fix: cosine distance

[kemingy]

• refer to feat: Compatibility with pgvector #227

[VoVAllen]

How about PQ and SQ distance?

[kemingy]

How about PQ and SQ distance?

Fixed in the new commit. PTAL

[usamoi]

Can we just use 1.0-cos and sqrt(squared l2) for PostgreSQL operator?

[usamoi]

I suggest add a display_distance function to the G trait and use it for displaying.

[kemingy]

Can we just use 1.0-cos and sqrt(squared l2) for PostgreSQL operator?

I don't have the plan to replace the squared L2 distance.

[gaocegege]

Please resolve the conflicts.

Can we just use 1.0-cos and sqrt(squared l2) for PostgreSQL operator?

It will add more overhead to the compute, but we do not see the benefits. Thus let's keep no sqrt for now.

[kemingy]

Please resolve the conflicts.

fixed

[gaocegege]

PTAL @usamoi @VoVAllen

[usamoi]

Please fix:
https://github.com/tensorchord/pgvecto.rs/blob/a7a9a592c1ef8db0c41ff106b62c85359c10a3b1/README.md?plain=1#L109
https://github.com/tensorchord/pgvecto.rs/blob/a7a9a592c1ef8db0c41ff106b62c85359c10a3b1/README.md?plain=1#L118
https://github.com/tensorchord/pgvecto.rs/blob/a7a9a592c1ef8db0c41ff106b62c85359c10a3b1/crates/service/src/instance/metadata.rs#L21

[VoVAllen]

Can we just use 1.0-cos and sqrt(squared l2) for PostgreSQL operator?

I think this is the better way. We can add sqrt to PostgresSQL operator. If user didn't use L2 instance in the query, no extra cost will be introduced? And we can make it consistent with pgvector at all.

@usamoi Will the cost of sqrt happen when user doesn't need distance in the query?

[usamoi]

@usamoi Will the cost of sqrt happen when user doesn't need distance in the query?

No, there will be no extra cost. What I cares is consistency: we use l2 for opeator name but it's l2s(l2-squared) in fact.

[kemingy]

Cosine distance has to be fixed because it's defined as 1 - cosine_similarity. For the L2 distance, if you choose to impl squared L2, it's also widely accepted. I don't think it's necessary to get the square root.

[usamoi]

We didn't implement Cosine distance. What we implemented before is negative cosine similarity and we take cos for short. It's consistent with users who choose to normalize the vector and use negative dot product.