interesting points in there about the next pgvector release:
> The upcoming pgvector 0.7.0 release is planning to include functionality that allows for you to use both scalar and binary quantization as part of your indexing strategy. Specifically, pgvector 0.7.0 adds support for indexing 2-byte floats (halfvec) and bit/binary vectors (using the PostgreSQL bit data type). The 0.7.0 will also add support for sparse vectors via sparsevec, but that will be for a future blog post.
> ASIDE: These quantization techniques also let you index vectors that are larger than 2,000 dimensions, which has been a major request for pgvector. halfvec can store up to 4,000 dimensions, bit can store up to 64,000 dimensions, and sparsevec can store up to 1,000 nonzero elements (which means it can store vectors with very high dimensionality).
1 comment
[ 3.0 ms ] story [ 13.0 ms ] thread> The upcoming pgvector 0.7.0 release is planning to include functionality that allows for you to use both scalar and binary quantization as part of your indexing strategy. Specifically, pgvector 0.7.0 adds support for indexing 2-byte floats (halfvec) and bit/binary vectors (using the PostgreSQL bit data type). The 0.7.0 will also add support for sparse vectors via sparsevec, but that will be for a future blog post.
> ASIDE: These quantization techniques also let you index vectors that are larger than 2,000 dimensions, which has been a major request for pgvector. halfvec can store up to 4,000 dimensions, bit can store up to 64,000 dimensions, and sparsevec can store up to 1,000 nonzero elements (which means it can store vectors with very high dimensionality).