Memory hook: Ground the answer; constrain the action; measure both.
Must remember
Google's current Gemini Enterprise Agent Platform evolves Vertex AI capabilities into an integrated model/agent platform. Older course and API terminology may still say Vertex AI. Model Garden offers model choices; managed APIs, customization and custom training address different levels of control and effort. Choose a prebuilt API when its capability fits rather than training unnecessarily.
Model choice depends on quality, modality, latency, context, cost, data handling and deployment constraints. RAG retrieves external knowledge at request time; tuning changes model behavior/weights through supported training. Grounding can improve factual relevance but does not guarantee correctness or enforce authorization by itself.
Agent systems add planning, tools, memory and actions. Keep tool authority narrow, authenticate workload identities, validate inputs/outputs and gate high-impact operations. Treat retrieved documents and tool results as untrusted content. Model Armor and Sensitive Data Protection can contribute filtering and sensitive-data controls; application policy and evaluation remain necessary.
ML pipelines orchestrate repeatable data preparation, training, evaluation and deployment. Track datasets, features, experiments and artifacts to reproduce results. Separate training from evaluation data to avoid leakage; monitor production drift and quality. GPUs/TPUs and AI Hypercomputer infrastructure fit different training/serving requirements; expensive hardware alone does not solve bad data or inefficient inference.
Managed search/conversation, vision, document, image, video and audio APIs reduce implementation work for suitable tasks. Gemini Enterprise and NotebookLM-related capabilities support enterprise knowledge workflows under their actual access/governance model. Check supported data locations, quotas, retention and integration rather than assuming all products share one policy.
Evaluate task success, groundedness, safety, latency and cost on representative and adversarial cases. Human review belongs where incorrect output/action has high consequences. Version prompts, retrieval configuration, models and tools as production changes.
Choose under exam pressure
| Requirement | Choice and reason |
|---|---|
| Answers need current private knowledge | Permission-aware RAG with evaluation and source grounding. |
| Need a known vision/document capability | Assess a managed API before custom training. |
| Agent can update business records | Scoped tools, authorization, validation and auditable approval boundaries. |
Traps
- RAG does not automatically prevent unauthorized document disclosure.
- An AI-generated infrastructure proposal still needs human/automated verification.
Active recall
1. RAG versus tuning?
RAG adds retrieved context at request time; tuning changes supported model behavior through training.
2. Why version prompts and retrieval settings?
They can change production behavior as materially as code.
3. What is training/evaluation leakage?
Information from evaluation targets improperly influences training or selection, overstating expected performance.
4. Why constrain agent tools?
A mistaken or manipulated model should not gain unrestricted action authority.
5. Does a safety filter guarantee factual correctness?
No. Safety and factual/task quality require distinct evaluation.