Support tsvector columns for Postgres fulltext search

Problem

See https://github.com/prisma/prisma/issues/8950#issuecomment-1067395682 a search runs 1,000x faster with a tsvector column with an index, when compared to the “TEXT” field with an GIN index approach you currently recommend (which does not use the index due to the bug linked)

Suggested solution

Support tsvector columns

Alternatives

None

Additional context

n/a

[joshribakoff-sm]

Here’s the workaround I used for now:

model products {
    indexed_name Unsupported("tsvector")? @default(dbgenerated("''::tsvector")) /// Pre-computed tsvector search index, updated via trigger
}

-- AlterTable
ALTER TABLE "products" ADD COLUMN     "indexed_name" tsvector DEFAULT ''::tsvector;
CREATE INDEX products_indexed_name ON products USING GIN (indexed_name);

const myQuery = `'foo':*`
const query = `SELECT id from products where indexed_name @@ to_tsquery(${myQuery})`; // important, do not select indexed_name!
const res = await prisma.$queryRawUnsafe(query);
// use Prisma to hydrate the IDs
const hydratedMatches = await prisma.products.findMany({
        where: {
          id: {
            in: res.map((row) => row.id),
          },
        },
})

Performance is 🔥 :

josh=# explain analyze SELECT id, name from products where indexed_name @@ to_tsquery('slack:*');
                                                           QUERY PLAN                                                            
---------------------------------------------------------------------------------------------------------------------------------
 Bitmap Heap Scan on products  (cost=16.30..27.27 rows=6 width=64) (actual time=0.050..0.051 rows=0 loops=1)
   Recheck Cond: (indexed_name @@ to_tsquery('slack:*'::text))
   ->  Bitmap Index Scan on products_indexed_name  (cost=0.00..16.30 rows=6 width=0) (actual time=0.048..0.048 rows=0 loops=1)
         Index Cond: (indexed_name @@ to_tsquery('slack:*'::text))
 Planning Time: 1.211 ms
 Execution Time: 0.148 ms
(6 rows)

Compared to what Prisma does, with it’s “preview fulltext search” module. 214x slower.

josh=# explain analyze SELECT id, name from products where name @@ to_tsquery('slack:*');
                                               QUERY PLAN                                               
--------------------------------------------------------------------------------------------------------
 Seq Scan on products  (cost=0.00..363.72 rows=7 width=64) (actual time=21.444..30.549 rows=1 loops=1)
   Filter: (name @@ to_tsquery('slack:*'::text))
   Rows Removed by Filter: 912
 Planning Time: 0.129 ms
 Execution Time: 30.567 ms
(5 rows)

[AzSiAz]

That’s actually one of my proposed solution, using a tsvector stored generated column from one or more column + index

https://github.com/prisma/prisma/issues/8950#issuecomment-982915628