Vector type needed for storing OpenAI embeddings

[gablabelle]

Hello,

Still a Prisma newbie and don't know a great deal about DBs TBH. I was wondering if it's possible create a column of the Vector type needed for storing OpenAI embeddings as described in this Supabase blog post but using my current neon database that supports pgvector.

So I've added the extension to my neon database.

create extension vector;

And now would need to do the equivalent of the following via the Prisma Schema:

create table documents (
  id bigserial primary key,
  content text,
  embedding vector (1536)
);

So my guess is that this Vector type isn't supported? What happens then if I have a type supported by the DB (PostgreSQL) but not by Prisma? How do I keep all that in sync?

[janpio]

This should help:

The Unsupported attribute allows you to define fields in the Prisma schema for database types that are not yet supported by Prisma. For example, MySQL's POLYGON type is not currently supported by Prisma, but can now be added to the Prisma schema using the Unsupported("polygon") type.

https://www.prisma.io/docs/concepts/components/prisma-schema/data-model#:~:text=The%20Unsupported%20attribute%20allows%20you,(%22polygon%22)%20type.

[pranavp10]

embedding Unsupported("vector")? is not working with neon databse

[pranavp10]

I am getting this error

[janpio]

The error you are getting is missing.

[pranavp10]

got the answer (#18220 (comment))

[pranavp10]

it should be embedding Unsupported("vector(1536)")?

[gablabelle]

There is a complete prisma schema example using vector available here: https://github.com/vercel/examples/blob/main/storage/postgres-pgvector/prisma/schema.prisma

[pranavp10]

works thanks

[gaufranc]

hey @pranavp10 did you get this working?

if i do the following:
const vector = await prisma.$executeRawSELECT "mycolumn"::vector FROM "mytable" WHERE id = ${id};

i always get 1 as response (1 as type number) no matter if i already saved an embedding inside

[disbelief]

What you need to do is cast the column as ::text and then parse it into a vector (array of floats) yourself:

const result = await client.$queryRawUnsafe<{ mycolumn: string | null }[]>(`
  SELECT "mycolumn"::text FROM "mytable"
  WHERE "id" = '${id}'
  LIMIT 1;
`);
const embeddingString = result[0]?.mycolumn;
let embedding: number[] | null = null;
if (embeddingString !== null || typeof embeddingString !== 'undefined') {
  embedding = embeddingString
    .slice(1, embedding.length - 1) // remove the brackets []
    .split(',')
    .map((v) => parseFloat(v));
}