Vertex AI Vector Index
Does ZopNight manage Vertex AI Vector Index?
A Vertex AI Vector Search index bills for the volume of embedding data indexed, a storage meter that runs 24/7 whether anything queries it or not. Serving cost accrues on the index endpoint, not the index itself. ZopNight discovers indexes through Cloud Asset Inventory and attributes vector-search storage inside its ML cost views.
Rules that fire on Vertex AI Vector Index
A Vertex AI index stores vector embeddings for similarity search (Vector Search, formerly Matching Engine), billed for the data indexed. Index size scales with embedding volume.
Embedding bytes drive the index charge
A Vector Search index bills for the volume of vector data it holds. Every embedding added grows the indexed corpus, so the meter tracks dataset size, not query load. Query load belongs to the index endpoint that serves the index. Streaming updates and periodic rebuilds change the contents over time, but at any given moment the charge follows how much data stands indexed.
Attributing vector storage to its owners
ZopDev discovers indexes via Cloud Asset Inventory to attribute vector-search storage in ML cost views. Attribution matters here because embedding corpora are usually produced by one team’s ingestion pipeline and queried by another team’s application; the index row is where that storage cost gets pinned to the workload that created it rather than dissolving into a generic AI line item.
Stale corpora and duplicate embeddings
The characteristic leaks: an index built for a retired RAG prototype that nobody deleted, the same corpus embedded twice under different chunking experiments with both indexes retained, and indexes that keep growing because ingestion runs on while the consuming application has been shelved. Since serving cost accrues on the index endpoint rather than the index, undeploying ends only the node-hour charges; the storage meter persists until the index itself is deleted. Both halves of the teardown matter. A stale index with an active rebuild schedule pays twice for irrelevance: once in stored embeddings, once in the compute each rebuild cycle burns.
Vector Search assets in the console
Google Cloud console → Vertex AI → Vector Search lists indexes per region with their update method and deployment state. An index with no deployment and no recent updates is the strongest candidate for deletion, and confirming its pipeline has stopped writing closes the case.