Skip to main content

Binary quantization

Binary quantization (BQ), is a technique used to compress vectors. In Weaviate, it can be used to reduce the size of the in-memory HNSW index or the disk-based flat index.

For HNSW, BQ can decrease its memory footprint and thus improve performance and reduce resource requirements as well as costs. For the flat index, BQ can reduce the size of the index on disk, which can improve performance.

What is binary quantization?

Binary quantization compresses vectors by reducing each dimension to a single bit, either 0 or 1.

In other words, a n-dimensional vector composed of n floating point numbers is compressed to a n-dimensional vector composed of n bits.

This will reduce the size of the vector by a factor of 32 (from 32 bits per float to 1 bit per dimension).

Model suitability

BQ is a relatively simple algorithm, but can perform well in the right circumstances. It is particularly suitable for high-dimensional vectors, where even with BQ, the vector can retain a high degree of information.

We suggest using BQ for vectors that have been designed for, or been shown to perform well with, binary quantization. Anecdotally, we have seen encouraging recall with Cohere's V3 models (e.g. embed-multilingual-v3.0 or embed-english-v3.0), and OpenAI's ada-002 and larger text-embedding-3 models work well with BQ enabled.

Lossiness

BQ is a lossy compression technique, as the original floating point numbers are quantized a bit.

Weaviate compensates for this by overfetching vectors from the index, and then rescoring the vectors in the uncompressed space. In practice, we find that this compensates quite well for the lossiness of BQ.

Configure BQ

This example creates a collection with binary quantization (BQ) enabled, using default settings.

from weaviate.classes.config import Configure, DataType, Property

# Client instantiation not shown
collection_name = "BQExampleCollection"

client.collections.create(
name=collection_name,
# Other configuration not shown
vector_index_config=Configure.VectorIndex.hnsw(
quantizer=Configure.VectorIndex.Quantizer.bq()
),
)

client.close()

Explain the code

This will create a collection with BQ enabled, using the default settings.

With BQ, the compression begins immediately, as there is no need to wait for a training set to be reached.

Customize BQ

Some BQ parameters are configurable. An important one is rescore_limit, which is the minimum number of vectors to be fetched from the index before the rescore phase is triggered.

from weaviate.classes.config import Configure, DataType, Property

from weaviate.collections.classes.config import BQConfig

client = weaviate.connect_to_local()

print(client.get_meta())
print(weaviate.__version__)


# Client instantiation not shown
collection_name = "BQExampleCollection"

client.collections.create(
name=collection_name,
# Other configuration not shown
vector_index_config=Configure.VectorIndex.hnsw(
quantizer=Configure.VectorIndex.Quantizer.bq(
rescore_limit=150
)
),
)

client.close()

Questions and feedback

If you have any questions or feedback, please let us know on our forum. For example, you can: