Prepare and test your skills
Prepare and test your skills
Core functionality and grounding mechanisms in Vertex AI (opens in a new tab) Agent Builder describe how teams quickly assemble agents from models and data while keeping answers tied to trusted enterprise sources. The platform provides pre-trained models that work out of the box, user-trained models tuned to a specific use case, and an app graph that connects models to inputs and destinations. Grounding depends on the quality of the data and prompts that feed the model, because a model is only as good as the data it receives.
Pre-trained and user-trained models in Vertex AI are the two model choices for an agent, and the choice 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, then train a model 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 training data is available.
The low-code app graph in Vertex AI (opens in a new tab) is the visual structure where an application is built from connected nodes without custom integration code. Each node represents a model, an input source, or a destination, and connections between nodes define how data moves through the application. The builder flow runs in order:
A low-code app graph in Vertex AI Agent Builder: an input source node connects to an AI model node, and the model's output flows via connections to three destination nodes — streams, another AI model, and a media warehouse.
Once the model is a node in the graph, its output can flow to streams, to other models, or to a media warehouse. This lets one model's output feed another model's input or feed a storage destination.
The training workflow for a user-trained model is the longer path for adding a custom model to the app graph, because the model must first learn from a prepared dataset. For an object detection example, the ordered steps are:
AutoML vision models (opens in a new tab) created in October 2022 or later are supported in this flow, while compatibility is not guaranteed for models trained before that date.
Extensions (opens in a new tab) and remote inference extend Vertex AI models to tools and data platforms outside the app graph. An extension is a Vertex AI component that executes with an operation and parameters and returns structured output. Executing an extension requires a project, a location, and an extension ID, plus an operation ID with its parameters, and the response can include generated code, an execution result, an error field, and output files. 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 is then run with the ML.PREDICT function. In that arrangement, all inference processing happens in Vertex AI while the data stays in BigQuery, which avoids pipelines to move data back and forth and helps when GPU hardware support is needed.
Grounding through verified data and prompts is the practice of keeping model responses tied to accurate, authoritative sources instead of unsupported output. Inaccurate or incomplete input data and prompts lead to suboptimal performance or false outputs, so inputs must be verified for accuracy before use. Understanding the limitations of generative AI models helps developers mitigate misuse and unintended issues in the applications they build.
Exam tip: A pre-trained model is ready to use on a general objective, while a user-trained AutoML or custom-trained model requires you to supply sample data and train it for your specific use case.
Business value and operational impact from Vertex AI Agent Builder (opens in a new tab) describe how enterprises shorten delivery time, improve answer quality, and lower the skills needed to build agents. The platform combines ready-to-use pre-trained models with options to train custom models, so teams can start quickly and specialize later. Operations stay predictable because compute, throughput, and storage needs are defined per service.
Accelerated time-to-market means teams launch working agents faster by reusing pre-built capabilities instead of building models from scratch. Vertex AI (opens in a new tab) provides pre-trained models that perform a specific objective, are trained on generalized data, and are ready to use. This reuse reduces initial data collection, training time, and integration work, which accelerates time to value. Three platform features support faster delivery:
Because models, tools, and APIs are already available, a team can connect an input source to a model node and move toward production while planning later customization.
Improved accuracy in automated communications means agent answers stay correct, complete, and fair by grounding the agent in high-quality data and the right model choice. A generative AI model is only as good as the data used with it, so inaccurate or incomplete input data and prompts lead to suboptimal performance or false outputs. Teams must verify input data and prompts before trusting automated messages. Pre-trained models fit general tasks, while user-trained models fit unique cases that need specialized behavior and require sample data plus training through AutoML or custom training. Teams should also review regional regulations and fairness guidance, reduce bias in training datasets, and evaluate custom models for disparities in performance.
Exam tip: Choose a pre-trained model for a general task that is ready to use, and choose a user-trained AutoML or custom-trained model when the agent must reflect unique enterprise data.
Lower technical barriers with predictable operations means non-developer teams can build agents while capacity and storage remain plannable. A business user can prepare object detection data, create a dataset, select AutoML, choose where to use the model, and start training without building training infrastructure. After training, the model can be added to an application (opens in a new tab) and connected to streams, other models, or a media warehouse. Each pre-trained service has defined compute and throughput limits on an appliance with a single NVIDIA A100 80 GB graphics processing unit (GPU), so only one GPU-dependent API can run at a time:
| 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 |
Speech-to-Text can run alongside either OCR or Translation because it uses only CPU resources. If both Translation and OCR are enabled, the second API fails with insufficient GPU resources. Capacity limits are approximate and vary with input complexity, appliance configuration, and concurrent use, so teams should estimate peak usage and future growth and add appliance units for demanding workloads. Storage must also be planned, including 5 GB for the OCR backend, 1.5 GB for the Speech-to-Text backend, and 61.4 GB for the Translation backend plus smaller frontend and extractor components.
Strategic use cases for custom agents are organizational applications where a generative AI agent carries out a defined task, such as answering customer questions or giving employees natural-language access to a complex knowledge base. Business value comes from fitting the specific use case, but results depend on data quality, model choice, and the environment where the agent runs. Governance over models, data, and monitoring determines whether the use case can scale safely.
Data quality and responsible use for support assistants set the ceiling on answer quality for customer support and internal knowledge agents. A generative AI model is only as good as the data used with it, so inaccurate or incomplete input data or prompts produce suboptimal performance or false outputs. Input data and prompts must be verified as accurate before the agent handles real work, because a weak or outdated knowledge base produces weak or wrong answers. Organizations should plan for model limitations up front to mitigate misuse and unintended issues. Fairness planning includes reviewing regulations in the deployment region and industry guidance, reducing bias in training datasets, and evaluating custom models for performance disparities.
Model build choice for domain fit determines how much work the organization does and how closely the agent matches its task. A pre-trained model performs a specific objective, is trained on generalized data, and is ready to use. A user-trained model requires the organization to identify and provide sample data and then train a model suited to its unique use case, using either AutoML trained models or custom-trained models. Choose a pre-trained model when the task is common and general behavior is good enough, and choose a user-trained model when the agent must reflect the organization's own data and domain.
Deployment constraints in air-gapped environments describe what an agent can do when it must run inside an isolated device. A Google Distributed Cloud (GDC) air-gapped appliance (opens in a new tab) runs Vertex AI services entirely inside a portable device, which provides data sovereignty and compliance for organizations that cannot send data outside their own environment. The appliance offers a select set of pre-trained Vertex AI APIs — Optical Character Recognition (OCR), Speech-to-Text, and Vertex AI Translation — with the same tools and APIs used in Google Cloud. Hardware limits shape use: the device includes one NVIDIA A100 80 GB GPU, so only one GPU-dependent API can be active at a time and a second one fails with an insufficient-GPU-resources error. Speech-to-Text needs only CPU, so it can run alongside either OCR or Translation.
Throughput per appliance unit is approximate and not guaranteed: OCR handles up to 30 images per minute, Speech-to-Text up to seven minutes of audio per minute, and Vertex AI Translation up to 61,000 characters per minute. Actual capacity varies with input complexity such as language, image quality, and audio clarity, plus appliance configuration and concurrent use. Teams should estimate peak usage and future growth and deploy multiple appliance units for demanding workloads.
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.
Governance controls and model lifecycle management let an organization adopt custom agents while keeping data inside approved boundaries and tracking behavior after deployment. Generative AI on Vertex AI (opens in a new tab) supports customer-managed encryption keys (CMEK), VPC Service Controls, and Data Residency, which control encryption, network boundaries, and where data is stored. The Vertex AI Model Registry is a searchable repository for managing the machine learning model lifecycle, where teams organize models, train new versions, and deploy directly to endpoints. Model evaluation supplies metrics such as precision and recall to judge performance, and Model Monitoring watches deployed models over time.
Pre-trained models in Vertex AI perform a specific objective, are trained on generalized data, and are ready to use without further training, while user-trained models require the team to identify and provide sample data and train a model for a unique use case using 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 training data is available.
An agent application is assembled as an app graph where each node represents a model, an input source, or a destination and connections between nodes define how data moves through the application. The builder flow is to select a model from the AI models list and review its configuration settings, connect the model to an input source by selecting the source node and choosing Add output, and add the model as a node by selecting it from Available nodes. Once the model is a node in the graph, its output can flow to streams, to other models, or to a media warehouse, which lets one model's output feed another model's input or feed a storage destination.
Grounding keeps model responses tied to accurate, authoritative sources because a model is only as good as the data it receives. Inaccurate or incomplete input data and prompts lead to suboptimal performance or false outputs, so inputs must be verified for accuracy before use.
A Google Distributed Cloud (GDC) air-gapped appliance runs Vertex AI services entirely inside a portable device and 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 Optical Character Recognition (OCR) or Vertex AI Translation. Throughput per appliance unit is approximate with Optical Character Recognition (OCR) handling up to 30 images per minute, Speech-to-Text handling up to 7 minutes of audio per minute, and Vertex AI Translation handling up to 61,000 characters per minute. Actual capacity varies with input complexity, appliance configuration, and concurrent use, so teams should estimate peak usage and future growth and add appliance units for demanding workloads.
An enterprise organization wants to rapidly build and prototype a custom customer support assistant. The team requires a visual, low-code design environment to quickly configure conversation flows and wants to connect the assistant to internal company knowledge bases using built-in grounding tools to ensure responses are accurate and factual.
Which Google Cloud offering meets these requirements?
Vertex AI Agent Builder
Cloud Natural Language API
Vertex AI Model Garden
BigQuery ML
Last updated