The main components of Vertex AI map to the ML lifecycle: data, training, registry, deployment, pipelines, and monitoring. Data lives in Cloud Storage and BigQuery, and you can define Datasets for labeling and validation. Training is handled by Custom Training Jobs (custom containers or prebuilt frameworks) and Hyperparameter Tuning. The Model Registry stores versioned artifacts, metadata, and lineage. Deployment happens via Endpoints for online serving or Batch Prediction for asynchronous jobs. Vertex Pipelines orchestrate multi-step workflows with reproducibility, and Model Monitoring plus Cloud Logging/Monitoring provide observability and alerts in production.
Model Garden and foundation model integrations let you use large models for generation and embeddings, while keeping the same control plane for quotas and security. TensorBoard integration captures metrics and visualizations during training. For compliance and governance, you can attach model cards, approvals, and evaluation reports to registry entries. Access control is handled with IAM, and networking with VPC SC and Private Service Connect helps restrict data movement. This stack gives developers a consistent way to move from experiment to production without stitching together ad-hoc scripts and servers.
In vector-focused systems, two components matter additionally: embedding endpoints and your vector store. Vertex AI provides the embedding generation (either with a hosted model or a custom one you deploy). Milvus acts as the external component for vector storage, indexing, and ANN search. You orchestrate periodic re-embedding and index refreshes with Pipelines, track versions and metrics in the Registry, and monitor retrieval performance alongside model metrics. This division ensures that each component does its job well: Vertex AI manages training/inference lifecycle; Milvus delivers low-latency semantic retrieval at scale.