AI Intermediate: Google Vertex AI & Gemini Practice Questions

Vertex AI Agent Builder supports an 'Agent as a Tool' pattern where a sub-agent is registered as a tool that can be invoked by an orchestrator (parent) agent. The orchestrator's reasoning loop can call the sub-agent tool with structured input and receive its output, enabling context passing and delegation across specialized agents — all within the Agent Builder framework. Using multiple Data Stores with routing rules handles knowledge retrieval, not agent-to-agent delegation with context passing. Deploying sub-agents as Cloud Run services with Pub/Sub is a custom engineering approach that bypasses Agent Builder's managed orchestration and is significantly more complex to maintain. Vertex AI Pipelines is designed for ML workflow orchestration (training, evaluation), not for wiring together live conversational agents at inference time.

The two valid risks are PII exposure and data quality degradation. Raw production chat logs frequently contain user PII (names, emails, account numbers) which a fine-tuned model can memorize and regurgitate in future responses — a serious compliance and privacy risk. Additionally, production logs contain the full distribution of agent responses, including poor, off-topic, or incorrect ones; using these as supervision targets trains the model to replicate bad behavior. Vertex AI SFT supports training data in Cloud Storage (as JSONL files), not exclusively in BigQuery, so that claim is false. Gemini SFT on Vertex AI supports multi-turn conversational formats, so the conversion requirement described is incorrect. There is no mechanism in Vertex AI that automatically switches a job from SFT to RLHF based on dataset size; these are separate, explicitly configured tuning methods.

In KFP v2, the @component decorator from kfp.dsl marks a Python function as a pipeline component. To output primitive values such as a float, the function should return a NamedTuple with the appropriately typed fields, which KFP serializes as component outputs. Using @kfp.v2.dsl.component is the older import path and the combination with Output[Metrics] is for artifact outputs, not primitive float outputs. The @kfp.v2.dsl.pipeline decorator is for defining the pipeline DAG itself, not individual components, and components cannot return Python primitives directly without a NamedTuple wrapper. Writing to BigQuery is a valid side effect but does not expose the value as a typed KFP component output for downstream steps.

Gemini 1.5 Pro on Vertex AI natively accepts video and audio as multimodal inputs is correct. The model can ingest video files (including their audio tracks) directly, reason over the visual and auditory content simultaneously, and produce timestamped summaries—no pre-transcription step is required. Gemini is explicitly a multimodal model that handles text, images, audio, and video. The claim that Gemini only accepts text is false. Multimodal support is fully available through the Vertex AI API; it is not restricted to Google AI Studio. The claim that video and audio are unsupported is also false—Gemini 1.5 Pro has a large context window specifically designed to handle lengthy video and audio files.

VPC Service Controls creates a security perimeter around GCP APIs (including Vertex AI) to prevent data exfiltration and ensure traffic does not leave the defined perimeter. Private Service Connect allows the application's VPC to reach Vertex AI APIs via a private, internal IP address, ensuring traffic never traverses the public internet. Assigning roles/aiplatform.user is about authorization, not network isolation, and does not prevent public internet traversal. Grounding with Google Search is a response-quality feature unrelated to network security. Using GOOGLE_API_KEY is an authentication method for the public Google AI API, not Vertex AI, and would actually route traffic through the public endpoint.