CrateDB 集成 - 以编程方式使用文档

> [CrateDB] 是一个分布式且可扩展的 SQL 数据库，用于存储和 > 近实时地分析海量数据，即使使用复杂 > 查询。它与 PostgreSQL 兼容，基于 Lucene，并继承 > 自 Elasticsearch。

安装和设置

### 设置 CrateDB 有两种方法可以快速开始使用 CrateDB。或者，选择其他 [CrateDB 安装选项]。

#### 在本地计算机上启动 CrateDB 示例：运行一个禁用安全性的单节点 CrateDB 实例，使用 Docker 或 Podman。这不推荐用于生产环境。

docker run --name=cratedb --rm \
  --publish=4200:4200 --publish=5432:5432 --env=CRATE_HEAP_SIZE=2g \
  crate:latest -Cdiscovery.type=single-node

#### 在 CrateDB Cloud 上部署集群 [CrateDB Cloud] 是一项托管式 CrateDB 服务。注册 [免费试用][CrateDB Cloud Console]。

### 安装客户端安装最新版本的 [langchain-cratedb] 包以及本教程所需的其他几个包。

pip install -U langchain-cratedb langchain-openai unstructured

uv add langchain-cratedb langchain-openai unstructured

## 文档有关 CrateDB 包装器的更详细说明，请参阅 [将 LangChain 与 CrateDB 结合使用]。另请参阅 [CrateDB 的所有功能] 以了解 CrateDB 提供的其他功能。

## 功能适用于 LangChain 的 CrateDB 适配器提供了将 CrateDB 用作向量存储的 API，文档加载器和聊天消息存储。

### 向量存储围绕 CrateDB 的向量存储功能使用 FLOAT_VECTOR 和 KNN_MATCH 进行相似性搜索和其他目的。另请参阅 [CrateDBVectorStore 教程]。

确保已配置有效的 OpenAI API 密钥。

from langchain_community.document_loaders import UnstructuredURLLoader
from langchain_cratedb import CrateDBVectorStore
from langchain_openai import OpenAIEmbeddings
from langchain.text_splitter import CharacterTextSplitter

loader = UnstructuredURLLoader(urls=["https://github.com/langchain-ai/langchain/raw/refs/tags/langchain-core==0.3.28/docs/docs/how_to/state_of_the_union.txt"])
documents = loader.load()
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
docs = text_splitter.split_documents(documents)

embeddings = OpenAIEmbeddings()

# Connect to a self-managed CrateDB instance on localhost.
CONNECTION_STRING = "crate://?schema=testdrive"

store = CrateDBVectorStore.from_documents(
    documents=docs,
    embedding=embeddings,
    collection_name="state_of_the_union",
    connection=CONNECTION_STRING,
)

query = "What did the president say about Ketanji Brown Jackson"
docs_with_score = store.similarity_search_with_score(query)

### 文档加载器使用基于 SQLAlchemy 的文档加载器从 CrateDB 数据库表加载文档， CrateDBLoader。另请参阅 [CrateDBLoader 教程]。

要在应用程序中使用文档加载器：

from langchain_community.utilities import SQLDatabase
from langchain_cratedb import CrateDBLoader

# Connect to a self-managed CrateDB instance on localhost.
CONNECTION_STRING = "crate://?schema=testdrive"

db = SQLDatabase(engine=sa.create_engine(CONNECTION_STRING))

loader = CrateDBLoader(
    'SELECT * FROM sys.summits LIMIT 42',
    db=db,
)
documents = loader.load()

### 聊天消息历史使用 CrateDB 作为聊天消息的存储。另请参阅 [CrateDBChatMessageHistory 教程]。

要在应用程序中使用聊天消息历史：

from langchain_cratedb import CrateDBChatMessageHistory

# Connect to a self-managed CrateDB instance on localhost.
CONNECTION_STRING = "crate://?schema=testdrive"

message_history = CrateDBChatMessageHistory(
    session_id="test-session",
    connection=CONNECTION_STRING,
)

message_history.add_user_message("hi!")

### 完整缓存 The standard / full cache avoids invoking the LLM when the supplied 当 prompt 与之前遇到的完全相同时。另请参阅 [CrateDBCache 示例]。

要在应用程序中使用完整缓存：

from langchain.globals import set_llm_cache
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_cratedb import CrateDBCache

# Configure cache.
engine = sa.create_engine("crate://crate@localhost:4200/?schema=testdrive")
set_llm_cache(CrateDBCache(engine))

# Invoke LLM conversation.
llm = ChatOpenAI(
    model_name="gpt-5.5",
    temperature=0.7,
)
print()
print("Asking with full cache:")
answer = llm.invoke("What is the answer to everything?")
print(answer.content)

语义缓存

语义缓存允许用户根据语义相似性检索缓存的 prompt，即用户输入与之前缓存的输入之间的相似性。它还避免了在不需要时调用 LLM。另请参阅 [CrateDBSemanticCache 示例]。

要在应用程序中使用语义缓存：

from langchain.globals import set_llm_cache
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_cratedb import CrateDBSemanticCache

# Configure embeddings.
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")

# Configure cache.
engine = sa.create_engine("crate://crate@localhost:4200/?schema=testdrive")
set_llm_cache(
    CrateDBSemanticCache(
        embedding=embeddings,
        connection=engine,
        search_threshold=1.0,
    )
)

# Invoke LLM conversation.
llm = ChatOpenAI(model_name="gpt-5.5")
print()
print("Asking with semantic cache:")
answer = llm.invoke("What is the answer to everything?")
print(answer.content)

[all features of CrateDB]: https://cratedb.com/docs/guide/feature/ [CrateDB]: https://cratedb.com/database [CrateDB Cloud]: https://cratedb.com/database/cloud [CrateDB Cloud Console]: https://console.cratedb.cloud/?utm_source=langchain&utm_content=documentation [CrateDB installation options]: https://cratedb.com/docs/guide/install/ [CrateDBCache Example]: https://github.com/crate/langchain-cratedb/blob/main/examples/basic/cache.py [CrateDBSemanticCache Example]: https://github.com/crate/langchain-cratedb/blob/main/examples/basic/cache.py [CrateDBChatMessageHistory Tutorial]: https://github.com/crate/cratedb-examples/blob/main/topic/machine-learning/llm-langchain/conversational_memory.ipynb [CrateDBLoader Tutorial]: https://github.com/crate/cratedb-examples/blob/main/topic/machine-learning/llm-langchain/document_loader.ipynb [CrateDBVectorStore Tutorial]: https://github.com/crate/cratedb-examples/blob/main/topic/machine-learning/llm-langchain/vector_search.ipynb [langchain-cratedb]: https://pypi.org/project/langchain-cratedb/ [using LangChain with CrateDB]: https://cratedb.com/docs/guide/integrate/langchain/