Prepare and test your skills
Prepare and test your skills
Enterprise data integration for generative AI connects large enterprise datasets to models so applications can analyze content, answer questions, and hold useful conversations. BigQuery ML (opens in a new tab) provides managed and task-specific AI functions for semantic analysis directly on rows and objects at scale. Agent Platform (opens in a new tab) provides model endpoints and evaluation metrics for multi-turn agents, while the AI Content Detection API helps identify AI-generated images.
Managed AI functions in BigQuery ML perform semantic analysis over structured data 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. This avoids per-row large language model calls, which reduces token consumption and latency when working with thousands or billions of rows. Tutorials show how to combine managed AI functions with general-purpose AI functions in queries and notebooks.
Task-specific AI functions in BigQuery ML handle enterprise data types that need specialized models beyond general text generation. BigQuery ML builds these solutions by using Cloud AI APIs, so a data warehouse query can depend on specialized models for language, documents, speech, and images. Data flows from BigQuery to the matching Cloud AI API model, and results flow back into the same BigQuery workflow.
These functions let the same workflow process text, documents, audio, and images by sending each type to its matching model.
Model access in BigQuery ML depends on Agent Platform endpoints, supported locations, and network boundaries. BigQuery ML supports the models gemini-3.1-flash-lite, gemini-3.5-flash, gemini-3.5-flash-lite, gemini-3.6-flash, and gemini-3.7-flash, and Agent Platform only supports multi-regional endpoints for these models. If a short endpoint name is used, BigQuery ML selects the us endpoint for queries in us or a single US region, the eu endpoint for queries in eu or a single EU region other than europe-west2 or europe-west6, and the global endpoint for all other locations. Queries in asia-south1 must use the fully qualified global endpoint name that includes the project ID and model ID.
Inside a Virtual Private Cloud Service Controls perimeter, AI functions cannot process ObjectRef values that use delegated access. Delegated access generates a signed HTTPS URL for the object, and Agent Platform blocks HTTP and HTTPS fetches for projects inside a perimeter, so the function returns an INVALID_ARGUMENT error about HTTP links. Because the ref column of an object table always uses the object table connection as the authorizer, passing ref to an AI function inside a perimeter always returns this error. To analyze the object, pass a single-argument OBJ.MAKE_REF(uri) value instead, which uses direct access and sends the Cloud Storage URI to the model without generating a signed URL.
Multi-turn agent metrics evaluate full conversation context to assess overall agent performance across multiple turns. These adaptive rubric metrics are reference-free and focus on what happened during the conversation rather than comparing output to a single correct answer. The registry distinguishes task outcome, tool behavior, and conversation path.
| Metric | Type | Assesses |
|---|---|---|
| Agent Multi-turn Task Success | Adaptive rubric | Whether the agent achieved the goal or goals of the conversation, not how it achieved them |
| Agent Multi-turn Tool Use Quality | Adaptive rubric | Whether the agent called the right tools with correct arguments at the right time |
| Agent Multi-turn Trajectory Quality | Adaptive rubric | Whether the reasoning path was logical and efficient in how the goal was achieved |
Exam tip: Task Success asks whether the goal was achieved, while Trajectory Quality asks how logically and efficiently the agent reached the goal.
Metric management for agents uses the Google Cloud console and the Agent Platform SDK to define, register, and run evaluations. Predefined metrics include final response quality, tool use quality, hallucination, and safety, and results can be visualized after running client.evals.evaluate with a dataset and selected metrics. Custom metrics extend the registry with organization-specific checks for tone, accuracy, or other behavior.
Managing metrics in the console follows an ordered workflow:
Registering metrics with the SDK sends the definition from code to the managed registry. A custom large language model metric defines a name, prompt template, and result-parsing function, then registers with create_evaluation_metric. A custom code metric defines a Python evaluate function that reads agent_eval_data and turns, then registers the same way for programmatic multi-turn checks.
The AI Content Detection API identifies AI-generated or modified images from Google and third-party models to support content trust. The API accepts JPEG, PNG, and WebP image payloads sent to the Gemini Enterprise Agent Platform endpoint using a REST API call. Machine learning models process the payload by analyzing pixel-level artifacts, noise patterns, and spectral anomalies, and the response aggregates findings from the models. The API does not store or retain processed images and does not include C2PA metadata detection.
The AI Content Detection API is in Private Preview and requires access through an application form. The API provides a probabilistic estimate and does not guarantee definitive identification, so models may produce false positives that flag authentic images and false negatives that miss AI-generated images. Output should serve as a supplementary signal and must not be used as the sole basis for critical decisions such as content takedowns or user sanctions. The service is provided as is, with no liability for damages or disputes from reliance on its outputs.
Vertex AI Model Garden (opens in a new tab) is the part of the Vertex AI Platform used to discover, manage, and deploy foundation models. A foundation model (FM) is trained on broad data and can be adapted, such as through fine-tuning, to many downstream tasks. Model Garden brings together Google proprietary models and third-party or open-source alternatives so teams can match a model to business and technical needs.
Model Garden discovery helps teams find foundation models for different tasks. Google Cloud provides state-of-the-art foundation models through Vertex AI for text, code, images, and embeddings. Third-party models can be deployed to Model Garden or self-hosted on other compute platforms, which lets teams compare Google proprietary models with open-source alternatives.
Open-source models, including HuggingFace models, can be deployed to Model Garden or to Google Kubernetes Engine (GKE). Third-party models can also be self-hosted on GKE, Cloud Run, or Compute Engine.
Model Garden management is the entry point for AI models that come from Google Cloud Marketplace (opens in a new tab). To manage a procured AI model on Vertex AI, open the product page for the AI model in Cloud Marketplace and click Manage on Vertex AI. This action connects the Marketplace model to Vertex AI so it can be managed there. Discovery therefore covers both Google and third-party models, while management connects a Marketplace purchase to Vertex AI.
Foundation model deployment choice depends on how much infrastructure management the team wants and on the model technical or licensing needs. Deployment infrastructure choice (opens in a new tab) determines whether Vertex AI manages the infrastructure or the team self-hosts it.
| Option | Management | Choose when |
|---|---|---|
| Vertex AI Model Garden | Managed by Vertex AI | The team wants fully managed infrastructure with first-class generative AI tools and APIs, and needs to deploy a third-party or open-source model |
| 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, a legacy OS, or special licensing terms |
After a model is selected, Vertex AI Model Registry can manage the model lifecycle. Vertex AI Model Registry is a central repository for organizing, tracking, and training new model versions. When a version is ready, it can be assigned to an endpoint directly from the registry or deployed to an endpoint using aliases.
Model Garden testing and business value come from separating discovery and deployment from quality assessment. Model Garden itself handles discovery, management, and deployment of foundation models. For testing, Vertex AI Model Evaluation is a managed service that assesses model performance and quality by generating evaluation metrics and visualizations, helping teams identify bias and make deployment decisions. The business value is access to a diverse set of Google proprietary and open-source models, so teams can choose based on task, infrastructure preference, licensing, and operating system needs. This reduces the need to build every model from scratch, because foundation models can be adapted or fine-tuned to specific downstream tasks.
Exam tip: Model Garden manages AI models from Google Cloud Marketplace, while Vertex AI Model Registry manages the lifecycle of ML models and deploys versions to endpoints.
Vertex AI AutoML (opens in a new tab) is a managed training service on the Vertex AI Platform (opens in a new tab) that builds high-quality custom machine learning models with minimal coding. A business user or developer supplies a labeled dataset, and AutoML handles model selection, training, and automated tuning to produce a deployable model. This approach lets teams without deep machine learning expertise create custom models for their own data and reach business value faster.
Vertex AI AutoML reduces custom model development to supplying data and setting a small number of training controls. Data flows from the customer dataset to the AutoML training job, then the resulting model flows to the Vertex AI Model Registry, and from the registry the model can be assigned directly to an endpoint for Vertex AI Inference. Inference is the step where the deployed model makes predictions on new, unseen data. The training set is typically about 80% of the data and is used to teach patterns, while separate validation and test sets are used to check performance during and after training.
A five-step left-to-right process flow showing labeled dataset feeding AutoML training, then Model Registry, endpoint deployment, and inference on new data, with data split and control notes.
The user controls only limited aspects of the training environment, mainly the budget for training hours and whether to enable early stopping. AutoML performs some automated tuning of hyperparameters, but the user cannot view or modify those values. This limited control is intentional, because it removes infrastructure and coding work and keeps the workflow accessible to non-specialists.
Vertex AI AutoML data types define the kind of custom model that can be trained. Each type maps to common business problems that use an organization own examples.
AutoML for tabular data also supports forecasting with time series data, where the model predicts a series of numeric values that extend into the future. Forecasting with AutoML is useful for demand planning, revenue projection, and other time-based predictions. 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 business value of Vertex AI AutoML is democratization of AI and faster time-to-value. Democratization means analysts, developers, and domain teams can build custom models without writing training code or managing servers. Because the service automates architecture choice and tuning, a team can move from labeled data to a working model in hours instead of building a modeling pipeline from scratch. This accelerates time-to-value for specific, well-defined use cases where the business already has examples.
Central management adds to that value because models flow into the same registry and endpoints as other Vertex AI models. Teams can organize versions, track lineage, and deploy to online or batch inference without a separate system. The result is broader participation in AI across the organization while governance and deployment stay consistent.
Vertex AI AutoML tradeoffs center on flexibility versus simplicity. AutoML requires the user to target one of its predefined objectives and enforces data-size limits that vary by dataset type. Custom training has no predefined-objective limit and no AutoML-style data-size limit, but it requires the team to provide training code and configure infrastructure.
| 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 |
Exam tip: Choose AutoML when the task fits a predefined objective and minimal coding matters; choose custom training when manual hyperparameter tuning or full control of machine type and container is required.
Vertex AI Model Garden is the part of the Vertex AI Platform used to discover, manage, and deploy foundation model (FM) technology. It brings together Google proprietary models such as Gemini, Codey, Imagen, and text-embedding alongside third-party and open-source alternatives so teams can match a model to business and technical needs. This reduces the need to build every model from scratch because foundation models can be adapted or fine-tuned to specific downstream tasks.
Vertex AI AutoML should be used when the task fits one of its predefined objectives and minimal coding matters, while custom training should be used when manual hyperparameter tuning or full control of machine type and container is required. Vertex AI AutoML builds high-quality custom models from a labeled dataset with automated model selection, training, and tuning and limits training environment control to training-hour budget and early stopping with automated-only hyperparameter tuning. Custom training has no predefined-objective limit and no AutoML-style data-size limit but requires training code and allows specifying Compute Engine machine type, disk size, and container image with tuning during each run.
Agent Multi-turn Task Success measures whether the agent achieved the goal or goals of the conversation, while Agent Multi-turn Trajectory Quality measures whether the reasoning path was logical and efficient in how the goal was achieved. Both are adaptive rubric, reference-free metrics that evaluate full conversation context rather than comparing output to a single correct answer.
Managed AI functions in BigQuery ML perform semantic analysis over structured data and object data by training a lightweight, distilled model on a sample of the data and then using that distilled model for inference on most rows. This efficient mode avoids calling a large language model for every row, which reduces token consumption and latency when working with thousands or billions of rows.
An enterprise organization wants to build an internal generative AI assistant that allows employees to find answers across various internal data sources, such as Jira, Confluence, and document repositories.
The organization requires the solution to:
How does Vertex AI Search and Conversation enable this capability?
It restricts retrieval to public web indices and disables internal single sign-on to avoid managing corporate identities.
It connects to enterprise data sources to index content and uses Retrieval-Augmented Generation (RAG) to ground conversational responses with access controls.
It converts all enterprise databases into static Looker dashboards that require manual keyword filtering instead of natural language querying.
It continually fine-tunes and retrains foundational models on internal datasets each time an employee submits a new conversational query.
Last updated