本笔记本介绍了如何开始使用 VDMS 作为向量存储。

>Intel 的 视觉数据管理系统 (VDMS) 是一种存储解决方案，用于高效访问大型"视觉"数据，旨在通过可视化元数据（存储为图形形式）搜索相关视觉数据来实现云规模，并为视觉数据提供机器友好的增强以加快访问速度。VDMS 采用 MIT 许可证。要了解更多关于 VDMS，访问 VDMS wiki.

VDMS 支持：

• * K 近邻搜索
• * 欧几里得距离 (L2) 和内积 (IP)
• * 索引和距离计算的库：FaissFlat（默认）、FaissHNSWFlat、FaissIVFFlat、Flinng、TileDBDense、TileDBSparse
• * 文本、图像和视频的嵌入
• * 向量和元数据搜索

设置

要访问 VDMS 向量存储，你需要安装 langchain-vdms 集成包，并通过公开可用的 Docker 镜像部署 VDMS 服务器。 为简单起见，本笔记本将在本地主机上使用端口 55555 部署 VDMS 服务器。

pip install -qU "langchain-vdms>=0.1.3"
!docker run --no-healthcheck --rm -d -p 55555:55555 --name vdms_vs_test_nb intellabs/vdms:latest
!sleep 5

凭证

你可以使用 VDMS 无需任何凭证。

要启用模型调用的自动追踪，请设置你的 LangSmith API 密钥：

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

初始化

使用 VDMS Client 连接到 VDMS 向量存储，使用 FAISS IndexFlat 索引（默认）和欧几里得距离（默认）作为相似性搜索的距离度量。

# | output: false
# | echo: false

! pip install -qU langchain-huggingface
from langchain_huggingface import HuggingFaceEmbeddings

embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-mpnet-base-v2")

from langchain_vdms.vectorstores import VDMS, VDMS_Client

collection_name = "test_collection_faiss_L2"

vdms_client = VDMS_Client(host="localhost", port=55555)

vector_store = VDMS(
    client=vdms_client,
    embedding=embeddings,
    collection_name=collection_name,
    engine="FaissFlat",
    distance_strategy="L2",
)

管理向量存储

向向量存储添加项

logging.basicConfig()
logging.getLogger("langchain_vdms.vectorstores").setLevel(logging.INFO)

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"},
    id=1,
)

document_2 = Document(
    page_content="The weather forecast for tomorrow is cloudy and overcast, with a high of 62 degrees.",
    metadata={"source": "news"},
    id=2,
)

document_3 = Document(
    page_content="Building an exciting new project with LangChain - come check it out!",
    metadata={"source": "tweet"},
    id=3,
)

document_4 = Document(
    page_content="Robbers broke into the city bank and stole $1 million in cash.",
    metadata={"source": "news"},
    id=4,
)

document_5 = Document(
    page_content="Wow! That was an amazing movie. I can't wait to see it again.",
    metadata={"source": "tweet"},
    id=5,
)

document_6 = Document(
    page_content="Is the new iPhone worth the price? Read this review to find out.",
    metadata={"source": "website"},
    id=6,
)

document_7 = Document(
    page_content="The top 10 soccer players in the world right now.",
    metadata={"source": "website"},
    id=7,
)

document_8 = Document(
    page_content="LangGraph is the best framework for building stateful, agentic applications!",
    metadata={"source": "tweet"},
    id=8,
)

document_9 = Document(
    page_content="The stock market is down 500 points today due to fears of a recession.",
    metadata={"source": "news"},
    id=9,
)

document_10 = Document(
    page_content="I have a bad feeling I am going to get deleted :(",
    metadata={"source": "tweet"},
    id=10,
)

documents = [
    document_1,
    document_2,
    document_3,
    document_4,
    document_5,
    document_6,
    document_7,
    document_8,
    document_9,
    document_10,
]

doc_ids = [str(i) for i in range(1, 11)]
vector_store.add_documents(documents=documents, ids=doc_ids)

['1', '2', '3', '4', '5', '6', '7', '8', '9', '10']

如果多次提供同一个 id， add_documents 不会检查 id 是否唯一。因此，请使用 upsert 在添加之前删除现有的 id 条目。

vector_store.upsert(documents, ids=doc_ids)

{'succeeded': ['1', '2', '3', '4', '5', '6', '7', '8', '9', '10'],
 'failed': []}

更新向量存储中的项

updated_document_1 = Document(
    page_content="I had chocolate chip pancakes and fried eggs for breakfast this morning.",
    metadata={"source": "tweet"},
    id=1,
)

updated_document_2 = Document(
    page_content="The weather forecast for tomorrow is sunny and warm, with a high of 82 degrees.",
    metadata={"source": "news"},
    id=2,
)

vector_store.update_documents(
    ids=doc_ids[:2],
    documents=[updated_document_1, updated_document_2],
    batch_size=2,
)

从向量存储中删除项

vector_store.delete(ids=doc_ids[-1])

True

查询向量存储

一旦您的向量存储已创建并添加了相关文档，您很可能希望在运行链或代理时对其进行查询。

直接查询

执行简单的相似性搜索可以按如下方式进行：

results = vector_store.similarity_search(
    "LangChain provides abstractions to make working with LLMs easy",
    k=2,
    filter={"source": ["==", "tweet"]},
)
for doc in results:
    print(f"* ID={doc.id}: {doc.page_content} [{doc.metadata}]")

INFO:langchain_vdms.vectorstores:VDMS similarity search took 0.0063 seconds

* ID=3: Building an exciting new project with LangChain - come check it out! [{'source': 'tweet'}]
* ID=8: LangGraph is the best framework for building stateful, agentic applications! [{'source': 'tweet'}]

如果您想执行相似性搜索并接收相应的分数，可以运行：

results = vector_store.similarity_search_with_score(
        "Will it be hot tomorrow?", k=1, filter={"source": ["==", "news"]}
)
for doc, score in results:
        print(f"* [SIM={score:3f}] {doc.page_content} [{doc.metadata}]")

INFO:langchain_vdms.vectorstores:VDMS similarity search took 0.0460 seconds

* [SIM=0.753577] The weather forecast for tomorrow is sunny and warm, with a high of 82 degrees. [{'source': 'news'}]

如果您想使用嵌入执行相似性搜索，可以运行：

results = vector_store.similarity_search_by_vector(
    embedding=embeddings.embed_query("I love green eggs and ham!"), k=1
)
for doc in results:
    print(f"* {doc.page_content} [{doc.metadata}]")

INFO:langchain_vdms.vectorstores:VDMS similarity search took 0.0044 seconds

* The weather forecast for tomorrow is sunny and warm, with a high of 82 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，以便在链中更方便地使用。

retriever = vector_store.as_retriever(
        search_type="similarity",
        search_kwargs={"k": 3},
)
results = retriever.invoke("Stealing from the bank is a crime")
for doc in results:
    print(f"* {doc.page_content} [{doc.metadata}]")

INFO:langchain_vdms.vectorstores:VDMS similarity search took 0.0042 seconds

* Robbers broke into the city bank and stole $1 million in cash. [{'source': 'news'}]
* The stock market is down 500 points today due to fears of a recession. [{'source': 'news'}]
* Is the new iPhone worth the price? Read this review to find out. [{'source': 'website'}]

retriever = vector_store.as_retriever(
    search_type="similarity_score_threshold",
    search_kwargs={
        "k": 1,
        "score_threshold": 0.0,  # >= score_threshold
    },
)
results = retriever.invoke("Stealing from the bank is a crime")
for doc in results:
    print(f"* {doc.page_content} [{doc.metadata}]")

INFO:langchain_vdms.vectorstores:VDMS similarity search took 0.0042 seconds

* Robbers broke into the city bank and stole $1 million in cash. [{'source': 'news'}]

retriever = vector_store.as_retriever(
        search_type="mmr",
        search_kwargs={"k": 1, "fetch_k": 10},
)
results = retriever.invoke(
        "Stealing from the bank is a crime", filter={"source": ["==", "news"]}
)
for doc in results:
    print(f"* {doc.page_content} [{doc.metadata}]")

INFO:langchain_vdms.vectorstores:VDMS mmr search took 0.0042 secs

* Robbers broke into the city bank and stole $1 million in cash. [{'source': 'news'}]

删除集合

之前，我们根据其 id删除文档。在这里，由于没有提供 ID，所有文档都被删除。

print("Documents before deletion: ", vector_store.count())

vector_store.delete(collection_name=collection_name)

print("Documents after deletion: ", vector_store.count())

Documents before deletion:  10
Documents after deletion:  0

用于检索增强生成

有关如何使用此向量存储进行检索增强生成 (RAG) 的指南，请参阅以下章节：

• - 检索文档
• * 使用 VDMS 的多模态 RAG
• * 使用 VDMS 的视觉 RAG
• - 使用 LangChain 构建 RAG 应用
• - 智能体 RAG

使用其他引擎进行相似性搜索

VDMS 支持多种用于索引和计算距离的库：FaissFlat（默认）、FaissHNSWFlat、FaissIVFFlat、Flinng、TileDBDense 和 TileDBSparse。 默认情况下，向量存储使用 FaissFlat。下面我们展示一些使用其他引擎的示例。

使用 faiss HNSWFlat 和欧氏距离进行相似性搜索

在这里，我们使用 Faiss IndexHNSWFlat 索引和 L2 作为距离度量将文档添加到 VDMS 进行相似性搜索。我们搜索与查询相关的三个文档（k=3），同时返回文档及其得分。

db_FaissHNSWFlat = VDMS.from_documents(
    documents,
    client=vdms_client,
    ids=doc_ids,
    collection_name="my_collection_FaissHNSWFlat_L2",
    embedding=embeddings,
    engine="FaissHNSWFlat",
    distance_strategy="L2",
)
# Query
k = 3
query = "LangChain provides abstractions to make working with LLMs easy"
docs_with_score = db_FaissHNSWFlat.similarity_search_with_score(query, k=k, filter=None)

for res, score in docs_with_score:
    print(f"* [SIM={score:3f}] {res.page_content} [{res.metadata}]")

INFO:langchain_vdms.vectorstores:Descriptor set my_collection_FaissHNSWFlat_L2 created
INFO:langchain_vdms.vectorstores:VDMS similarity search took 0.1272 seconds

* [SIM=0.716791] Building an exciting new project with LangChain - come check it out! [{'source': 'tweet'}]
* [SIM=0.936718] LangGraph is the best framework for building stateful, agentic applications! [{'source': 'tweet'}]
* [SIM=1.834110] Is the new iPhone worth the price? Read this review to find out. [{'source': 'website'}]

使用 faiss IVFFlat 和内积 (IP) 距离进行相似性搜索

我们使用 Faiss IndexIVFFlat 索引和 IP 作为距离度量将文档添加到 VDMS 进行相似性搜索。我们搜索与查询相关的三个文档（k=3），同时返回文档及其得分。

db_FaissIVFFlat = VDMS.from_documents(
    documents,
        client=vdms_client,
        ids=doc_ids,
        collection_name="my_collection_FaissIVFFlat_IP",
        embedding=embeddings,
        engine="FaissIVFFlat",
        distance_strategy="IP",
)

k = 3
query = "LangChain provides abstractions to make working with LLMs easy"
docs_with_score = db_FaissIVFFlat.similarity_search_with_score(query, k=k, filter=None)
for res, score in docs_with_score:
        print(f"* [SIM={score:3f}] {res.page_content} [{res.metadata}]")

INFO:langchain_vdms.vectorstores:Descriptor set my_collection_FaissIVFFlat_IP created
INFO:langchain_vdms.vectorstores:VDMS similarity search took 0.0052 seconds

* [SIM=0.641605] Building an exciting new project with LangChain - come check it out! [{'source': 'tweet'}]
* [SIM=0.531641] LangGraph is the best framework for building stateful, agentic applications! [{'source': 'tweet'}]
* [SIM=0.082945] Is the new iPhone worth the price? Read this review to find out. [{'source': 'website'}]

使用 FLINNG 和 IP 距离进行相似性搜索

在本节中，我们使用基于过滤器的近邻组识别（FLINNG）索引和 IP 作为距离度量将文档添加到 VDMS 进行相似性搜索。我们搜索与查询相关的三个文档（k=3），同时返回文档及其得分。

db_Flinng = VDMS.from_documents(
    documents,
    client=vdms_client,
    ids=doc_ids,
    collection_name="my_collection_Flinng_IP",
    embedding=embeddings,
    engine="Flinng",
    distance_strategy="IP",
)
# Query
k = 3
query = "LangChain provides abstractions to make working with LLMs easy"
docs_with_score = db_Flinng.similarity_search_with_score(query, k=k, filter=None)
for res, score in docs_with_score:
    print(f"* [SIM={score:3f}] {res.page_content} [{res.metadata}]")

INFO:langchain_vdms.vectorstores:Descriptor set my_collection_Flinng_IP created
INFO:langchain_vdms.vectorstores:VDMS similarity search took 0.0042 seconds

* [SIM=0.000000] I had chocolate chip pancakes and scrambled eggs for breakfast this morning. [{'source': 'tweet'}]
* [SIM=0.000000] I had chocolate chip pancakes and scrambled eggs for breakfast this morning. [{'source': 'tweet'}]
* [SIM=0.000000] I had chocolate chip pancakes and scrambled eggs for breakfast this morning. [{'source': 'tweet'}]

元数据过滤

在处理集合之前先缩小范围可能会很有帮助。

例如，可以使用 get_by_constraints 方法按元数据进行集合过滤。使用字典来过滤元数据。在这里我们检索 langchain_id = "2" 的文档并将其从向量存储中移除。

***NOTE:*** id 作为整数类型的额外元数据生成，而 langchain_id （内部 ID）是每个条目的唯一字符串。

response, response_array = db_FaissIVFFlat.get_by_constraints(
    db_FaissIVFFlat.collection_name,
        limit=1,
        include=["metadata", "embeddings"],
        constraints={"langchain_id": ["==", "2"]},
)

# Delete id=2
db_FaissIVFFlat.delete(collection_name=db_FaissIVFFlat.collection_name, ids=["2"])

print("Deleted entry:")
for doc in response:
        print(f"* ID={doc.id}: {doc.page_content} [{doc.metadata}]")

Deleted entry:
* ID=2: The weather forecast for tomorrow is cloudy and overcast, with a high of 62 degrees. [{'source': 'news'}]

response, response_array = db_FaissIVFFlat.get_by_constraints(
    db_FaissIVFFlat.collection_name,
        include=["metadata"],
)
for doc in response:
        print(f"* ID={doc.id}: {doc.page_content} [{doc.metadata}]")

* ID=10: I have a bad feeling I am going to get deleted :( [{'source': 'tweet'}]
* ID=9: The stock market is down 500 points today due to fears of a recession. [{'source': 'news'}]
* ID=8: LangGraph is the best framework for building stateful, agentic applications! [{'source': 'tweet'}]
* ID=7: The top 10 soccer players in the world right now. [{'source': 'website'}]
* ID=6: Is the new iPhone worth the price? Read this review to find out. [{'source': 'website'}]
* ID=5: Wow! That was an amazing movie. I can't wait to see it again. [{'source': 'tweet'}]
* ID=4: Robbers broke into the city bank and stole $1 million in cash. [{'source': 'news'}]
* ID=3: Building an exciting new project with LangChain - come check it out! [{'source': 'tweet'}]
* ID=1: I had chocolate chip pancakes and scrambled eggs for breakfast this morning. [{'source': 'tweet'}]

这里我们使用 id 来过滤 ID 范围，因为它是一个整数。

response, response_array = db_FaissIVFFlat.get_by_constraints(
    db_FaissIVFFlat.collection_name,
        include=["metadata", "embeddings"],
        constraints={"source": ["==", "news"]},
)
for doc in response:
        print(f"* ID={doc.id}: {doc.page_content} [{doc.metadata}]")

* ID=9: The stock market is down 500 points today due to fears of a recession. [{'source': 'news'}]
* ID=4: Robbers broke into the city bank and stole $1 million in cash. [{'source': 'news'}]

停止 VDMS 服务器

!docker kill vdms_vs_test_nb

vdms_vs_test_nb