Prepare and test your skills
Prepare and test your skills
The Vertex AI Platform (opens in a new tab) is a managed platform for discovering foundation models, training custom models, evaluating quality, and deploying models to endpoints for inference. Data flows from customer datasets to training or retrieval services, models flow into the Vertex AI Model Registry for versioning, and deployed versions serve predictions to applications. The business value is faster delivery without building every model from scratch, with consistent governance for deployment.
Vertex AI Model Garden is the part of the Vertex AI Platform used to discover, manage, and deploy foundation models. A foundation model is trained on broad data and can be adapted through fine-tuning to many downstream tasks. Model Garden brings together Google proprietary models and third-party and open-source alternatives so a team can match a model to task, licensing, and infrastructure needs.
Discovery covers text, code, image, and embedding tasks:
Open-source models, including HuggingFace models, can also be deployed to Model Garden. To manage a procured AI model from Google Cloud Marketplace on Vertex AI, open the product page for the AI model in Cloud Marketplace and click Manage on Vertex AI.
Deployment choice depends on how much infrastructure management the team wants to keep:
| Option | Management | Choose when |
|---|---|---|
| Vertex AI Model Garden | Managed by Vertex AI | The team wants fully managed infrastructure with generative AI tools and APIs for a third-party or open-source model |
| Google Kubernetes Engine (GKE) | Self-hosted | The team wants to deploy an open-source model, including HuggingFace models |
| Cloud Run | Self-hosted | The team wants to self-host a third-party model |
| Compute Engine | Self-hosted | The model requires a specialized kernel, legacy operating system, or special licensing terms |
After selection, Vertex AI Model Registry organizes and tracks model versions and assigns versions to endpoints directly or with aliases. Vertex AI Model Evaluation then assesses performance and quality with metrics and visualizations to identify bias and support deployment decisions.
Exam tip: Model Garden manages AI models from Google Cloud Marketplace, while Vertex AI Model Registry manages the lifecycle of machine learning models and deploys versions to endpoints.
Vertex AI AutoML is a managed training service that builds custom machine learning models from labeled data with minimal coding. A developer supplies a labeled dataset, and AutoML handles model selection, training, and automated tuning to produce a deployable model. Data flows from the customer dataset to the AutoML training job, the resulting model flows to the Vertex AI Model Registry, and the registry version flows to an endpoint for Vertex AI Inference, where the deployed model makes predictions on new data.
AutoML supports four main data types, each mapped to common business problems:
Training uses about 80% of the data to teach patterns, with separate validation and test sets to check performance. The user controls only the training-hour budget and early stopping, while AutoML performs automated hyperparameter tuning that the user cannot view or modify. Training or updating new AutoML models for text classification, entity extraction, and sentiment analysis is being replaced by Gemini prompts and tuning, while existing AutoML text models remain usable for a transition period.
The tradeoff between AutoML and custom training is simplicity versus control:
| Capability | Vertex AI AutoML | Custom training |
|---|---|---|
| Machine learning objectives | Must use one of AutoML predefined objectives | No predefined limit |
| Hyperparameter tuning | Automated only, values cannot be manually modified | Can be tuned during each run for experimentation |
| Training environment control | Limited to training-hour budget and early stopping | Can specify Compute Engine machine type, disk size, and container image |
BigQuery ML brings semantic analysis and specialized artificial intelligence (AI) functions directly into the data warehouse workflow. Managed AI functions perform semantic analysis over structured rows and object data without calling a large language model for every row. The efficient mode trains a lightweight, distilled model on a sample of the data and then uses that distilled model for inference on most rows, which reduces token consumption and latency for thousands or billions of rows.
Task-specific AI functions handle enterprise data types by sending each type from BigQuery to its matching Cloud AI API model, with results flowing back into the same BigQuery workflow. These functions cover natural language processing for understanding text, machine translation between languages, document processing for extraction, audio transcription from speech to text, and computer vision for images. This lets one warehouse query process text, documents, audio, and images through specialized models.
Agent evaluation measures full conversation quality across multiple turns without comparing output to a single correct answer. Three adaptive rubric metrics separate what was achieved from how it was achieved:
| Metric | Assesses |
|---|---|
| Agent Multi-turn Task Success | Whether the agent achieved the conversation goal or goals, not how it achieved them |
| Agent Multi-turn Tool Use Quality | Whether the agent called the right tools with correct arguments at the right time |
| Agent Multi-turn Trajectory Quality | Whether the reasoning path was logical and efficient |
Teams define and register metrics in the Google Cloud console and with the Agent Platform software development kit (SDK). The console workflow runs in order: first open Agent Platform Agents Evaluation, then open the Metrics tab to view the registry, then click New metric and select Custom large language model (LLM) metric or Custom code metric, then define instructions, criteria, and rating scores. The AI Content Detection API supports trust by identifying AI-generated or modified JPEG, PNG, and WebP images sent to the Gemini Enterprise Agent Platform endpoint. The API analyzes pixel-level artifacts, noise patterns, and spectral anomalies, provides only a probabilistic estimate with possible false positives and false negatives, and must not be used as the sole basis for critical decisions such as takedowns.
Retrieval-Augmented Generation (RAG) is the method that grounds a generative model in company or real-time data before it answers. The retrieval layer finds relevant content, the retrieved content flows into the model prompt, and the model generates a response based on that context. This pattern matters when answers must come from current documents, customer records, or internal knowledge rather than only from model training.
Grounding improves model responses by adding retrieved facts to the prompt so the model does not have to guess. A hallucination happens when a model generates fluent but incorrect content because its training knowledge is incomplete or out of date. Gemini models can evaluate whether model knowledge is sufficient or whether grounding with Google Search is required for up-to-date internet information. For enterprise data, grounding follows a fixed order:
A left-to-right process flow showing enterprise data indexed as embeddings, retrieved as relevant passages for a user question, then added to the prompt for Gemini to generate a grounded answer.
An embedding is a numerical representation of text that lets a search engine compare meaning, not just keywords. Many engines store embeddings in a vector database, which is optimized for similarity search. Grounding can also use non-vector databases by querying Cloud SQL or Firestore and placing the query result in the prompt, or by calling external system APIs through Vertex AI Extensions.
Google Cloud RAG offerings provide different levels of management and control for different data sources and performance needs. Vertex AI Search is a fully managed and optimized solution that supports most data sources and prevents direct access to underlying embeddings. Vertex AI RAG Engine (opens in a new tab) is a managed orchestrator with a LlamaIndex-like interface for building a search engine for RAG. Vertex AI Vector Search provides low-latency vector search and large-scale serving as a specialized vector database.
High-impact enterprise use cases for RAG include semantic search across document repositories by meaning rather than keywords, context-aware virtual assistants for customer support that retrieve help articles and policies, and internal knowledge management that answers employee questions without manual search across repositories. The business value comes from more accurate and consistent answers, up-to-date grounding, secure access to enterprise data, and faster launch by reusing managed retrieval instead of building it from scratch.
| Option | Primary role | Choose when |
|---|---|---|
| Vertex AI Search | Fully managed optimized solution | You want support for most data sources quickly with no direct access to embeddings |
| Vertex AI RAG Engine | Managed orchestrator with developer interface | You want to build a search engine for RAG with customizable retrieval logic |
| Vertex AI Vector Search | Specialized vector database | You need low-latency, large-scale vector serving for a tailor-made engine |
| AlloyDB / Cloud SQL | SQL databases for operational data | Data is accessed programmatically, you already use SQL, or you need low latency and Google AI models directly from the database |
| BigQuery | Analytical data warehouse | You have a large analytical dataset needing batch processing and frequent SQL access |
Exam tip: Choose Vertex AI Search when you want a fully managed search engine without direct embedding access, and choose RAG Engine or Vector Search when you need to customize orchestration or low-latency vector serving.
Vertex AI Agent Builder (opens in a new tab) is a platform for quickly assembling custom generative AI agents from models and enterprise data while keeping answers tied to trusted sources. The platform provides pre-trained models that work out of the box, user-trained models tuned to a specific use case, and a low-code app graph that connects models to inputs and destinations. Grounding depends on verified data and prompts, because inaccurate or incomplete inputs lead to suboptimal performance or false outputs.
Model choice for Vertex AI Agent Builder depends on whether a general model fits or domain-specific training is needed. Pre-trained models perform a specific objective, are trained on generalized data, and are ready to use without further training. User-trained models require the team to identify and provide sample data and then train for a unique use case, using either AutoML trained models or custom-trained models. Choose a pre-trained model when speed matters and the task is general, and choose a user-trained model when general behavior is not accurate enough and representative data is available.
The low-code app graph is the visual structure where nodes for models, input sources, and destinations define how data moves through the application. The builder flow runs in order: first select a model from the AI models list and review its configuration, then connect the model to an input source by selecting the source node and choosing Add output, then add the model as a node from the list of Available nodes. Once the model is a node, its output can flow to streams, to other models, or to a media warehouse, so one model's output can feed another model's input.
The longer path for a user-trained model, such as object detection, runs in order: first prepare the data and create a dataset in Vertex AI, then open the dataset, click Train new model, choose AutoML, and select where the model will be used, then wait for training to complete, which can take many hours, until completion is signaled, then open the Applications tab and add the finished model to the app graph. Extensions extend agents to outside tools by executing with a project, location, extension ID, operation ID, and parameters and returning structured output. A remote model can also be created in BigQuery with a CREATE MODEL statement that references a model hosted in Vertex AI Inference, and inference runs with the ML.PREDICT function while data stays in BigQuery and processing happens in Vertex AI.
Business value from Vertex AI Agent Builder comes from faster launch, better accuracy in automated communications, and lower technical barriers with predictable operations. Reusing pre-trained models for text extraction, transcription, and translation reduces data collection, training time, and integration work, which accelerates time to value. Teams can start with a general model connected to an input source and specialize later with a user-trained model. Governance controls support safe scaling, including customer-managed encryption keys (CMEK) for encryption control, VPC Service Controls for network boundaries, and Data Residency for storage location, plus the Vertex AI Model Registry for lifecycle management, model evaluation with precision and recall, and Model Monitoring for deployed behavior.
Strategic use cases include customer support assistants and internal knowledge agents that give natural-language access to complex knowledge bases. Results depend on data quality, model choice, and deployment environment, including fairness work such as reviewing regional regulations, reducing bias in training datasets, and evaluating custom models for performance disparities. For isolated environments, a Google Distributed Cloud (GDC) air-gapped appliance runs Vertex AI services entirely inside a portable device for data sovereignty, offering Optical Character Recognition (OCR), Speech-to-Text, and Vertex AI Translation with the same tools and APIs as Google Cloud.
Hardware and throughput limits shape air-gapped design. The appliance includes one NVIDIA A100 80 GB graphics processing unit (GPU), so only one GPU-dependent API can run at a time, while CPU-only Speech-to-Text can run alongside either GPU workload:
| Pre-trained API | Computing resources | Capacity per appliance unit |
|---|---|---|
| Optical Character Recognition (OCR) | GPU, 40 GB GPU memory | Up to 30 images per minute |
| Speech-to-Text | CPU only | Up to 7 minutes of audio transcribed per minute |
| Vertex AI Translation | GPU, 50 GB GPU memory | Up to 61,000 characters per minute |
If both Translation and OCR are enabled, the second API fails with insufficient GPU resources. Limits are approximate and vary with input complexity, configuration, and concurrent use, so teams should estimate peak usage and add appliance units for demanding workloads, plus plan storage including 5 GB for the OCR backend, 1.5 GB for Speech-to-Text, and 61.4 GB for Translation.
Exam tip: Only one GPU-dependent pre-trained API — OCR or Vertex AI Translation — can run at a time on a GDC air-gapped appliance, but Speech-to-Text can run alongside either because it uses only CPU.
Developers move from foundation to production by discovering a model in Model Garden, grounding it with RAG so answers reflect enterprise facts, and assembling it into a custom agent with Agent Builder. Vertex AI Search and RAG Engine reduce retrieval work, AutoML reduces training work, and the Model Registry, evaluation, monitoring, and governance controls keep versions, quality, and data boundaries consistent from prototype to deployment.
Vertex AI Model Garden is used to discover, manage, and deploy foundation models including Google proprietary models and third-party and open-source alternatives, while Vertex AI Model Registry organizes and tracks model versions and assigns versions to endpoints directly or with aliases. Model Garden brings together models such as Gemini, Codey, Imagen, and text-embedding models for different tasks, and the Registry supports deployment decisions with versioning and Vertex AI Model Evaluation for performance and quality.
Vertex AI AutoML should be used when you want to build a custom machine learning model from labeled data with minimal coding using one of its predefined objectives for image, text, tabular, and video data. AutoML handles model selection, training, and automated tuning with control limited to the training-hour budget and early stopping, while custom training should be used when you need unrestricted objectives and control over hyperparameter tuning, machine type, disk size, and container image.
Vertex AI Search is a fully managed and optimized Retrieval-Augmented Generation solution that supports most data sources and prevents direct access to underlying embeddings, while Vertex AI RAG Engine is a managed orchestrator for building a search engine for RAG with customizable retrieval logic and Vertex AI Vector Search is a specialized vector database for low-latency, large-scale vector serving. Choose Vertex AI Search when you want support for most data sources quickly with no direct embedding access, and choose RAG Engine or Vector Search when you need to customize orchestration or low-latency vector serving.
In Vertex AI Agent Builder, a pre-trained model should be used when speed matters and the task is general because pre-trained models are trained on generalized data and are ready to use without further training. A user-trained model should be used when general behavior is not accurate enough and representative data is available, because it requires the team to identify and provide sample data and then train for a unique use case using either AutoML trained models or custom-trained models.
Last updated