Vertex AI

Vertex AI Agent Builder is Google Cloud's platform for building and deploying AI agents — ReAct agents that combine reasoning with tool use (web search, code execution, custom functions, RAG retrieval). Key 2026 updates: Duo Agent Platform reached GA in January 2026 with multi-agent orchestration, ADK (Agent Development Kit) enables single-command production deployment, observability dashboard tracks token usage/latency/error rates per agent, evaluation layer simulates user interactions for automated agent testing, and Memory Bank (GA) persists agent state across sessions at $0.25/1,000 events.

Vertex AI RAG Engine is a managed RAG (Retrieval-Augmented Generation) pipeline that handles: document ingestion from Cloud Storage or Google Drive, automatic chunking, embedding with text-embedding-004, vector indexing in Vector Search, and Gemini response grounding with retrieved context. RAG Cross Corpus (preview) retrieves across multiple document corpora simultaneously. Use RAG Engine for enterprise knowledge bases, internal Q&A systems, and document-grounded AI assistants — without building the ingestion, embedding, and retrieval pipeline yourself.

Model Garden provides 100+ models including: Google's Gemini family (3.5 Flash, 3.1 Pro, 2.5 Pro/Flash), Anthropic Claude 4 series (Opus/Sonnet/Haiku), Meta Llama 4 (Scout, Maverick), Google's open models (Gemma 3), specialized models (Imagen 3 for images, Veo 3.1 for video, Lyria 3 for audio), code models (Codey), and embedding models (text-embedding-004). All models use IAM authentication, support VPC-private endpoints, and are billed through GCP — no separate API keys per provider.

Vector Search 2.0 is Google Cloud's managed approximate nearest-neighbor search service — sub-10ms latency at billion-scale vector indexes. It stores high-dimensional embeddings (from text-embedding-004, multimodal embeddings, or custom models) and returns semantically similar results on query. In Vertex AI RAG Engine, Vector Search is the retrieval backend. For standalone use: create an index from a Cloud Storage JSONL of embeddings, deploy to an IndexEndpoint, and query via API. Pricing is per index node-hour plus per query. GA since late 2025 with simplified deployment.

Vertex AI pricing varies by component: Custom Training charges per TPU/GPU node-hour (TPU v5 ~$2.80/chip-hour, A3 Mega ~$15/GPU-hour); Prediction endpoints charge per node-hour; Model Garden (Gemini) charges per token (Gemini 2.5 Pro ~$1.25/1M input); Vector Search ~$0.70/node-hour plus $0.06/1M queries; Agent Engine Sessions $0.25/1,000 events. GCP free tier and credits apply. Share your expected usage across training, prediction, Vector Search queries, and Model Garden token volume for a detailed cost model.

Vertex AI Pipelines is a managed Kubeflow Pipelines service — define ML workflows as Python component graphs, run them on managed GCP infrastructure, and track all artifacts, parameters, and runs in Vertex ML Metadata. Use Vertex AI Pipelines when: models need regular retraining on new data, you want reproducible training with full artifact tracking, or automated evaluation gates before deployment are required. For one-off training, Vertex AI Custom Training is simpler. TFX Pipelines integrates natively with Vertex AI Pipelines for TensorFlow ML workflows.

Both are managed ML platforms with comparable capabilities: custom training, AutoML, model serving, pipelines, and MLOps. Key differences: Vertex AI has access to Gemini/Claude/Imagen/Veo via Model Garden (no SageMaker equivalent for Google models), TPU training is Google Cloud-exclusive (SageMaker uses GPUs), and Vertex AI Agent Builder is more integrated for multi-agent systems. SageMaker has more established documentation, larger community, and deeper AWS service integrations (Lambda, ECS, DynamoDB). Choose based primarily on which cloud your application stack runs on.

Yes — Vertex AI supports supervised fine-tuning for Gemini models (upload JSONL dataset, configure training parameters, deploy fine-tuned model to Prediction endpoint). For open-source LLMs (Llama 4, Gemma 3), Vertex AI Custom Training runs HuggingFace PEFT/LoRA fine-tuning on A3 GPU or TPU v5 clusters. The Gemini fine-tuning API is simpler than custom container training but less flexible. For complex LoRA fine-tuning with custom training loops, custom training containers on Vertex AI provide full flexibility with managed GPU infrastructure.

Enterprise Vertex AI security configuration: VPC Service Controls creates a network perimeter restricting data access to GCP resources within your VPC; CMEK (Customer-Managed Encryption Keys) via Cloud KMS encrypts model artifacts and training data; IAM Conditions provide fine-grained access control per model, endpoint, and dataset; Cloud Audit Logs record every API call with requester identity; Vertex AI Explainability generates feature attribution explanations for compliance documentation; data residency is enforced via region selection and VPC Service Control boundaries.

We provide Vertex AI managed support: Model Garden model version upgrades (models deprecate on Google's schedule), Agent Builder ADK updates as the SDK evolves, RAG Engine corpus maintenance for new document ingestion, Vector Search index optimization as corpus grows, Vertex AI Pipelines maintenance for data schema changes, and cost optimization reviews across training/serving/retrieval components. We also provide architecture reviews as new Vertex AI features (new model families, Agent Builder capabilities) create opportunities to improve existing deployments.