One solution could be integrating duckdb in a similar way as pgvector. You need to map duckdb storage to Postgres storage and reuse duckdb query processor. I believe it's the fastest way to get Postgres to have competitive columnstores.