Domain 1 of 4 · Chapter 3 of 3

Apply generative AI concepts and techniques

Unlock the complete study guide.

Bundled into the existing AWS Certified AI Business Strategist premium course — no separate purchase.

14-day refund guarantee on one-time purchases and your first subscription charge.

Included in this chapter:

  • Three layers that shape a generative AI response
  • Writing prompts that state task, evidence, and format
  • Examples, templates, and iterative prompt refinement
  • Tokens, context windows, and generation limits
  • RAG: supplying evidence at response time
  • Fine-tuning and other ways to customize a model
  • Choosing and evaluating an adaptation
  • Applying GenAI techniques to business gaps

Prompt engineering vs RAG vs fine-tuning: which layer each changes

Decision factorPrompt engineeringRAGFine-tuning
Layer it changesThe request (written instructions, evidence, examples)The request (evidence retrieved automatically)The trained parameters
Gap it closesWrong task, format, audience, or styleMissing private or frequently changing knowledgePersistent behavior or style that prompts did not fix
What you must supplyClear instructions, relevant evidence, examplesAuthoritative, current, permission-controlled sourcesReliable examples (or quality feedback) of the desired output
When facts changeEdit the evidence in the promptRefresh the sources and their index; updates can be available in minutesRetrain, which can take hours to days
Source citationsCan cite supplied sources when promptedCan return references to the retrieved sourcesNo built-in reference to sources
Effort and skillsLowest; no trainingManaged options such as Amazon Bedrock Knowledge BasesTraining and evaluation work, often a data scientist

Decision tree

Needs private orchanging facts?YesRAGNoTask, format, oraudience wrong?YesPromptengineeringNoBehavior gap remains;training evidence?YesFine-tuningNoGather trainingevidence firstAlways: evaluate on the same cases

Cheat sheet

  • A prompt should name the exact task and any allowed answer choices
  • Relevant reference context supplies information that the task instructions alone cannot provide
  • Specify the output shape and limits when the business needs a particular deliverable
  • Audience and role context guide how an answer is framed rather than establish its truth
  • Source boundaries and an uncertainty response help keep answers within available evidence
  • Temperature influences response variation rather than factual correctness
  • Few-shot prompting demonstrates a desired pattern without fine-tuning the model
  • Prompt templates separate reusable instructions from changing business inputs
  • Complex requests can benefit from explicit intermediate tasks and dependencies
  • Useful refinement identifies a specific deficiency rather than asking for arbitrary regeneration
  • A tested prompt template needs reevaluation when its underlying model changes
  • The context budget includes more than the user's visible document
  • Missing input and a cut-off response point to different token constraints
  • A conversation's visible history is not automatically available in every model request
  • Long-content handling should preserve the evidence required by the particular task
  • Fitting more content into context does not guarantee that every detail will be used well
  • RAG supplies relevant external knowledge without retraining the answering model
  • A RAG answer can only be as well grounded as the evidence retrieved for it
  • RAG can provide source attribution, but citations still need to support the answer
  • Retrieval must respect access permissions before private information reaches the model
  • Semantic retrieval finds related evidence, while generation turns evidence into an answer
  • Fine-tuning adapts learned behavior rather than simply attaching reference documents
  • A fine-tuning proposal needs suitable examples and a justified training commitment
  • RAG and fine-tuning can address different gaps in the same business application
  • An adaptation succeeds only if evidence shows that it closes the intended business gap
  • Model distillation transfers demonstrated capability to a smaller model rather than refreshing its knowledge sources
  • Domain-language adaptation can learn from specialist text without paired ideal answers
  • Reinforcement fine-tuning adapts a model using feedback on response quality

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

Also tested in

References

  1. https://aws.amazon.com/what-is/retrieval-augmented-generation/
  2. https://docs.aws.amazon.com/bedrock/latest/userguide/design-a-prompt.html
  3. https://docs.aws.amazon.com/aws-certification/latest/ai-business-strategist-01/ai-business-strategist-01-domain1.html
  4. https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-engineering-guidelines.html
  5. https://docs.aws.amazon.com/bedrock/latest/userguide/inference-parameters.html
  6. https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html
  7. https://aws.amazon.com/what-is/prompt-engineering/
  8. https://docs.aws.amazon.com/nova/latest/userguide/prompting.html
  9. https://docs.aws.amazon.com/nova/latest/userguide/prompting-tools-rag.html
  10. https://docs.aws.amazon.com/nova/latest/userguide/prompting-focus.html
  11. https://docs.aws.amazon.com/connect/latest/adminguide/upgrade-models-ai-prompts-agents.html
  12. https://docs.aws.amazon.com/wellarchitected/latest/generative-ai-lens/genperf01-bp01.html
  13. https://aws.amazon.com/blogs/security/context-window-overflow-breaking-the-barrier/
  14. https://aws.amazon.com/blogs/machine-learning/break-the-context-window-barrier-with-amazon-bedrock-agentcore/
  15. https://docs.aws.amazon.com/wellarchitected/latest/generative-ai-lens/gencost01-bp01.html
  16. https://aws.amazon.com/blogs/machine-learning/amazon-bedrock-agentcore-memory-building-context-aware-agents/
  17. https://docs.aws.amazon.com/prescriptive-guidance/latest/retrieval-augmented-generation-options/rag-vs-fine-tuning.html
  18. https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
  19. https://docs.aws.amazon.com/prescriptive-guidance/latest/retrieval-augmented-generation-options/what-is-rag.html
  20. https://aws.amazon.com/bedrock/evaluations/
  21. https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html
  22. https://aws.amazon.com/blogs/security/authorizing-access-to-data-with-rag-implementations/
  23. https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-contextual-grounding-check.html
  24. https://docs.aws.amazon.com/aws-certification/latest/ai-business-strategist-01/ai-business-strategist-01-domain3.html
  25. https://docs.aws.amazon.com/nova/latest/nova2-userguide/nova-cpt.html
  26. https://docs.aws.amazon.com/nova/latest/userguide/nova-forge-sft.html
  27. https://docs.aws.amazon.com/bedrock/latest/userguide/reinforcement-fine-tuning.html
  28. https://docs.aws.amazon.com/bedrock/latest/userguide/model-distillation.html