Vertex AI Index Endpoint

An index endpoint serves vector similarity queries from deployed indexes on dedicated machines billed per node-hour continuously. Abandoned vector-search endpoints are a fast-growing category of idle AI spend.

Query-serving machines that never clock out

Deploying an index to an index endpoint provisions serving nodes, and those nodes bill per node-hour from that moment on, continuously, with query volume playing no part in the charge. Vector similarity serving is memory-hungry, so the machine shapes involved sit toward the expensive end of the catalog, and a forgotten deployment can quietly outbill the application it once supported.

Silence detection on the query path

ZopDev discovers index endpoints via Cloud Asset Inventory and flags endpoints with no query traffic in recommendations. Because the meter belongs to the deployed-index-on-endpoint pairing, the recommendation targets exactly that pairing: an endpoint still holding serving capacity for an index nothing queries. The type is not schedulable: vector serving is either needed or it is not, and the decisive action is undeployment rather than a nightly stop.

The retired RAG stack pattern

This type’s waste story is recent but consistent: a retrieval-augmented prototype stands up an index, an index endpoint, and an application; the application is abandoned; the endpoint keeps serving to nobody. A second pattern is scale left over from load testing, with node counts raised for a benchmark and never lowered afterwards. Undeploying the index releases the serving nodes and their charges; deleting the index separately ends its storage meter, and skipping either half leaves money running.

Index endpoints in the console

Google Cloud console → Vertex AI → Vector Search, on the index endpoints view, shows each endpoint’s deployed indexes and machine configuration per region. Those two facts determine what it costs and whether anything still needs it. The public or private connectivity setting appears there as well, which helps trace which network still expects to route queries at the deployment.