- RagCorpus REST body uses vectorDbConfig.ragManagedDb (not ragManagedDbConfig)
- RagManagedDb has no tier field - empty object {} = serverless
- Add GET+verify step after PATCH to confirm spanner key is gone before proceeding
- Print actual engine config state so failures are immediately visible
The v1beta1 API uses ragManagedDbConfig (not vectorDbConfig) on the RagCorpus
resource itself. Omitting it causes the engine to default to Spanner. Also the
engine config PATCH must use {"ragManagedDbConfig":{"basic":{}}} with the
correct field name matching v1beta1 schema.
Two root causes fixed:
1. ragEngineConfig is a project-level control-plane endpoint that only exists in
us-central1. Patching it against europe-west1 (or any other region) returns
'Invalid endpoint name'. We now always use us-central1 for the config PATCH
regardless of RAG_LOCATION.
2. Including vectorDbConfig.ragManagedDb in the corpus creation payload explicitly
requests Spanner mode, which is allowlist-restricted for new projects. Omitting
vectorDbConfig entirely lets the engine honour the project-level 'basic' config
and use serverless mode — the correct approach per Google docs."
SDK 1.153.1 always defaults to Spanner when backend_config is omitted,
and crashes when it's provided. Use REST POST directly with
vectorDbConfig.ragManagedDb set to force serverless.
SDK 1.153.1 has a bug in set_backend_config where it calls .CopyFrom()
on RagManagedDb which is not a real protobuf message.
The project-level ragEngineConfig PATCH already sets basic tier,
so passing backend_config is unnecessary and breaks corpus creation.
- Pass RagVectorDbConfig with RagManagedDbConfig basic tier explicitly
to avoid Spanner allowlist errors on new projects
- Add pre-flight ensure_serverless_engine_config() via REST PATCH
- Cleaner error messages with actionable hints
- Idempotent: skips creation if corpus already exists
- Add agents/core-logic/Dockerfile.base with heavy deps pre-baked
- Update agents/core-logic/Dockerfile to build FROM base image
- Update cloudbuild.yaml with step to build/push base (manual only)
- Typical deploy time reduction: ~3 min → ~30 sec
Updated the status and added LLM routing strategy details, including models, pricing, and configuration. Improved cost estimates and region availability information.
