VectorChord extensions instead of PGVector

[qdrddr]

Most of the current SotA embedding models are 3k, 4k, or even 8k dimensions. With PGVector you are limited to use with embedding models.

I'd like to benefit from VectorChord's PG extensions that provide superior search speed than PGvector and support 65535 dimensions, whereas PGVector supports only 2000.

Note VectorChord also supports HybridSearch (BM25), data and index quantization and multi-vectors.

https://docs.vectorchord.ai/faqs/benchmark.html
https://docs.vectorchord.ai/faqs/comparison-pgvector.html#vector-dimensions
https://docs.vectorchord.ai/use-case/hybrid-search.html
https://docs.vectorchord.ai/usage/quantization.html
https://docs.vectorchord.ai/vectorchord/getting-started/vectorchord-suite.html#how-to-use-the-vectorchord-suite

[adsharma]

This is currently blocked on mac CI issues, which has nothing to do with VectorChord. Stay tuned or please chime in if you know what's wrong.

[adsharma]

ok - that issue is fixed now. Could you please look into the build failures here?

• avx512f, avx512bw are not enabled. Are there any specific flags I need to pass?
• aarch64 builds failed with: non_null_from_ref unstable feature not enabled

Note that we use manylinux containers to build the binaries. These use an old glibc and an updated toolchain for max compatibility across linux distros. But they may lag in terms of features supported.