Production boundaries

Provision explicit resources

Use stable, unique labels and index names for each managed graph. Match vector dimensions to the embedding model and size capacity, resizeCoefficient, and scalarKind for the expected corpus. Memgraph vector indexes use READ_UNCOMMITTED visibility, while text search uses its own index snapshot; do not assume cross-index snapshot isolation during concurrent ingestion.

Keep writes explicit

Choose a conflict and orphanEntities policy for every managed write. Anvia runs provisioning in auto-commit sessions and document replacement or deletion in explicit atomic transactions. If a graph and vector store must be updated together, persist application-owned reconciliation state because those writes are not one distributed transaction.

Own credentials and native queries

Use scoped credentials and private networking. Never expose nativeDriver() or arbitrary Cypher as an Agent tool. A driver created from uri is closed with MemgraphClient; a supplied driver remains caller-owned. Both ownership modes support await using.

Use opaque explorer IDs only for follow-up expansion. Application logic should rely on the stable identity properties declared in the graph schema.