Google Vertex AI 向量搜索

本笔记本展示了如何使用与 Google Cloud Vertex AI Vector Search 向量数据库相关的功能。

Google Vertex AI Vector Search, formerly known as Vertex AI Matching Engine, provides the industry's leading high-scale low latency vector database. These vector databases are commonly referred to as vector similarity-matching or an approximate nearest neighbor (ANN) service.

注意：Langchain API 需要一个已创建的端点和部署的索引。创建索引可能需要长达一个小时的时间。

To see how to create an index refer to the section Create Index and deploy it to an Endpoint
If you already have an index deployed , skip to Create VectorStore from texts

创建索引并将其部署到端点

• 本节演示如何创建一个新索引并将其部署到端点。

# TODO : Set values as per your requirements
# Project and Storage Constants
PROJECT_ID = "<my_project_id>"
REGION = "<my_region>"
BUCKET = "<my_gcs_bucket>"
BUCKET_URI = f"gs://{BUCKET}"

# The number of dimensions for the textembedding-gecko@003 is 768
# If other embedder is used, the dimensions would probably need to change.
DIMENSIONS = 768

# Index Constants
DISPLAY_NAME = "<my_matching_engine_index_id>"
DEPLOYED_INDEX_ID = "<my_matching_engine_endpoint_id>"

# Create a bucket.
! gsutil mb -l $REGION -p $PROJECT_ID $BUCKET_URI

使用 VertexAIEmbeddings 作为嵌入模型

from google.cloud import aiplatform
from langchain_google_vertexai import VertexAIEmbeddings

API 参考：VertexAIEmbeddings

aiplatform.init(project=PROJECT_ID, location=REGION, staging_bucket=BUCKET_URI)

embedding_model = VertexAIEmbeddings(model_name="text-embedding-005")

创建一个空的索引

注意： 在创建索引时，您需要从“BATCH_UPDATE”或“STREAM_UPDATE”中指定一个“index_update_method”

A batch index is for when you want to update your index in a batch, with data which has been stored over a set amount of time, like systems which are processed weekly or monthly. A streaming index is when you want index data to be updated as new data is added to your datastore, for instance, if you have a bookstore and want to show new inventory online as soon as possible. Which type you choose is important, since setup and requirements are different.

请参阅官方文档以获取有关配置索引的更多详细信息

# NOTE : This operation can take upto 30 seconds
my_index = aiplatform.MatchingEngineIndex.create_tree_ah_index(
    display_name=DISPLAY_NAME,
    dimensions=DIMENSIONS,
    approximate_neighbors_count=150,
    distance_measure_type="DOT_PRODUCT_DISTANCE",
    index_update_method="STREAM_UPDATE",  # allowed values BATCH_UPDATE , STREAM_UPDATE
)

创建一个端点

# Create an endpoint
my_index_endpoint = aiplatform.MatchingEngineIndexEndpoint.create(
    display_name=f"{DISPLAY_NAME}-endpoint", public_endpoint_enabled=True
)

部署索引到端点

# NOTE : This operation can take upto 20 minutes
my_index_endpoint = my_index_endpoint.deploy_index(
    index=my_index, deployed_index_id=DEPLOYED_INDEX_ID
)

my_index_endpoint.deployed_indexes

从文本创建向量存储

注意：如果您已有索引和端点，可以使用以下代码加载它们。

# TODO : replace 1234567890123456789 with your acutial index ID
my_index = aiplatform.MatchingEngineIndex("1234567890123456789")

# TODO : replace 1234567890123456789 with your acutial endpoint ID
my_index_endpoint = aiplatform.MatchingEngineIndexEndpoint("1234567890123456789")

from langchain_google_vertexai import (
    VectorSearchVectorStore,
    VectorSearchVectorStoreDatastore,
)

API 参考：VectorSearchVectorStore | VectorSearchVectorStoreDatastore

创建简单的向量存储（不带过滤器）

# Input texts
texts = [
    "The cat sat on",
    "the mat.",
    "I like to",
    "eat pizza for",
    "dinner.",
    "The sun sets",
    "in the west.",
]

# Create a Vector Store
vector_store = VectorSearchVectorStore.from_components(
    project_id=PROJECT_ID,
    region=REGION,
    gcs_bucket_name=BUCKET,
    index_id=my_index.name,
    endpoint_id=my_index_endpoint.name,
    embedding=embedding_model,
    stream_update=True,
)

# Add vectors and mapped text chunks to your vectore store
vector_store.add_texts(texts=texts)

可选：您还可以创建向量并将块存储在数据存储中

# NOTE : This operation can take upto 20 mins
vector_store = VectorSearchVectorStoreDatastore.from_components(
    project_id=PROJECT_ID,
    region=REGION,
    index_id=my_index.name,
    endpoint_id=my_index_endpoint.name,
    embedding=embedding_model,
    stream_update=True,
)

vector_store.add_texts(texts=texts, is_complete_overwrite=True)

# Try running a simialarity search
vector_store.similarity_search("pizza")

使用元数据过滤器创建向量存储

# Input text with metadata
record_data = [
    {
        "description": "A versatile pair of dark-wash denim jeans."
        "Made from durable cotton with a classic straight-leg cut, these jeans"
        " transition easily from casual days to dressier occasions.",
        "price": 65.00,
        "color": "blue",
        "season": ["fall", "winter", "spring"],
    },
    {
        "description": "A lightweight linen button-down shirt in a crisp white."
        " Perfect for keeping cool with breathable fabric and a relaxed fit.",
        "price": 34.99,
        "color": "white",
        "season": ["summer", "spring"],
    },
    {
        "description": "A soft, chunky knit sweater in a vibrant forest green. "
        "The oversized fit and cozy wool blend make this ideal for staying warm "
        "when the temperature drops.",
        "price": 89.99,
        "color": "green",
        "season": ["fall", "winter"],
    },
    {
        "description": "A classic crewneck t-shirt in a soft, heathered blue. "
        "Made from comfortable cotton jersey, this t-shirt is a wardrobe essential "
        "that works for every season.",
        "price": 19.99,
        "color": "blue",
        "season": ["fall", "winter", "summer", "spring"],
    },
    {
        "description": "A flowing midi-skirt in a delicate floral print. "
        "Lightweight and airy, this skirt adds a touch of feminine style "
        "to warmer days.",
        "price": 45.00,
        "color": "white",
        "season": ["spring", "summer"],
    },
]

# Parse and prepare input data

texts = []
metadatas = []
for record in record_data:
    record = record.copy()
    page_content = record.pop("description")
    texts.append(page_content)
    if isinstance(page_content, str):
        metadata = {**record}
        metadatas.append(metadata)

# Inspect metadatas
metadatas

# NOTE : This operation can take more than 20 mins
vector_store = VectorSearchVectorStore.from_components(
    project_id=PROJECT_ID,
    region=REGION,
    gcs_bucket_name=BUCKET,
    index_id=my_index.name,
    endpoint_id=my_index_endpoint.name,
    embedding=embedding_model,
)

vector_store.add_texts(texts=texts, metadatas=metadatas, is_complete_overwrite=True)

from google.cloud.aiplatform.matching_engine.matching_engine_index_endpoint import (
    Namespace,
    NumericNamespace,
)

# Try running a simple similarity search

# Below code should return 5 results
vector_store.similarity_search("shirt", k=5)

# Try running a similarity search with text filter
filters = [Namespace(name="season", allow_tokens=["spring"])]

# Below code should return 4 results now
vector_store.similarity_search("shirt", k=5, filter=filters)

# Try running a similarity search with combination of text and numeric filter
filters = [Namespace(name="season", allow_tokens=["spring"])]
numeric_filters = [NumericNamespace(name="price", value_float=40.0, op="LESS")]

# Below code should return 2 results now
vector_store.similarity_search(
    "shirt", k=5, filter=filters, numeric_filter=numeric_filters
)

使用向量存储作为检索器

# Initialize the vectore_store as retriever
retriever = vector_store.as_retriever()

# perform simple similarity search on retriever
retriever.invoke("What are my options in breathable fabric?")

# Try running a similarity search with text filter
filters = [Namespace(name="season", allow_tokens=["spring"])]

retriever.search_kwargs = {"filter": filters}

# perform similarity search with filters on retriever
retriever.invoke("What are my options in breathable fabric?")

# Try running a similarity search with combination of text and numeric filter
filters = [Namespace(name="season", allow_tokens=["spring"])]
numeric_filters = [NumericNamespace(name="price", value_float=40.0, op="LESS")]


retriever.search_kwargs = {"filter": filters, "numeric_filter": numeric_filters}

retriever.invoke("What are my options in breathable fabric?")

在问答链中使用检索器的过滤器

from langchain_google_vertexai import VertexAI

llm = VertexAI(model_name="gemini-pro")

API 参考：VertexAI

from langchain.chains import RetrievalQA

filters = [Namespace(name="season", allow_tokens=["spring"])]
numeric_filters = [NumericNamespace(name="price", value_float=40.0, op="LESS")]

retriever.search_kwargs = {"k": 2, "filter": filters, "numeric_filter": numeric_filters}

retrieval_qa = RetrievalQA.from_chain_type(
    llm=llm,
    chain_type="stuff",
    retriever=retriever,
    return_source_documents=True,
)

question = "What are my options in breathable fabric?"
response = retrieval_qa({"query": question})
print(f"{response['result']}")
print("REFERENCES")
print(f"{response['source_documents']}")

API 参考：RetrievalQA

读取、分块、向量化并索引PDF文件

!pip install pypdf

from langchain_community.document_loaders import PyPDFLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter

API 参考：PyPDFLoader | RecursiveCharacterTextSplitter

loader = PyPDFLoader("https://arxiv.org/pdf/1706.03762.pdf")
pages = loader.load()

text_splitter = RecursiveCharacterTextSplitter(
    # Set a really small chunk size, just to show.
    chunk_size=1000,
    chunk_overlap=20,
    length_function=len,
    is_separator_regex=False,
)
doc_splits = text_splitter.split_documents(pages)

texts = [doc.page_content for doc in doc_splits]
metadatas = [doc.metadata for doc in doc_splits]

texts[0]

# Inspect Metadata of 1st page
metadatas[0]

vector_store = VectorSearchVectorStore.from_components(
    project_id=PROJECT_ID,
    region=REGION,
    gcs_bucket_name=BUCKET,
    index_id=my_index.name,
    endpoint_id=my_index_endpoint.name,
    embedding=embedding_model,
)

vector_store.add_texts(texts=texts, metadatas=metadatas, is_complete_overwrite=True)

vector_store = VectorSearchVectorStore.from_components(
    project_id=PROJECT_ID,
    region=REGION,
    gcs_bucket_name=BUCKET,
    index_id=my_index.name,
    endpoint_id=my_index_endpoint.name,
    embedding=embedding_model,
)

混合搜索

向量搜索支持混合搜索，这是信息检索（IR）中一种流行的架构模式，它结合了语义搜索和关键词搜索（也称为基于标记的搜索）。通过混合搜索，开发者可以充分利用这两种方法的优点，从而有效提高搜索质量。 点击 这里 了解更多信息。

为了使用混合搜索，我们需要拟合一个稀疏嵌入向量生成器，并在向量搜索集成之外处理嵌入。稀疏嵌入向量生成器的一个例子是 sklearn 的 TfidfVectorizer，但也可以使用其他技术，例如 BM25。

# Define some sample data
texts = [
    "The cat sat on",
    "the mat.",
    "I like to",
    "eat pizza for",
    "dinner.",
    "The sun sets",
    "in the west.",
]

# optional IDs
ids = ["i_" + str(i + 1) for i in range(len(texts))]

# optional metadata
metadatas = [{"my_metadata": i} for i in range(len(texts))]

from sklearn.feature_extraction.text import TfidfVectorizer

# Fit the TFIDF Vectorizer (This is usually done on a very large corpus of data to make sure that word statistics generalize well on new data)
vectorizer = TfidfVectorizer()
vectorizer.fit(texts)

# Utility function to transform text into a TF-IDF Sparse Vector
def get_sparse_embedding(tfidf_vectorizer, text):
    tfidf_vector = tfidf_vectorizer.transform([text])
    values = []
    dims = []
    for i, tfidf_value in enumerate(tfidf_vector.data):
        values.append(float(tfidf_value))
        dims.append(int(tfidf_vector.indices[i]))
    return {"values": values, "dimensions": dims}

# semantic (dense) embeddings
embeddings = embedding_model.embed_documents(texts)
# tfidf (sparse) embeddings
sparse_embeddings = [get_sparse_embedding(vectorizer, x) for x in texts]

sparse_embeddings[0]

# Add the dense and sparse embeddings in Vector Search

vector_store.add_texts_with_embeddings(
    texts=texts,
    embeddings=embeddings,
    sparse_embeddings=sparse_embeddings,
    ids=ids,
    metadatas=metadatas,
)

# Run hybrid search
query = "the cat"
embedding = embedding_model.embed_query(query)
sparse_embedding = get_sparse_embedding(vectorizer, query)

vector_store.similarity_search_by_vector_with_score(
    embedding=embedding,
    sparse_embedding=sparse_embedding,
    k=5,
    rrf_ranking_alpha=0.7,  # 0.7 weight to dense and 0.3 weight to sparse
)

相关

• 向量存储 概念指南
• 向量存储 操操作指南