Changelog category (leave one):

• Experimental Feature

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

Implement QBit data type that stores vectors in bit-sliced format and L2DistanceTransposed function that allows approximate vector search where precision-speed trade-off is controlled by a parameter.

Closes: #77088

Documentation entry for user-facing changes

• Documentation is written (mandatory for new features)

QBit type

The QBit data type provides efficient vector storage for approximate similarity search. Under the hood, QBit stores data as a tuple of FixedStrings. Each FixedString column stores one bit position from all vector elements. For example:

• Column 1 stores the 1st bit from every element of every vector
• Column 2 stores the 2nd bit from every element of every vector
• And so on...

Within each column, each row contains the bits from one vector's elements at that bit position. This allows reading only the most significant bits for faster approximate calculations, or more bits for higher accuracy.

Key features:

• Vectors are stored at full precision while allowing fine-grained quantization at search time
• Choose how many bits to read during search – fewer bits for faster queries, more bits for higher accuracy

Creating QBit

column_name QBit(element_type, dimension)

Where:

• element_type: BFloat16, Float32, or Float64
• dimension: number of elements in each vector

Example:

CREATE TABLE test (id UInt32, vec QBit(Float64, 3)) ENGINE = Memory;
INSERT INTO test VALUES (1, [0, 1, 2]);
SELECT * FROM test;

┌─id─┬─vec─────┐
│  1 │ [0,1,2] │
└────┴─────────┘

QBit subcolumns

Access individual bit planes using .N syntax where N is the bit position:

• BFloat16: 16 subcolumns (.1 to .16)
• Float32: 32 subcolumns (.1 to .32)
• Float64: 64 subcolumns (.1 to .64)

Vector search functions

L2DistanceTransposed(vector1, vector2, p) - Calculates approximate Euclidean distance between vectors

• vector1: QBit vector
• vector2: Array vector
• p: Number of bits to use (1 to element bit-width) - controls precision vs speed. The data will be read up to precision p, saving I/O.

Example:

-- Quantize Float64 to 16 bits
SELECT L2DistanceTransposed(vec, array(1.0, 2.0, 3.0), 16) FROM qbit;

┌─L2DistanceTr⋯., 3.], 16)─┐
│       1.7320508075688772 │
└──────────────────────────┘