Aerospike
Aerospike Vector Search (AVS) 是 Aerospike 資料庫的擴充功能，可在 Aerospike 中儲存的超大型資料集上進行搜尋。這項新服務位於 Aerospike 之外，並建立索引以執行這些搜尋。
本筆記本展示 LangChain Aerospike VectorStore 整合的功能。
安裝 AVS
在使用本筆記本之前，我們需要執行 AVS 執行個體。使用可用的安裝方法之一。
完成後，儲存您的 AVS 執行個體的 IP 位址和埠，以便在本示範稍後使用
AVS_HOST = "<avs-ip>"
AVS_PORT = 5000
安裝相依性
sentence-transformers 相依性較大。此步驟可能需要幾分鐘才能完成。
!pip install --upgrade --quiet aerospike-vector-search==3.0.1 langchain-community sentence-transformers langchain
下載引言資料集
我們將下載約 100,000 則引言的資料集，並使用這些引言的子集進行語意搜尋。
!wget https://github.com/aerospike/aerospike-vector-search-examples/raw/7dfab0fccca0852a511c6803aba46578729694b5/quote-semantic-search/container-volumes/quote-search/data/quotes.csv.tgz
--2024-05-10 17:28:17-- https://github.com/aerospike/aerospike-vector-search-examples/raw/7dfab0fccca0852a511c6803aba46578729694b5/quote-semantic-search/container-volumes/quote-search/data/quotes.csv.tgz
Resolving github.com (github.com)... 140.82.116.4
Connecting to github.com (github.com)|140.82.116.4|:443... connected.
HTTP request sent, awaiting response... 302 Found
Location: https://raw.githubusercontent.com/aerospike/aerospike-vector-search-examples/7dfab0fccca0852a511c6803aba46578729694b5/quote-semantic-search/container-volumes/quote-search/data/quotes.csv.tgz [following]
--2024-05-10 17:28:17-- https://raw.githubusercontent.com/aerospike/aerospike-vector-search-examples/7dfab0fccca0852a511c6803aba46578729694b5/quote-semantic-search/container-volumes/quote-search/data/quotes.csv.tgz
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 185.199.110.133, 185.199.109.133, 185.199.111.133, ...
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|185.199.110.133|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 11597643 (11M) [application/octet-stream]
Saving to: ‘quotes.csv.tgz’
quotes.csv.tgz 100%[===================>] 11.06M 1.94MB/s in 6.1s
2024-05-10 17:28:23 (1.81 MB/s) - ‘quotes.csv.tgz’ saved [11597643/11597643]
將引言載入文件
我們將使用 CSVLoader 文件載入器載入我們的引言資料集。在本例中，lazy_load 傳回迭代器，以更有效率地擷取我們的引言。在本範例中，我們僅載入 5,000 則引言。
import itertools
import os
import tarfile
from langchain_community.document_loaders.csv_loader import CSVLoader
filename = "./quotes.csv"
if not os.path.exists(filename) and os.path.exists(filename + ".tgz"):
# Untar the file
with tarfile.open(filename + ".tgz", "r:gz") as tar:
tar.extractall(path=os.path.dirname(filename))
NUM_QUOTES = 5000
documents = CSVLoader(filename, metadata_columns=["author", "category"]).lazy_load()
documents = list(
itertools.islice(documents, NUM_QUOTES)
) # Allows us to slice an iterator
API 參考：CSVLoader
print(documents[0])
page_content="quote: I'm selfish, impatient and a little insecure. I make mistakes, I am out of control and at times hard to handle. But if you can't handle me at my worst, then you sure as hell don't deserve me at my best." metadata={'source': './quotes.csv', 'row': 0, 'author': 'Marilyn Monroe', 'category': 'attributed-no-source, best, life, love, mistakes, out-of-control, truth, worst'}
建立您的嵌入器
在此步驟中，我們使用 HuggingFaceEmbeddings 和 "all-MiniLM-L6-v2" 句子轉換器模型來嵌入我們的文件，以便我們可以執行向量搜尋。
from aerospike_vector_search.types import VectorDistanceMetric
from langchain_community.embeddings import HuggingFaceEmbeddings
MODEL_DIM = 384
MODEL_DISTANCE_CALC = VectorDistanceMetric.COSINE
embedder = HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2")
API 參考：HuggingFaceEmbeddings
modules.json: 0%| | 0.00/349 [00:00<?, ?B/s]
config_sentence_transformers.json: 0%| | 0.00/116 [00:00<?, ?B/s]
README.md: 0%| | 0.00/10.7k [00:00<?, ?B/s]
sentence_bert_config.json: 0%| | 0.00/53.0 [00:00<?, ?B/s]
/opt/conda/lib/python3.11/site-packages/huggingface_hub/file_download.py:1132: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
warnings.warn(
config.json: 0%| | 0.00/612 [00:00<?, ?B/s]
/opt/conda/lib/python3.11/site-packages/huggingface_hub/file_download.py:1132: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
warnings.warn(
model.safetensors: 0%| | 0.00/90.9M [00:00<?, ?B/s]
tokenizer_config.json: 0%| | 0.00/350 [00:00<?, ?B/s]
vocab.txt: 0%| | 0.00/232k [00:00<?, ?B/s]
tokenizer.json: 0%| | 0.00/466k [00:00<?, ?B/s]
special_tokens_map.json: 0%| | 0.00/112 [00:00<?, ?B/s]
1_Pooling/config.json: 0%| | 0.00/190 [00:00<?, ?B/s]
建立 Aerospike 索引並嵌入文件
在新增文件之前，我們需要在 Aerospike 資料庫中建立索引。在以下範例中，我們使用一些便利程式碼來檢查預期的索引是否已存在。
from aerospike_vector_search import AdminClient, Client, HostPort
from aerospike_vector_search.types import VectorDistanceMetric
from langchain_community.vectorstores import Aerospike
# Here we are using the AVS host and port you configured earlier
seed = HostPort(host=AVS_HOST, port=AVS_PORT)
# The namespace of where to place our vectors. This should match the vector configured in your docstore.conf file.
NAMESPACE = "test"
# The name of our new index.
INDEX_NAME = "quote-miniLM-L6-v2"
# AVS needs to know which metadata key contains our vector when creating the index and inserting documents.
VECTOR_KEY = "vector"
client = Client(seeds=seed)
admin_client = AdminClient(
seeds=seed,
)
index_exists = False
# Check if the index already exists. If not, create it
for index in admin_client.index_list():
if index["id"]["namespace"] == NAMESPACE and index["id"]["name"] == INDEX_NAME:
index_exists = True
print(f"{INDEX_NAME} already exists. Skipping creation")
break
if not index_exists:
print(f"{INDEX_NAME} does not exist. Creating index")
admin_client.index_create(
namespace=NAMESPACE,
name=INDEX_NAME,
vector_field=VECTOR_KEY,
vector_distance_metric=MODEL_DISTANCE_CALC,
dimensions=MODEL_DIM,
index_labels={
"model": "miniLM-L6-v2",
"date": "05/04/2024",
"dim": str(MODEL_DIM),
"distance": "cosine",
},
)
admin_client.close()
docstore = Aerospike.from_documents(
documents,
embedder,
client=client,
namespace=NAMESPACE,
vector_key=VECTOR_KEY,
index_name=INDEX_NAME,
distance_strategy=MODEL_DISTANCE_CALC,
)
API 參考：Aerospike
quote-miniLM-L6-v2 does not exist. Creating index
搜尋文件
現在我們已嵌入向量，我們可以在引言上使用向量搜尋。
query = "A quote about the beauty of the cosmos"
docs = docstore.similarity_search(
query, k=5, index_name=INDEX_NAME, metadata_keys=["_id", "author"]
)
def print_documents(docs):
for i, doc in enumerate(docs):
print("~~~~ Document", i, "~~~~")
print("auto-generated id:", doc.metadata["_id"])
print("author: ", doc.metadata["author"])
print(doc.page_content)
print("~~~~~~~~~~~~~~~~~~~~\n")
print_documents(docs)
~~~~ Document 0 ~~~~
auto-generated id: f53589dd-e3e0-4f55-8214-766ca8dc082f
author: Carl Sagan, Cosmos
quote: The Cosmos is all that is or was or ever will be. Our feeblest contemplations of the Cosmos stir us -- there is a tingling in the spine, a catch in the voice, a faint sensation, as if a distant memory, of falling from a height. We know we are approaching the greatest of mysteries.
~~~~~~~~~~~~~~~~~~~~
~~~~ Document 1 ~~~~
auto-generated id: dde3e5d1-30b7-47b4-aab7-e319d14e1810
author: Elizabeth Gilbert
quote: The love that moves the sun and the other stars.
~~~~~~~~~~~~~~~~~~~~
~~~~ Document 2 ~~~~
auto-generated id: fd56575b-2091-45e7-91c1-9efff2fe5359
author: Renee Ahdieh, The Rose & the Dagger
quote: From the stars, to the stars.
~~~~~~~~~~~~~~~~~~~~
~~~~ Document 3 ~~~~
auto-generated id: 8567ed4e-885b-44a7-b993-e0caf422b3c9
author: Dante Alighieri, Paradiso
quote: Love, that moves the sun and the other stars
~~~~~~~~~~~~~~~~~~~~
~~~~ Document 4 ~~~~
auto-generated id: f868c25e-c54d-48cd-a5a8-14bf402f9ea8
author: Thich Nhat Hanh, Teachings on Love
quote: Through my love for you, I want to express my love for the whole cosmos, the whole of humanity, and all beings. By living with you, I want to learn to love everyone and all species. If I succeed in loving you, I will be able to love everyone and all species on Earth... This is the real message of love.
~~~~~~~~~~~~~~~~~~~~
將其他引言嵌入為文字
我們可以使用 add_texts 新增其他引言。
docstore = Aerospike(
client,
embedder,
NAMESPACE,
index_name=INDEX_NAME,
vector_key=VECTOR_KEY,
distance_strategy=MODEL_DISTANCE_CALC,
)
ids = docstore.add_texts(
[
"quote: Rebellions are built on hope.",
"quote: Logic is the beginning of wisdom, not the end.",
"quote: If wishes were fishes, we’d all cast nets.",
],
metadatas=[
{"author": "Jyn Erso, Rogue One"},
{"author": "Spock, Star Trek"},
{"author": "Frank Herbert, Dune"},
],
)
print("New IDs")
print(ids)
New IDs
['972846bd-87ae-493b-8ba3-a3d023c03948', '8171122e-cbda-4eb7-a711-6625b120893b', '53b54409-ac19-4d90-b518-d7c40bf5ee5d']
使用最大邊際相關性搜尋來搜尋文件
我們可以使用最大邊際相關性搜尋來尋找與我們的查詢相似但不彼此相似的向量。在本範例中，我們使用 as_retriever 建立檢索器物件，但這也可以直接呼叫 docstore.max_marginal_relevance_search 來輕鬆完成。lambda_mult 搜尋引數決定查詢回應的多樣性。0 對應於最大多樣性，而 1 對應於最小多樣性。
query = "A quote about our favorite four-legged pets"
retriever = docstore.as_retriever(
search_type="mmr", search_kwargs={"fetch_k": 20, "lambda_mult": 0.7}
)
matched_docs = retriever.invoke(query)
print_documents(matched_docs)
~~~~ Document 0 ~~~~
auto-generated id: 67d5b23f-b2d2-4872-80ad-5834ea08aa64
author: John Grogan, Marley and Me: Life and Love With the World's Worst Dog
quote: Such short little lives our pets have to spend with us, and they spend most of it waiting for us to come home each day. It is amazing how much love and laughter they bring into our lives and even how much closer we become with each other because of them.
~~~~~~~~~~~~~~~~~~~~
~~~~ Document 1 ~~~~
auto-generated id: a9b28eb0-a21c-45bf-9e60-ab2b80e988d8
author: John Grogan, Marley and Me: Life and Love With the World's Worst Dog
quote: Dogs are great. Bad dogs, if you can really call them that, are perhaps the greatest of them all.
~~~~~~~~~~~~~~~~~~~~
~~~~ Document 2 ~~~~
auto-generated id: ee7434c8-2551-4651-8a22-58514980fb4a
author: Colleen Houck, Tiger's Curse
quote: He then put both hands on the door on either side of my head and leaned in close, pinning me against it. I trembled like a downy rabbit caught in the clutches of a wolf. The wolf came closer. He bent his head and began nuzzling my cheek. The problem was…I wanted the wolf to devour me.
~~~~~~~~~~~~~~~~~~~~
~~~~ Document 3 ~~~~
auto-generated id: 9170804c-a155-473b-ab93-8a561dd48f91
author: Ray Bradbury
quote: Stuff your eyes with wonder," he said, "live as if you'd drop dead in ten seconds. See the world. It's more fantastic than any dream made or paid for in factories. Ask no guarantees, ask for no security, there never was such an animal. And if there were, it would be related to the great sloth which hangs upside down in a tree all day every day, sleeping its life away. To hell with that," he said, "shake the tree and knock the great sloth down on his ass.
~~~~~~~~~~~~~~~~~~~~
使用相關性閾值搜尋文件
另一個實用的功能是具有相關性閾值的相似性搜尋。一般而言，我們只想要與我們的查詢最相似，但也與某個鄰近範圍內的結果。相關性 1 是最相似，而相關性 0 是最不相似。
query = "A quote about stormy weather"
retriever = docstore.as_retriever(
search_type="similarity_score_threshold",
search_kwargs={
"score_threshold": 0.4
}, # A greater value returns items with more relevance
)
matched_docs = retriever.invoke(query)
print_documents(matched_docs)
~~~~ Document 0 ~~~~
auto-generated id: 2c1d6ee1-b742-45ea-bed6-24a1f655c849
author: Roy T. Bennett, The Light in the Heart
quote: Never lose hope. Storms make people stronger and never last forever.
~~~~~~~~~~~~~~~~~~~~
~~~~ Document 1 ~~~~
auto-generated id: 5962c2cf-ffb5-4e03-9257-bdd630b5c7e9
author: Roy T. Bennett, The Light in the Heart
quote: Difficulties and adversities viciously force all their might on us and cause us to fall apart, but they are necessary elements of individual growth and reveal our true potential. We have got to endure and overcome them, and move forward. Never lose hope. Storms make people stronger and never last forever.
~~~~~~~~~~~~~~~~~~~~
~~~~ Document 2 ~~~~
auto-generated id: 3bbcc4ca-de89-4196-9a46-190a50bf6c47
author: Vincent van Gogh, The Letters of Vincent van Gogh
quote: There is peace even in the storm
~~~~~~~~~~~~~~~~~~~~
~~~~ Document 3 ~~~~
auto-generated id: 37d8cf02-fc2f-429d-b2b6-260a05286108
author: Edwin Morgan, A Book of Lives
quote: Valentine WeatherKiss me with rain on your eyelashes,come on, let us sway together,under the trees, and to hell with thunder.
~~~~~~~~~~~~~~~~~~~~
清除
我們需要確保關閉用戶端以釋放資源並清除執行緒。
client.close()
準備好了。開始。搜尋！
既然您已快速掌握 Aerospike Vector Search 的 LangChain 整合，您就掌握了 Aerospike 資料庫和 LangChain 生態系統的力量。祝您建置愉快！