pqlite 0.1.1

Fast and Light Approximate Nearest Neighbor Search Database integrated with the Jina Ecosystem

PQLite

PQLite is an Approximate Nearest Neighbor Search (ANNS) library integrated with the Jina ecosystem. The PQLite class partitions the data into cells at index time, and instantiates a "sub-indexer" in each cell. Search is performed aggregating results retrieved from cells.

This indexer is recommended to be used when an application requires search with filters applied on Document tags. The filtering query language is based on MongoDB's query and projection operators. We currently support a subset of those selectors. The tags filters can be combined with $and and $or:

• $eq - Equal to (number, string)
• $ne - Not equal to (number, string)
• $gt - Greater than (number)
• $gte - Greater than or equal to (number)
• $lt - Less than (number)
• $lte - Less than or equal to (number)

For example, we want to search for a product with a price no more than 50$.

index.search(query, filter={"price": {"$lte": 50}})

More example filter expresses

• A Nike shoes with white color

{
  "brand": {"$eq": "Nike"},
  "category": {"$eq": "Shoes"},
  "color": {"$eq": "White"}
}

Or

{
  "$and":
    {
      "brand": {"$eq": "Nike"},
      "category": {"$eq": "Shoes"},
      "color": {"$eq": "White"}
    }
}

• A Nike shoes or price less than 100$

{
  "$or":
    {
      "brand": {"$eq": "Nike"},
      "price": {"$lt": 100}
    }
}

WARNING

• PQLite is still in the very early stages of development. APIs can and will change (now is the time to make suggestions!). Important features are missing. Documentation is sparse.
• PQLite contains code that must be compiled to be used. The build is prepared in setup.py. Users only need to pip install . from the root directory.

Quick Start

Setup

$ git clone https://github.com/jina-ai/pqlite.git \
  && cd pqlite \
  && pip install -e .

How to use?

1. Create a new pqlite

import random
import numpy as np
from jina import Document, DocumentArray
from pqlite import PQLite

N = 10000 # number of data points
Nq = 10 # number of query data
D = 128 # dimentionality / number of features

# the column schema: (name:str, dtype:type, create_index: bool)
pqlite = PQLite(dim=D, columns=[('price', float)], data_path='./workspace_data')

Note that this will create a folder ./workspace_data where indexed data will be stored. If there is already a folder with this name and the code presented here is not working remove that folder.

2. Add new data

X = np.random.random((N, D)).astype(np.float32)  # 10,000 128-dim vectors to be indexed
docs = DocumentArray(
    [
        Document(id=f'{i}', embedding=X[i], tags={'price': random.random()})
        for i in range(N)
    ]
)
pqlite.index(docs)

3. Search with filtering

Xq = np.random.random((Nq, D)).astype(np.float32)  # a 128-dim query vector
query = DocumentArray([Document(embedding=Xq[i]) for i in range(Nq)])

# without filtering
pqlite.search(query, limit=10)

print(f'the result without filtering:')
for i, q in enumerate(query):
    print(f'query [{i}]:')
    for m in q.matches:
        print(f'\t{m.id} ({m.scores["euclidean"].value})')

# with filtering
pqlite.search(query, filter={"price": {"$lte": 50}}, limit=10)
print(f'the result with filtering:')
for i, q in enumerate(query):
    print(f'query [{i}]:')
    for m in q.matches:
        print(f'\t{m.id} {m.scores["euclidean"].value} (price={m.tags["x"]})')

4. Update data

Xn = np.random.random((10, D)).astype(np.float32)  # 10,000 128-dim vectors to be indexed
docs = DocumentArray(
    [
        Document(id=f'{i}', embedding=Xn[i], tags={'price': random.random()})
        for i in range(10)
    ]
)
pqlite.update(docs)

5. Delete data

pqlite.delete(['1', '2'])

Benchmark

One can run executor/benchmark.py to get a quick performance overview.

Stored data	Indexing time	Query size=1	Query size=8	Query size=64
10000	2.970	0.002	0.013	0.100
100000	76.474	0.011	0.078	0.649
500000	467.936	0.046	0.356	2.823
1000000	1025.506	0.091	0.695	5.778

Research foundations of PQLite

• Xor Filters Faster and Smaller Than Bloom Filters
• CVPR20 Tutorial Billion-scale Approximate Nearest Neighbor Search
• XOR-Quantization Fast top-K Cosine Similarity Search through XOR-Friendly Binary Quantization on GPUs
• NeurIPS21 Challenge Billion-Scale Approximate Nearest Neighbor Search Challenge NeurIPS'21 competition track
• PAMI 2011 Product quantization for nearest neighbor search
• CVPR 2016 Efficient Indexing of Billion-Scale Datasets of Deep Descriptors
• NIPs 2017 Multiscale Quantization for Fast Similarity Search
• NIPs 2018 Non-metric Similarity Graphs for Maximum Inner Product Search
• ACMMM 2018 Reconfigurable Inverted Index code
• ECCV 2018 Revisiting the Inverted Indices for Billion-Scale Approximate Nearest Neighbors
• CVPR 2019 Unsupervised Neural Quantization for Compressed-Domain Similarity Search
• ICML 2019 Learning to Route in Similarity Graphs
• ICML 2020 Graph-based Nearest Neighbor Search: From Practice to Theory