Managing and provisioning a solution infrastructure · Topic 6 of 6
Prebuilt AI APIs, Gemini Enterprise and Model Garden
Choosing between Google's AI APIs for search, conversation, vision, video, image and audio, using Gemini Enterprise agents and NotebookLM, and Model Garden models.
- 4 min
- Prebuilt APIs for fixed tasks; Gemini for custom rules, reasons and mixed media.
- Structured output constrains Gemini to allowed values.
- Conversational Agents: webhooks, live agent handoff, versions and environments, redaction settings.
What it is
Many AI needs are met without training anything: a prebuilt API, a Gemini prompt, a ready-made agent or a model from Model Garden. The exam asks which one fits a requirement, and why a narrower API sometimes loses to Gemini, or the other way round.
How it works
The Google AI APIs
| Area | Service | What it does |
|---|---|---|
| Search | Agent Search (Google-quality search over your content, including commerce search) | Natural-language search and recommendations over websites, documents and catalogs |
| Conversation | Conversational Agents (Dialogflow CX), Agent Assist, Conversational Insights | Virtual agents for chat and voice, help for human agents, analysis of contact center conversations |
| Vision | Cloud Vision API | Labels, text detection (OCR), SafeSearch, objects in images |
| Documents | Document AI | Parsing forms, invoices and layouts, including tables, from scanned documents |
| Video | Video Intelligence API | Labels, shot changes, explicit content detection in video |
| Image generation and editing | Imagen on Agent Platform | Generating images, variations and edits from prompts |
| Audio | Speech-to-Text, Text-to-Speech | Transcription and spoken output |
| Language | Natural Language API, Cloud Translation | Entities, sentiment, classification; translation with glossaries |
Prebuilt API or Gemini?
| Situation | Better choice |
|---|---|
| A fixed, well-defined task with built-in categories (OCR, SafeSearch, translation) | The prebuilt API: predictable output, simple pricing |
| Your own rules or categories, written reasons, several media types at once | Gemini with instructions and a response schema |
| Output must match an allowed list exactly | Gemini structured output with enum fields |
| Consistent translation of brand terms | Cloud Translation Advanced with a glossary |
Built-in detectors return their own categories and scores. When auditors need a written reason against a company policy, Gemini with the policy in the prompt fits better.
Conversational Agents in practice
Conversational Agents handle chat and voice. Telephony integrations take calls; webhooks call your APIs (for example to place an order); handoff to a live agent passes the conversation context. Security settings can apply Sensitive Data Protection templates to redact data before it is stored, and set retention for conversation history. Versions and environments let you test a flow before it reaches customers.
Gemini Enterprise features
Gemini Enterprise brings Google's AI to employees: AI agents that work across company data and applications, and NotebookLM for research over uploaded sources. Use it to enhance internal workflows (searching policies, drafting, summarizing) without building a custom app.
Model Garden
Model Garden on Agent Platform is a catalog of Google models (such as Gemini, Imagen and others), partner models and open models. You can call managed models by API, or deploy open models to endpoints or GKE, and tune them when needed. Choose an open model when you need to run it in your own environment, control the weights or fine-tune deeply.
Integration patterns
| Pattern | Example |
|---|---|
| Event-driven | A file lands in Cloud Storage, Eventarc triggers a Cloud Run service that calls Gemini |
| Batch | A back catalog is processed as a batch inference job |
| Human in the loop | Generated attributes wait in a review app; associates approve, edit or reject |
| Grounding | Answers grounded in company data or search results to reduce invented facts |
| Safety | Model Armor screens prompts and responses for prompt injection and unsafe content |
How the exam asks about this
- "Virtual agent takes orders by voice and hands off to a human with context." Conversational Agents with telephony, webhook and live agent handoff.
- "Scanned PDF spec sheets with tables, no labeled data." Document AI Layout Parser.
- "Flag scenes against our own written policy with a reason." Gemini with the policy as instructions and a response schema.
What to remember for the exam
- Prebuilt APIs for fixed tasks; Gemini for custom rules, reasons and mixed media.
- Structured output constrains Gemini to allowed values.
- Conversational Agents: webhooks, live agent handoff, versions and environments, redaction settings.
- Gemini Enterprise and NotebookLM enhance employee workflows.
- Model Garden for Google, partner and open models.
Agent Assist helps human agents during calls; it does not answer callers by itself.
Common mistakes
Training a custom model when a prebuilt API or prompt already does the job.
Using regex checks on free-form output instead of a response schema.
Confusing Agent Assist with a virtual agent.
Check yourself
Which service extracts table rows from scanned PDF documents without custom parsing code?
Document AI, for example its Layout Parser.
How can Gemini be forced to return only values from an allowed list?
Use structured output with a response schema whose fields are enums of the allowed values.
What does Model Garden offer?
A catalog of Google, partner and open models to call, deploy or tune on Agent Platform.
Topic 6 of 6 in Managing and provisioning a solution infrastructure
Test yourself on Managing and provisioning a solution infrastructure
The full Google Cloud Professional Cloud Architect question bank has 877 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by Google Cloud. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.