Prompt, Retrieve or Fine-Tune?

The situation

The Bedrock prototype is good at general questions and wrong about Code Nexus. Asked about the content-reshare policy, it confidently said Code Nexus allows 'unlimited free resharing of any content'. Code Nexus's actual policy caps resharing another creator's content at three per day, with permission. The team has three ways to fix a model that doesn't know your specifics — write a better prompt, retrieve the real documents, or fine-tune — and Amara wants the right one chosen on evidence, not preference.

What you'll learn

• AIF-C01 generative AI in practice: prompting and context, retrieval-augmented generation (RAG), embeddings and vector search, model adaptation and fine-tuning, evaluation, and choosing the lightest approach that meets a knowledge or style requirement.
• Prompt engineering techniques (zero-, single- and few-shot, chain-of-thought, templates, negative prompts) and prompt risks (injection, jailbreaking, poisoning, exposure); Bedrock Prompt Management; inference parameters; the cost ladder from prompting to pre-training; instruction tuning, RLHF and fine-tuning data; Knowledge Bases and vector stores; MCP and agentic AI concepts; and evaluating a foundation model with ROUGE, BLEU, BERTScore, LLM-as-a-judge, benchmarks and human review.

Who you work with

• 

  Maya Chen

  Senior AI Engineer · Your mentor

• 

  Amara Okafor

  ML and Data Engineer

Scenes

1. 1.It doesn't know our policies
2. 2.Same question, three configs
3. 3.Ground it
4. 4.Same task, six prompts
5. 5.Which evaluation proves it?
6. 6.The house-style request
7. 7.Generative AI check

What you leave with

The reshare-policy gap is fixed with retrieval — the real policy chunk attached as context, the answer correct and cited, the 'unlimited' fabrication gone. The house-style request gets a different answer: prompting first, with example emails and a format checklist to measure against, and fine-tuning held back until that measurement shows prompting isn't enough. Both rationales rest on the comparison evidence, not on preference.