DashVector:Dynamic vector quantization

Background

Quantization is an optimization method commonly used in vector search technologies. It greatly improves the search performance and reduces the memory space occupied by index files at the cost of acceptable loss of precision, or recall rate.

DashVector supports dynamic vector quantization. To use this feature, you can specify a proper quantization policy when you create a collection. For information about how to create a collection, see Create a collection.

Enable dynamic quantization

Prerequisites

• A cluster is created. For more information, see Create a cluster.

• An API key is obtained. For more information, see Manage API keys.

• The SDK of the latest version is installed. For more information, see Install DashVector SDK.

Sample code

import dashvector
import numpy as np

client = dashvector.Client(
    api_key='YOUR_API_KEY',
    endpoint='YOUR_CLUSTER_ENDPOINT'
)
assert client

# Create a collection with a quantization policy.
ret = client.create(
    'quantize_demo',
    dimension=768,
    extra_params={
        'quantize_type': 'DT_VECTOR_INT8'
    }
)
print(ret)

collection = client.get('quantize_demo')

# Insert a document, which will be automatically quantized based on the quantization policy defined during collection creation.
collection.insert(('1', np.random.rand(768).astype('float32')))

# Obtain the document based on the ID. Note that the obtained vector data is not the original value inserted, but a dequantized approximate value.
doc = collection.fetch('1')

# If you specify to return vector data in a search, the returned vector data is not the original value inserted, but a dequantized approximate value.
docs = collection.query(
    vector=np.random.rand(768).astype('float32'),
    include_vector=True
)

Parameter description

When creating a collection, you can use the quantize_type field of the extra_params: Dict[str, str] parameter to define the quantization policy. The following valid value of quantize_type is supported:

• DT_VECTOR_INT8: Quantizes an FP32 vector to an INT8 vector.

Performance and recall rate

Based on a dataset of one million 768-dimensional vectors

• DashVector cluster specification: P.large

• Metric: cosine

• TopK: 100

References