Implement casting for Vector to HalfVec

[A2va]

This will enable some usage, as disccused in this issue

CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(3072));
CREATE INDEX ON items USING hnsw ((embedding::halfvec(3072)) halfvec_cosine_ops);

-- and query with
SELECT * FROM items ORDER BY embedding::halfvec(3072) <=> '[1,2,3]' LIMIT 5;

Taken this snippet

let mut conn = establish_connection();
let embed = HalfVector::from_f32_slice(&[1.0, 2.0, 5.0]);
let neighbors = items::table
    .order(
        items::embedding
            .cast::<Nullable<pgvector::sql_types::HalfVector>>()
            .l2_distance(embed.clone()),
    )
    .load::<Item>(&mut conn)
    .unwrap();

It produces the following query:
SELECT "items"."id", "items_diesel"."embedding" FROM "items" ORDER BY CAST("items_diesel"."embedding" AS halfvec) <-> $1 LIMIT $2 -- binds: [HalfVector([1.0, 2.0, 5.0]), 5]
But I actually don't now if not providing the dimension have an impact ?

Second question on the binary vector, for the example in the following query present in the docs:
SELECT * FROM items ORDER BY binary_quantize(embedding)::bit(3) <~> binary_quantize('[1,-2,3]') LIMIT 5;
Is the cast necessary? The documentation mentioned that binary_quantize already returns a bit vector.

[ankane]

Hi @A2va, thanks for the PR. Casts need to include dimensions in both cases for Postgres to use a vector index, so I'm not sure it makes sense to add without that.

[A2va]

Ok, I will see what I can do.

[A2va]

Since it's not possible to use the diesel bultin cast function, I have added a cast_to_halfvec on VectorExpressionMethods

[ankane]

Thanks @A2va. I think it'd be better to submit a generic casting function to Diesel for this (as casting with a type modifier isn't specific to pgvector or Postgres).