pgvector typmod lost: vector/halfvec dimensions dropped in plan/apply, causing HNSW index failure ("column does not have dimensions")

[KrisBraun]

Summary

When using pgvector types with dimensions (e.g., halfvec(384) / vector(384)), pgschema appears to drop the typmod during plan/apply. The generated plan SQL creates the column as halfvec (no dimensions). This causes CREATE INDEX ... hnsw to fail with ERROR: column does not have dimensions.

Reproduction

1. Ensure pgvector extension is available:

CREATE EXTENSION IF NOT EXISTS vector;

2. Desired state SQL (schema.sql):

CREATE TABLE activity (
  id bigserial PRIMARY KEY,
  embedding halfvec(384)
);

CREATE INDEX activity_embedding_idx
  ON activity USING hnsw (embedding halfvec_cosine_ops);

3. Run pgschema:

pgschema apply --file schema.sql --schema public --auto-approve \
  --host localhost --db postgres --user postgres

Optional: pgschema plan --file schema.sql --output-sql plan.sql and inspect plan.sql.

Expected

• Column is created as halfvec(384)
• HNSW index is created successfully

Actual

• Plan SQL creates column as halfvec (no dimensions)
• Apply fails with:

ERROR: column does not have dimensions (SQLSTATE 22023)

at the HNSW index creation

Environment

• pgschema 1.7.1 (github.com/pgplex/pgschema)
• PostgreSQL 18.x (Docker)
• pgvector enabled
• macOS (darwin/arm64)

Additional Notes

It looks like typmod for pgvector types is being dropped somewhere in the IR or normalization. Happy to provide full plan SQL output if needed.