Templates

A template defines the text that is embedded. The source column is still what full-text search uses.

Use this when a short body needs title or other row context in the vector.

SELECT postvec.set_format(
  'public.docs', 'body',
  E'$title\n\n$body'
);

Or pass format => E'$title\n\n$body' to enable() / adopt().

Expected

set_format() returns void, then status().pending_jobs jumps. Every row is re-enqueued, including NULL-source rows (those clear a stale vector). Until the queue drains, search sees a mix of old-template and new-template vectors. A byte-identical template is a no-op.

Grammar

Token	Meaning
$name	ASCII identifier [A-Za-z_][A-Za-z0-9_]*
${exact column name}	Quoted / awkward names. }} inside braces is one }
$$	A literal $
anything else	Literal

\n is two characters unless the SQL string itself interprets it. Newlines come from SQL string syntax (E'\n' or dollar-quoting).

Limits: non-empty, <= 16 KiB, <= 64 distinct columns. Must reference the source column. Referencing the vector column is a feedback loop and is refused. A chunked template must reference $chunk and omit the source column.

NULL context columns render as ''. A NULL source becomes a NULL vector, with no inference call.

Triggers fire on a change to any referenced column. The template is applied when the worker embeds the row, so a retry uses the current template.

set_format() is bulk maintenance

The call takes SHARE ROW EXCLUSIVE (blocks writes) while it replaces triggers and enqueues every row. Treat it as a maintenance window.

Clear a template with set_format(..., NULL).

migrate(strategy => 'reembed') renders the template. strategy => 'convert' translates existing vectors.

On adopt(format => ...) the template describes how existing vectors were produced. Use backfill => 'all' if that is uncertain.

Refused forms

Newlines come from SQL string syntax

Use E'$title\n\n$body' or dollar-quoting. In a plain SQL string, \n is a backslash and an n.

The template cannot name the vector column

That would create a continuous re-enqueue loop.

Where inference runs

The worker renders the template inside the database, in both modes. The embedding call then goes to the inference host: the PostgreSQL process by default, or postvec-server in remote mode.