Google Vertex AI

Provides access to Gemini 3 and earlier versions (PaLM) via REST API and SDKs, supporting text, image, video, and code inputs in a single request. Models are hosted on Google's managed infrastructure with automatic scaling and pay-per-token pricing. Requests are routed through Vertex AI's inference endpoints with optional request/response logging and monitoring via Cloud Logging.

Centralized registry of 200+ models spanning first-party (Gemini, Imagen, Lyria, Chirp, Veo), third-party (Anthropic Claude), and open-source (Gemma, Llama) artifacts. Model Garden provides filtering, comparison, and one-click deployment to Vertex AI endpoints. Each model includes metadata (task type, input/output specs, pricing estimates) and links to documentation and sample notebooks.

Managed vector database for storing and searching high-dimensional embeddings at scale. Supports approximate nearest neighbor (ANN) search with low latency and high throughput. Vector Search integrates with Vertex AI embeddings (from Gemini or custom models) and can be used for semantic search, recommendation systems, and similarity matching. Indexes are automatically optimized for query performance.

Native integration between Vertex AI and BigQuery enabling seamless data pipelines from data warehouse to ML models. BigQuery tables can be used directly as training data sources, feature computation sources, and prediction input. Vertex AI notebooks have native BigQuery connectors for exploratory analysis. Feature Store and RAG Engine integrate with BigQuery for feature retrieval and document indexing.

Vertex AI Model Garden includes third-party models (Anthropic Claude) alongside first-party models (Gemini, Imagen). Third-party models are accessed through unified Vertex AI APIs without requiring separate accounts or API keys. Billing is consolidated through Google Cloud. Model selection and switching is simplified through Model Garden discovery.

Vertex AI supports enterprise security controls including VPC Service Controls (VPC-SC) for network isolation and Customer-Managed Encryption Keys (CMEK) for data encryption. Models and data can be isolated within a VPC perimeter, preventing unauthorized access. Encryption keys are managed by the customer, meeting compliance requirements (HIPAA, FedRAMP, etc.). Audit logging via Cloud Audit Logs tracks all API calls and data access.

Managed Jupyter notebook environments for exploratory ML development. Vertex AI Workbench provides pre-configured notebooks with Vertex AI SDKs and BigQuery connectors. Colab Enterprise offers a lightweight alternative with similar integrations. Notebooks can be scheduled to run as jobs, enabling automated data exploration and model training workflows. Notebooks are stored in Cloud Storage with version control.

Unified environment for building, testing, and deploying custom AI agents using Gemini as the reasoning engine. Agents are registered in the Gemini Enterprise app with governance controls (access policies, audit logs). Agent Studio provides a prompt testing interface supporting text, image, video, and code inputs. Agents can be extended with custom tools (function calling) and real-time data retrieval via the Extensions system (mechanism not detailed).

Managed RAG service that integrates document ingestion, embedding generation, vector storage, and retrieval into a unified pipeline. RAG Engine handles chunking, embedding (using Google's embedding models), and semantic search over indexed documents. Retrieved context is automatically injected into Gemini prompts for grounded generation. Integration with BigQuery for structured data retrieval is mentioned but not detailed.

Vertex AI supports both AutoML (automated training for structured data and images) and custom training (bring-your-own-code with TensorFlow, PyTorch, scikit-learn). Training jobs run on managed infrastructure with automatic scaling, distributed training support, and hyperparameter tuning. Models are registered in the Model Registry and can be deployed to endpoints. Fine-tuning options for foundation models mentioned but specifics unknown.

Deploy trained models or pre-built models from Model Garden to managed Vertex AI endpoints with automatic scaling based on traffic. Endpoints support both online (real-time) and batch prediction. Models are containerized (Docker) and served via REST or gRPC APIs. Endpoints include built-in monitoring for latency, throughput, and error rates. VPC-SC and CMEK support for enterprise security (mentioned but not detailed).

Enterprise-grade evaluation service for assessing generative AI models against custom metrics and benchmarks. Evaluates models on dimensions like accuracy, safety, latency, and cost. Supports both automated evaluation (using rubrics and metrics) and human-in-the-loop review. Results are compared across model versions to identify the best model for your use case. Integration with Model Garden for model selection.

Managed feature store for managing, serving, and reusing ML features across training and prediction. Features are stored in a centralized repository with versioning and lineage tracking. Feature Store integrates with BigQuery for feature computation and Vertex AI Training/Prediction for feature retrieval. Supports both batch and online feature serving with low-latency access.

Automated monitoring service that tracks model performance in production, detecting input skew (distribution shift in features) and prediction drift (changes in model outputs). Monitoring is configured at deployment time with thresholds for alerts. Integrates with Cloud Logging and Cloud Monitoring for alerting and dashboards. Supports custom metrics and comparison against baseline distributions.

Workflow orchestration service for building and executing multi-step ML pipelines using Kubeflow Pipelines DSL or Python SDK. Pipelines define DAGs of tasks (training, evaluation, deployment) that run on managed infrastructure. Pipelines integrate with Vertex AI services (Training, Prediction, Feature Store) and external systems via custom containers. Execution history, logs, and artifacts are tracked automatically.