
    metadata={"source": "website"},
)

document_7 = Document(
    page_content="The top 10 soccer players in the world right now.",
    metadata={"source": "website"},
)

document_8 = Document(
    page_content="LangGraph is the best framework for building stateful, agentic applications!",
    metadata={"source": "tweet"},
)

document_9 = Document(
    page_content="The stock market is down 500 points today due to fears of a recession.",
    metadata={"source": "news"},
)

document_10 = Document(
    page_content="I have a bad feeling I am going to get deleted :(",
    metadata={"source": "tweet"},
)

documents = [
    document_1,
    document_2,
    document_3,
    document_4,
    document_5,
    document_6,
    document_7,
    document_8,
    document_9,
    document_10,
]
uuids = [str(uuid4()) for _ in range(len(documents))]

vector_store.add_documents(documents=documents, ids=uuids)
['21cca03c-9089-42d2-b41c-3d156be2b519',
 'a6ceb967-b552-4802-bb06-c0e95fce386e',
 '3a35fac4-e5f0-493b-bee0-9143b41aedae',
 '176da099-66b1-4d6a-811b-dfdfe0808d30',
 'ecfa1a30-3c97-408b-80c0-5c43d68bf5ff',
 'c0f08baa-e70b-4f83-b387-c6e0a0f36f73',
 '489b2c9c-1925-43e1-bcf0-0fa94cf1cbc4',
 '408c6503-9ba4-49fd-b1cc-95584cd914c5',
 '5248c899-16d5-4377-a9e9-736ca443ad4f',
 'ca182769-c4fc-4e25-8f0a-8dd0a525955c']

从向量存储删除项目

vector_store.delete(ids=[uuids[-1]])
True

查询向量存储

创建向量存储并添加相关文档后,您很可能希望在链或代理运行期间对其进行查询。这些示例还展示了在搜索时如何使用过滤。

直接查询

相似性搜索

执行带元数据过滤的简单相似性搜索可按以下方式进行:

results = vector_store.similarity_search(
    query="LangChain provides abstractions to make working with LLMs easy",
    k=2,
    filter=[{"term": {"metadata.source.keyword": "tweet"}}],
)
for res in results:
    print(f"* {res.page_content} [{res.metadata}]")
* Building an exciting new project with LangChain - come check it out! [{'source': 'tweet'}]
* LangGraph is the best framework for building stateful, agentic applications! [{'source': 'tweet'}]

带评分的相似性搜索

如果您想执行相似性搜索并接收相应的评分,可以运行:

results = vector_store.similarity_search_with_score(
    query="Will it be hot tomorrow",
    k=1,
    filter=[{"term": {"metadata.source.keyword": "news"}}],
)
for doc, score in results:
    print(f"* [SIM={score:3f}] {doc.page_content} [{doc.metadata}]")
* [SIM=0.765887] The weather forecast for tomorrow is cloudy and overcast, with a high of 62 degrees. [{'source': 'news'}]

通过转换为检索器进行查询

您还可以将向量存储转换为检索器,以便在链中更轻松地使用。

retriever = vector_store.as_retriever(
    search_type="similarity_score_threshold", search_kwargs={"score_threshold": 0.2}
)
retriever.invoke("Stealing from the bank is a crime")
[Document(metadata={'source': 'news'}, page_content='Robbers broke into the city bank and stole $1 million in cash.'),
 Document(metadata={'source': 'news'}, page_content='The stock market is down 500 points today due to fears of a recession.'),
 Document(metadata={'source': 'website'}, page_content='Is the new iPhone worth the price? Read this review to find out.'),
 Document(metadata={'source': 'tweet'}, page_content='Building an exciting new project with LangChain - come check it out!')]

距离相似性算法

Elasticsearch 支持以下向量距离相似性算法:

  • - 余弦
  • - 欧几里得
  • - 点积_乘积

余弦相似性算法是默认算法。

您可以通过 similarity 参数指定所需的相似性算法。

NOTE:根据检索策略,相似性算法在查询时无法更改。需要在创建字段索引映射时设置。如果您需要更改相似性算法,需要删除索引并使用正确的距离重新创建_strategy.

db = ElasticsearchStore.from_documents(
    docs,
    embeddings,
    es_url="http://localhost:9200",
    index_name="test",
    distance_strategy="COSINE",
    # distance_strategy="EUCLIDEAN_DISTANCE"
    # distance_strategy="DOT_PRODUCT"
)

检索策略

Elasticsearch 相对于其他纯向量数据库具有很大优势,因为它能够支持多种检索策略。在本笔记本中,我们将配置 ElasticsearchStore 以支持一些最常用的检索策略。

默认情况下, ElasticsearchStore 使用 DenseVectorStrategy (之前称为 ApproxRetrievalStrategy 0.2.0 版本之前)。

DenseVectorStrategy

这将返回与查询向量最相似的 top k 个向量。 k 参数在 ElasticsearchStore 初始化时设置。默认值为 10。

from langchain_elasticsearch import DenseVectorStrategy

db = ElasticsearchStore.from_documents(
    docs,
    embeddings,
    es_url="http://localhost:9200",
    index_name="test",
    strategy=DenseVectorStrategy(),
)

docs = db.similarity_search(
    query="What did the president say about Ketanji Brown Jackson?", k=10
)

示例:使用密集向量和关键词搜索的混合检索

本示例将展示如何配置 ElasticsearchStore 执行混合检索,结合近似语义搜索和基于关键词的搜索。

我们使用 RRF 来平衡来自不同检索方法的两个分数。

要启用混合检索,我们需要设置 hybrid=TrueDenseVectorStrategy constructor.

db = ElasticsearchStore.from_documents(
    docs,
    embeddings,
    es_url="http://localhost:9200",
    index_name="test",
    strategy=DenseVectorStrategy(hybrid=True),
)

启用混合后,执行的查询将是近似语义搜索和基于关键词搜索的组合。

它将使用 rrf(倒数排名融合)来平衡来自不同检索方法的两个分数。

注意:RRF 需要 Elasticsearch 8.9.0 或更高版本。

{
    "retriever": {
        "rrf": {
            "retrievers": [
                {
                    "standard": {
                        "query": {
                            "bool": {
                                "filter": [],
                                "must": [{"match": {"text": {"query": "foo"}}}],
                            }
                        },
                    },
                },
                {
                    "knn": {
                        "field": "vector",
                        "filter": [],
                        "k": 1,
                        "num_candidates": 50,
                        "query_vector": [1.0, ..., 0.0],
                    },
                },
            ]
        }
    }
}

示例:在 Elasticsearch 中使用嵌入模型的稠密向量搜索

本示例将展示如何配置 ElasticsearchStore 使用 Elasticsearch 中部署的嵌入模型进行稠密向量检索。

要使用此功能,请通过_id in DenseVectorStrategy 构造函数指定模型 query_model_id argument.

NOTE:这需要模型已部署并在 Elasticsearch ML 节点上运行。参见 笔记本示例 了解如何使用 eland.

DENSE_SELF_DEPLOYED_INDEX_NAME = "test-dense-self-deployed"

# Note: This does not have an embedding function specified
# Instead, we will use the embedding model deployed in Elasticsearch
db = ElasticsearchStore(
    es_cloud_id="<your cloud id>",
    es_user="elastic",
    es_password="<your password>",
    index_name=DENSE_SELF_DEPLOYED_INDEX_NAME,
    query_field="text_field",
    vector_query_field="vector_query_field.predicted_value",
    strategy=DenseVectorStrategy(model_id="sentence-transformers__all-minilm-l6-v2"),
)

# Setup a Ingest Pipeline to perform the embedding
# of the text field
db.client.ingest.put_pipeline(
    id="test_pipeline",
    processors=[
        {
            "inference": {
                "model_id": "sentence-transformers__all-minilm-l6-v2",
                "field_map": {"query_field": "text_field"},
                "target_field": "vector_query_field",
            }
        }
    ],
)

# creating a new index with the pipeline,
# not relying on langchain to create the index
db.client.indices.create(
    index=DENSE_SELF_DEPLOYED_INDEX_NAME,
    mappings={
        "properties": {
            "text_field": {"type": "text"},
            "vector_query_field": {
                "properties": {
                    "predicted_value": {
                        "type": "dense_vector",
                        "dims": 384,
                        "index": True,
                        "similarity": "l2_norm",
                    }
                }
            },
        }
    },
    settings={"index": {"default_pipeline": "test_pipeline"}},
)

db.from_texts(
    ["hello world"],
    es_cloud_id="<cloud id>",
    es_user="elastic",
    es_password="<cloud password>",
    index_name=DENSE_SELF_DEPLOYED_INDEX_NAME,
    query_field="text_field",
    vector_query_field="vector_query_field.predicted_value",
    strategy=DenseVectorStrategy(model_id="sentence-transformers__all-minilm-l6-v2"),
)

# Perform search
db.similarity_search("hello world", k=10)

SparseVectorStrategy (ELSER)

此策略使用 Elasticsearch 的稀疏向量检索来检索前 k 个结果。目前我们仅支持自己的 "ELSER" 嵌入模型。

NOTE:这需要 ELSER 模型已部署并在 Elasticsearch ML 节点上运行。

要使用此功能,请指定 SparseVectorStrategy (在 0.2.0 版本之前称为 SparseVectorRetrievalStrategy )在 ElasticsearchStore 构造函数中。您需要提供一个模型 ID。

from langchain_elasticsearch import SparseVectorStrategy

# Note that this example doesn't have an embedding function. This is because we infer the tokens at index time and at query time within Elasticsearch.
# This requires the ELSER model to be loaded and running in Elasticsearch.
db = ElasticsearchStore.from_documents(
    docs,
    es_cloud_id="<cloud id>",
    es_user="elastic",
    es_password="<cloud password>",
    index_name="test-elser",
    strategy=SparseVectorStrategy(model_id=".elser_model_2"),
)

db.client.indices.refresh(index="test-elser")

results = db.similarity_search(
    "What did the president say about Ketanji Brown Jackson", k=4
)
print(results[0])

DenseVectorScriptScoreStrategy

此策略使用 Elasticsearch 的脚本评分查询执行精确向量检索(也称为暴力搜索)来检索前 k 个结果。(此策略在 0.2.0 版本之前称为 ExactRetrievalStrategy 。)

要使用此功能,请指定 DenseVectorScriptScoreStrategy in ElasticsearchStore constructor.

from langchain_elasticsearch import SparseVectorStrategy

db = ElasticsearchStore.from_documents(
    docs,
    embeddings,
    es_url="http://localhost:9200",
    index_name="test",
    strategy=DenseVectorScriptScoreStrategy(),
)

BM25Strategy

最后,您可以使用全文关键词搜索。

要使用此功能,请指定 BM25Strategy in ElasticsearchStore constructor.

from langchain_elasticsearch import BM25Strategy

db = ElasticsearchStore.from_documents(
    docs,
    es_url="http://localhost:9200",
    index_name="test",
    strategy=BM25Strategy(),
)

BM25RetrievalStrategy

此策略允许用户使用纯 BM25 执行搜索,而不进行向量搜索。

要使用此功能,请指定 BM25RetrievalStrategy in ElasticsearchStore constructor.

请注意,在下面的示例中未指定嵌入选项,表明搜索是在不使用嵌入的情况下进行的。

from langchain_elasticsearch import ElasticsearchStore

db = ElasticsearchStore(
    es_url="http://localhost:9200",
    index_name="test_index",
    strategy=ElasticsearchStore.BM25RetrievalStrategy(),
)

db.add_texts(
    ["foo", "foo bar", "foo bar baz", "bar", "bar baz", "baz"],
)

results = db.similarity_search(query="foo", k=10)
print(results)

自定义查询

通过 custom_query 搜索参数,您可以调整用于从 Elasticsearch 检索文档的查询。如果您想使用更复杂的查询、支持字段的线性提升,这非常有用。

# Example of a custom query that's just doing a BM25 search on the text field.
def custom_query(query_body: dict, query: str):
    """Custom query to be used in Elasticsearch.
    Args:
        query_body (dict): Elasticsearch query body.
        query (str): Query string.
    Returns:
        dict: Elasticsearch query body.
    """
    print("Query Retriever created by the retrieval strategy:")
    print(query_body)
    print()

    new_query_body = {"query": {"match": {"text": query}}}

    print("Query that's actually used in Elasticsearch:")
    print(new_query_body)
    print()

    return new_query_body


results = db.similarity_search(
    "What did the president say about Ketanji Brown Jackson",
    k=4,
    custom_query=custom_query,
)
print("Results:")
print(results[0])

自定义文档构建器

通过 doc_builder 搜索参数,您可以调整如何使用从 Elasticsearch 检索的数据构建 Document。如果您有不是使用 LangChain 创建的索引,这尤其有用。

from typing import Dict

from langchain_core.documents import Document


def custom_document_builder(hit: Dict) -> Document:
    src = hit.get("_source", {})
    return Document(
        page_content=src.get("content", "Missing content!"),
        metadata={
            "page_number": src.get("page_number", -1),
            "original_filename": src.get("original_filename", "Missing filename!"),
        },
    )


results = db.similarity_search(
    "What did the president say about Ketanji Brown Jackson",
    k=4,
    doc_builder=custom_document_builder,
)
print("Results:")
print(results[0])

用于检索增强生成的用法

有关如何使用此向量存储进行检索增强生成(RAG)的指南,请参见以下部分:

FAQ

问题:我在将文档索引到 Elasticsearch 时遇到超时错误。如何解决?

一个可能的问题是您的文档可能需要更长时间才能索引到 Elasticsearch。ElasticsearchStore 使用 Elasticsearch 的批量 API,其中有多个默认值可以调整以减少超时错误的发生。

当您使用 SparseVectorRetrievalStrategy 时,这也是一个好的做法。

默认值如下:

  • - chunk_size: 500
  • - max_chunk_bytes: 100MB

要调整这些,您可以通过 chunk_sizemax_chunk_bytes 参数传递给 ElasticsearchStore add_texts method.

    vector_store.add_texts(
        texts,
        bulk_kwargs={
            "chunk_size": 50,
            "max_chunk_bytes": 200000000
        }
    )

升级到 ElasticsearchStore

如果您已在基于 langchain 的项目中使用 Elasticsearch,您可能正在使用旧版实现: ElasticVectorSearchElasticKNNSearch 这些现在已弃用。我们引入了一个新的实现,称为 ElasticsearchStore 它更加灵活且更易于使用。本笔记本将引导您完成升级到新实现的过程。

新增功能有哪些?

新实现现在是一个名为 ElasticsearchStore 的类,可通过策略用于近似密集向量、精确密集向量、稀疏向量 (ELSER)、BM25 检索和混合检索。

我正在使用 ElasticKNNSearch

from langchain_elasticsearch import ElasticsearchStore, DenseVectorStrategy

db = ElasticsearchStore(
  es_url="http://localhost:9200",
  index_name="test_index",
  embedding=embedding,
  # if you use the model_id
  # strategy=DenseVectorStrategy(model_id="test_model")
  # if you use hybrid search
  # strategy=DenseVectorStrategy(hybrid=True)
)

我正在使用 ElasticVectorSearch

from langchain_elasticsearch import ElasticsearchStore, DenseVectorScriptScoreStrategy

db = ElasticsearchStore(
  es_url="http://localhost:9200",
  index_name="test_index",
  embedding=embedding,
  strategy=DenseVectorScriptScoreStrategy()
)
db.client.indices.delete(
    index="test-metadata, test-elser, test-basic",
    ignore_unavailable=True,
    allow_no_indices=True,
)

API 参考

有关所有 ElasticSearchStore 功能和配置的详细文档,请参阅 API 参考