Domain 3 of 4 · Chapter 1 of 3

Overcoming Model Limitations

Unlock the complete study guide + 1,040 practice questions across 16 full exams.

Bundled into the existing Generative AI Leader premium course — no separate purchase.

14-day money-back guarantee — no questions asked.

Included in this chapter:

  • Why foundation models fail, by root cause
  • Google's recommended fixes, and what each is for
  • Keep it working: monitoring and evaluation
  • How exam questions test model limitations

Limitation to Google-recommended mitigation map

LimitationWhat goes wrongPrimary Google-recommended fixSupporting practice
Data dependencyOnly knows its training data, so it may miss your domainFine-tuning on your own labeled dataData quality and curation upstream
Knowledge cutoffBlind to events after the training snapshotGrounding / RAGUpgrading to newer model versions
Bias and fairnessInherits and can amplify bias present in training dataFairness evaluation and data curationHuman in the loop review
HallucinationConfident but fabricated or unsupported claimsGrounding / RAGPrompt engineering; HITL for high stakes
Edge casesRare or adversarial inputs produce unreliable outputHuman in the loopPrompt engineering and drift monitoring

Decision tree

Which fix for which failureWrong on facts?stale or made-upWrong form?format, tone, stepsNeeds domain fit?or house styleHigh stakes?or rare edge caseGrounding / RAG+ HITL if high-stakesPrompt engineeringFine-tuningHuman in the loopBase model is fineAlways: monitor and evaluate continuouslyYesNoYesNoYesNoYesNo
Cheat sheet

Cheat sheet

  • Foundation-model limits are inherent, so plan mitigations up front
  • A knowledge cutoff blinds the model to anything after its training snapshot
  • Hallucinations are fluent and confident, which is what makes them dangerous
  • Grounding and RAG are the primary fix for stale or fabricated facts
  • A model inherits the quality and bias of its training data
  • Prompt engineering fixes wrong form, not missing facts
  • Fine-tune to lock in domain terms and house style
  • Keep a human in the loop for high-stakes and edge-case decisions
  • No single mitigation covers every limitation, so layer them
  • Reach for the lightest fix first, escalate only as needed
  • Evaluate and monitor continuously; launch is not the finish line
  • Drift and training-serving skew quietly degrade a live model
  • Vertex AI Feature Store serves one feature to both training and serving
  • Tie gen-AI KPIs to the business outcome and trend them over time
  • Version your model use and plan upgrades; models have a lifecycle

Unlock with Premium — includes all practice exams and the complete study guide.

Also tested in

References

  1. What is machine learning?
  2. Responsible AI for generative AI on Vertex AI
  3. Grounding overview (Vertex AI)
  4. Introduction to prompt design (Vertex AI)
  5. Tune generative AI models (Vertex AI)
  6. What is human-in-the-loop (HITL) in AI?
  7. Gen AI evaluation service overview (Vertex AI)
  8. Vertex AI Model Monitoring overview
  9. Vertex AI Feature Store overview
  10. Generative AI model versions and lifecycle (Vertex AI)