Production

• Manage the extension, table, and ANN index through migrations.
• Use validate() after schema deployment.
• Inject a pool configured for TLS, limits, timeouts, and shutdown.
• Match vector dimension, adapter distance, and index operator class.
• Inspect query plans after realistic data volume is loaded.
• Add indexes for high-frequency metadata filters and monitor table bloat.

The adapter's SQL upsert makes repeated physical IDs replace stored content and embeddings, but a changed chunking strategy can produce obsolete IDs. Include cleanup in corpus-version rollouts.

When the adapter creates its own pool, PgVectorClient.close() ends it explicitly; inject a pool in lifecycle-managed services where connection setup and teardown belong to the application.