Binary vectors and Index about pgvecto.rs

[gaocegege]

We wrote a post for the feature: https://blog.pgvecto.rs/my-binary-vector-search-is-better-than-your-fp32-vectors

[VoVAllen]

Thank you for your interest and contributions are greatly appreciated.

For the binary support, there're actually two part, how to store the binary vector inside postgres, and the implementation of binary vector search at the vector search part. And we can reuse the binary vector search part for the binary quantization, like qdrant's latest blog.

I think we can start with the vector search part by supporting binary vector search first, and then try to store it efficiently in postgres. @usamoi can introduce more on the technical details

[AmineDiro]

Thanks @VoVAllen for the response, you're correct I did miss the quantization part of distance computation. The Distance does implement scalar_quantization_distance and scalar_quantization_distance2 that takes two '&[u8]. The issue here is that I don't really think we can reuse those for computing the hamming distance between binary data if we don't have them stored as binary, or at least I don't know how to do so. The issue here is that hamming should be a separate distance from L2 and should be added as a variant to Distance.

pub enum Distance {
    L2,
    Cosine,
    Dot,
   Hamming,
}

But the impl distance is tightly coupled with the Scalar struct. For hamming distance, you shouldn't have to operate on the Scalar data. Maybe having a generic T : Data where Data ressembles RawData trait 🤔 could resolve this ?

[gaocegege]

@usamoi could you please help answer the question?

[usamoi]

Binary vectors require a lot of work. Maybe we should implement binary quantization. It is simpler and saves memory too.

[usamoi]

Completed in #368

[gaocegege]

Here is an example:

CREATE TABLE items (
  id bigserial PRIMARY KEY,
  embedding bvector(3) NOT NULL
);

INSERT INTO items (embedding) VALUES ('[1,0,1]'), ('[0,1,0]');