随着 LLM（大型语言模型）和 RAG（检索增强生成）的兴起，像 Milvus 和 Pinecone 这样的矢量数据库越来越受到关注。传统数据库也通过第三方扩展在矢量搜索支持方面迎头赶上，例如用于 PostgreSQL 的 pgvector 和用于 SQLite 的 sqlite-vss。
在本文中，我将介绍 Vectorlite，这是我为 SQLite 编写的另一个快速且可调的矢量搜索扩展，它刚刚发布了第一个测试版。现在可以使用 pip 安装
pip install vectorlite-py
SQLite 扩展是一个动态库，可以由 SQLite 在运行时加载。下面是在 SQLite CLI shell 中使用 Vectorlite 的示例：
-- Load vectorlite
.load path/to/vectorlite.[so|dll|dylib]
-- shows vectorlite version and build info.
select vectorlite_info();
-- Calculate vector l2(squared) distance
select vector_distance(vector_from_json('[1,2,3]'), vector_from_json('[3,4,5]'), 'l2');
-- Create a virtual table named my_table with one vector column my_embedding with a dimention of 3 using default HNSW parameters.
create virtual table my_table using vectorlite(my_embedding float32[3], hnsw(max_elements=100));
-- Insert vectors into my_table. rowid can be used to relate to a vector's metadata stored elsewhere, e.g. another table.
insert into my_table(rowid, my_embedding) values (0, vector_from_json('[1,2,3]'));
insert into my_table(rowid, my_embedding) values (1, vector_from_json('[2,3,4]'));
insert into my_table(rowid, my_embedding) values (2, vector_from_json('[7,7,7]'));
-- Find 2 approximate nearest neighbors of vector [3,4,5] with distances
select rowid, distance from my_table where knn_search(my_embedding, knn_param(vector_from_json('[3,4,5]'), 2));
-- Find the nearest neighbor of vector [3,4,5] among vectors with rowid 0 and 1. (requires sqlite_version>=3.38)
-- It is called metadata filter in vectorlite, because you could get rowid set beforehand based on vectors' metadata and then perform vector search.
-- Metadata filter is pushed down to the underlying index when traversing the HNSW graph.
select rowid, distance from my_table where knn_search(my_embedding, knn_param(vector_from_json('[3,4,5]'), 1)) and rowid in (0, 1) ;
有关完整的 API 参考，请查看此处。
首先，SQLite 对于 LLM 支持的应用程序来说是很有价值的工具，因为它是轻量级的并且在本地存储数据，使您的数据更加安全。
虽然已经有 sqlite-vss 可以支持 SQLite 的矢量搜索，但它的一些技术决策值得讨论，这就是我编写 Vectorlite 的原因。 sqlite-vss 的作者已经转向另一个矢量搜索扩展项目，并写了一篇文章从他的角度解释了 sqlite-vss 的问题所在。我来分享一下我的。
Sqlite-vss 使用 faiss 进行向量搜索。它是 Meta（facebook）开源的一个很棒的库，提供了广泛的矢量搜索算法。但是，它针对大型数据集上的批量操作进行了优化，这使得 CPU 上的单个向量查询和增量索引速度变慢。但是，SQLite 的可扩展性模型（称为虚拟表）不提供用于批处理操作的 API，仅公开用于一次插入/更新/删除一行的 API。另外sqlite-vss只支持单向量搜索，faiss不擅长。因此，sqlite-vss 无法充分发挥 faiss 的性能。
Vectorlite 使用 hnswlib，它提供了 HNSW 算法的快速实现，并针对增量索引构建和单向量查询进行了优化，与 SQLite 的虚拟表 API 配合良好。
在我的基准测试中，与 sqlite-vss 相比，Vectorlite 在插入向量方面快了 10 倍，在搜索方面快了 2 倍到 40 倍（取决于具有速度精度权衡的 HNSW 参数），这主要是由于向量搜索库的选择不同.
sqlite-vss 错过的另一个重要功能是矢量元数据过滤。在现实场景中，向量总是被标记一些元数据，包括日期、流派、类别、名称、类型、内容等。毫无疑问，在执行向量搜索之前根据其标签过滤向量是向量搜索的必备功能。一个功能向量数据库。
Vectorlite 自首次发布以来就支持矢量元数据过滤器（需要 sqlite 版本 >= 3.38）。请检查这个例子。
sqlite-vss 和 vectorlite 都利用 ANN（近似最近邻）算法进行矢量搜索，这意味着结果不能保证 100% 正确，但搜索速度相对较快。在生产过程中，需要根据您的工作负载、嵌入和特定需求对 ANN 算法的性能进行基准测试和调整。例如，一个人可能需要快速向量查询，但代价是较低的召回率，因为他们正在构建实时语义搜索应用程序，而另一个人可能需要高召回率，但不关心搜索速度，因为他们正在做一些离线分析。
In a typical recommendation system, its vector index is often built and tuned offline using libraries like faiss and hnswlib for acceptable speed and accuracy, and then served in production.
With sqlite-vss, you have to build vector index using it, making performance tuning very difficult. One cannot build the vector index offline with faiss and then serve the index with sqlite-vss. Though you do get to pass faiss strings to it to tune the index, non-default indexes are so slow that they barely work (probably due to the nature of the algorithm or improper use of faiss I mentioned above).
With vectorlite, you have full control over the HNSW paramters to tune the search speed and recall rate. Please check this example and the result.
┏━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━┳━━━━┳━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━┓ ┃ distance ┃ vector ┃ ef ┃ ┃ ef ┃ insert_time ┃ search_time ┃ recall ┃ ┃ type ┃ dimension ┃ construction ┃ M ┃ search ┃ per vector ┃ per query ┃ rate ┃ ┡━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━╇━━━━╇━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━┩ │ l2 │ 256 │ 200 │ 32 │ 10 │ 291.13 us │ 35.70 us │ 31.60% │ │ l2 │ 256 │ 200 │ 32 │ 50 │ 291.13 us │ 99.50 us │ 72.30% │ │ l2 │ 256 │ 200 │ 32 │ 100 │ 291.13 us │ 168.80 us │ 88.60% │ │ l2 │ 256 │ 200 │ 32 │ 150 │ 291.13 us │ 310.53 us │ 95.50% │ │ l2 │ 256 │ 200 │ 48 │ 10 │ 286.92 us │ 37.79 us │ 37.30% │ │ l2 │ 256 │ 200 │ 48 │ 50 │ 286.92 us │ 117.73 us │ 80.30% │ │ l2 │ 256 │ 200 │ 48 │ 100 │ 286.92 us │ 196.01 us │ 93.80% │ │ l2 │ 256 │ 200 │ 48 │ 150 │ 286.92 us │ 259.88 us │ 98.20% │ │ l2 │ 256 │ 200 │ 64 │ 10 │ 285.82 us │ 50.26 us │ 42.60% │ │ l2 │ 256 │ 200 │ 64 │ 50 │ 285.82 us │ 138.83 us │ 84.00% │ │ l2 │ 256 │ 200 │ 64 │ 100 │ 285.82 us │ 253.18 us │ 95.40% │ │ l2 │ 256 │ 200 │ 64 │ 150 │ 285.82 us │ 316.45 us │ 98.70% │ │ l2 │ 1024 │ 200 │ 32 │ 10 │ 1395.02 us │ 158.75 us │ 23.50% │ │ l2 │ 1024 │ 200 │ 32 │ 50 │ 1395.02 us │ 564.27 us │ 60.30% │ │ l2 │ 1024 │ 200 │ 32 │ 100 │ 1395.02 us │ 919.26 us │ 79.30% │ │ l2 │ 1024 │ 200 │ 32 │ 150 │ 1395.02 us │ 1232.40 us │ 88.20% │ │ l2 │ 1024 │ 200 │ 48 │ 10 │ 1489.91 us │ 252.91 us │ 28.50% │ │ l2 │ 1024 │ 200 │ 48 │ 50 │ 1489.91 us │ 848.13 us │ 69.40% │ │ l2 │ 1024 │ 200 │ 48 │ 100 │ 1489.91 us │ 1294.02 us │ 86.80% │ │ l2 │ 1024 │ 200 │ 48 │ 150 │ 1489.91 us │ 1680.97 us │ 94.20% │ │ l2 │ 1024 │ 200 │ 64 │ 10 │ 1412.03 us │ 273.36 us │ 33.30% │ │ l2 │ 1024 │ 200 │ 64 │ 50 │ 1412.03 us │ 899.13 us │ 75.50% │ │ l2 │ 1024 │ 200 │ 64 │ 100 │ 1412.03 us │ 1419.61 us │ 90.10% │ │ l2 │ 1024 │ 200 │ 64 │ 150 │ 1412.03 us │ 1821.85 us │ 96.00% │ │ cosine │ 256 │ 200 │ 32 │ 10 │ 255.22 us │ 28.66 us │ 38.60% │ │ cosine │ 256 │ 200 │ 32 │ 50 │ 255.22 us │ 85.39 us │ 75.90% │ │ cosine │ 256 │ 200 │ 32 │ 100 │ 255.22 us │ 137.31 us │ 91.10% │ │ cosine │ 256 │ 200 │ 32 │ 150 │ 255.22 us │ 190.87 us │ 95.30% │ │ cosine │ 256 │ 200 │ 48 │ 10 │ 259.62 us │ 57.31 us │ 46.60% │ │ cosine │ 256 │ 200 │ 48 │ 50 │ 259.62 us │ 170.54 us │ 84.80% │ │ cosine │ 256 │ 200 │ 48 │ 100 │ 259.62 us │ 221.11 us │ 94.80% │ │ cosine │ 256 │ 200 │ 48 │ 150 │ 259.62 us │ 239.90 us │ 97.90% │ │ cosine │ 256 │ 200 │ 64 │ 10 │ 273.21 us │ 49.34 us │ 48.10% │ │ cosine │ 256 │ 200 │ 64 │ 50 │ 273.21 us │ 139.07 us │ 88.00% │ │ cosine │ 256 │ 200 │ 64 │ 100 │ 273.21 us │ 242.51 us │ 96.30% │ │ cosine │ 256 │ 200 │ 64 │ 150 │ 273.21 us │ 296.21 us │ 98.40% │ │ cosine │ 1024 │ 200 │ 32 │ 10 │ 1192.27 us │ 146.86 us │ 27.40% │ │ cosine │ 1024 │ 200 │ 32 │ 50 │ 1192.27 us │ 451.61 us │ 66.10% │ │ cosine │ 1024 │ 200 │ 32 │ 100 │ 1192.27 us │ 826.40 us │ 83.30% │ │ cosine │ 1024 │ 200 │ 32 │ 150 │ 1192.27 us │ 1199.33 us │ 90.00% │ │ cosine │ 1024 │ 200 │ 48 │ 10 │ 1337.96 us │ 200.14 us │ 33.10% │ │ cosine │ 1024 │ 200 │ 48 │ 50 │ 1337.96 us │ 654.35 us │ 72.60% │ │ cosine │ 1024 │ 200 │ 48 │ 100 │ 1337.96 us │ 1091.57 us │ 88.90% │ │ cosine │ 1024 │ 200 │ 48 │ 150 │ 1337.96 us │ 1429.51 us │ 94.50% │ │ cosine │ 1024 │ 200 │ 64 │ 10 │ 1287.88 us │ 257.67 us │ 38.20% │ │ cosine │ 1024 │ 200 │ 64 │ 50 │ 1287.88 us │ 767.61 us │ 77.00% │ │ cosine │ 1024 │ 200 │ 64 │ 100 │ 1287.88 us │ 1250.36 us │ 92.10% │ │ cosine │ 1024 │ 200 │ 64 │ 150 │ 1287.88 us │ 1699.57 us │ 96.50% │ └──────────┴───────────┴──────────────┴────┴────────┴─────────────┴─────────────┴────────┘
The same benchmark is also run for sqlite-vss using its default index:
┏━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓ ┃ vector dimension ┃ insert_time(per vector) ┃ search_time(per query) ┃ recall_rate ┃ ┡━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩ │ 256 │ 3644.42 us │ 1483.18 us │ 55.00% │ │ 1024 │ 18466.91 us │ 3412.92 us │ 52.20% │ └──────────────────┴─────────────────────────┴────────────────────────┴─────────────┘
Vectorlite is not only much faster, but also offers much better recall rate if properly tuned. Besides, you can also build the index using hnswlib directly and serve the index using vectorlite as vectorlite can be considered a thin wrapper around hnswlib with a SQL API.
The quickest way to get started is to install vectorlite using python.
# Note: vectorlite-py not vectorlite. vectorlite is another project. pip install vectorlite-py apsw numpy
Vectorlite's metadata filter feature requires sqlite>=3.38. Python's builtin sqlite module is usually built with old sqlite versions. So apsw is used here as the SQLite driver, as it provides bindings to newer SQLite. Vectorlite still works with old SQLite versions if metadata filter support is not required.
Below is a minimal example of using vectorlite.
import vectorlite_py
import apsw
import numpy as np
"""
Quick start of using vectorlite extension.
"""
conn = apsw.Connection(':memory:')
conn.enable_load_extension(True) # enable extension loading
conn.load_extension(vectorlite_py.vectorlite_path()) # load vectorlite
cursor = conn.cursor()
# check if vectorlite is loaded
print(cursor.execute('select vectorlite_info()').fetchall())
# Vector distance calculation
for distance_type in ['l2', 'cosine', 'ip']:
v1 = "[1, 2, 3]"
v2 = "[4, 5, 6]"
# Note vector_from_json can be used to convert a JSON string to a vector
distance = cursor.execute(f'select vector_distance(vector_from_json(?), vector_from_json(?), "{distance_type}")', (v1, v2)).fetchone()
print(f'{distance_type} distance between {v1} and {v2} is {distance[0]}')
# generate some test data
DIM = 32 # dimension of the vectors
NUM_ELEMENTS = 10000 # number of vectors
data = np.float32(np.random.random((NUM_ELEMENTS, DIM))) # Only float32 vectors are supported by vectorlite for now
# Create a virtual table using vectorlite using l2 distance (default distance type) and default HNSW parameters
cursor.execute(f'create virtual table my_table using vectorlite(my_embedding float32[{DIM}], hnsw(max_elements={NUM_ELEMENTS}))')
# Vector distance type can be explicitly set to cosine using:
# cursor.execute(f'create virtual table my_table using vectorlite(my_embedding float32[{DIM}] cosine, hnsw(max_elements={NUM_ELEMENTS}))')
# Insert the test data into the virtual table. Note that the rowid MUST be explicitly set when inserting vectors and cannot be auto-generated.
# The rowid is used to uniquely identify a vector and serve as a "foreign key" to relate to the vector's metadata.
# Vectorlite takes vectors in raw bytes, so a numpy vector need to be converted to bytes before inserting into the table.
cursor.executemany('insert into my_table(rowid, my_embedding) values (?, ?)', [(i, data[i].tobytes()) for i in range(NUM_ELEMENTS)])
# Query the virtual table to get the vector at rowid 12345. Note the vector needs to be converted back to json using vector_to_json() to be human-readable.
result = cursor.execute('select vector_to_json(my_embedding) from my_table where rowid = 1234').fetchone()
print(f'vector at rowid 1234: {result[0]}')
# Find 10 approximate nearest neighbors of data[0] and there distances from data[0].
# knn_search() is used to tell vectorlite to do a vector search.
# knn_param(V, K, ef) is used to pass the query vector V, the number of nearest neighbors K to find and an optional ef parameter to tune the performance of the search.
# If ef is not specified, ef defaults to 10. For more info on ef, please check https://github.com/nmslib/hnswlib/blob/v0.8.0/ALGO_PARAMS.md
result = cursor.execute('select rowid, distance from my_table where knn_search(my_embedding, knn_param(?, 10))', [data[0].tobytes()]).fetchall()
print(f'10 nearest neighbors of row 0 is {result}')
# Find 10 approximate nearest neighbors of the first embedding in vectors with rowid within [1000, 2000) using metadata(rowid) filtering.
rowids = ','.join([str(rowid) for rowid in range(1000, 2000)])
result = cursor.execute(f'select rowid, distance from my_table where knn_search(my_embedding, knn_param(?, 10)) and rowid in ({rowids})', [data[0].tobytes()]).fetchall()
print(f'10 nearest neighbors of row 0 in vectors with rowid within [1000, 2000) is {result}')
conn.close()
More examples can be found in examples and integration_test folder.
Vectorlite is created with the hope that it could be the go-to vector search solution for SQLite like pgvector for PostgreSQL.
It is still in early stage. Any suggestions are welcome and appreciated.
以上是Vectorlite 简介：SQLite 的快速且可调的矢量搜索扩展的详细内容。更多信息请关注PHP中文网其他相关文章！
