Applications of Foundation Models
What this domain tests
Choosing and operating the technique that makes a foundation model useful in production - prompt engineering, RAG, agents, or customization - and proving it works through evaluation.
Core rules
- Prompt engineering is cheapest and fastest: zero-shot, few-shot (examples embedded in the prompt), chain-of-thought (request explicit steps), and structured output templates.
- RAG retrieves relevant chunks from your corpus at query time and inserts them into the prompt. It reduces hallucination and keeps knowledge current without retraining. Pipeline: ingest → chunk → embed → index in a vector store (OpenSearch Serverless, Aurora pgvector) → retrieve top-k → generate. Amazon Bedrock Knowledge Bases manages this flow.
- Agents add tool use: Bedrock Agents invoke action groups (Lambda functions or APIs) to transact, not merely answer.
- Customization: continued pre-training on unlabeled domain text, fine-tuning on labeled examples, and distillation. Use it when style or format must be learned rather than facts supplied.
- Evaluation: automatic metrics (ROUGE, BERTScore, accuracy) plus human review of helpfulness, harmlessness, and faithfulness to sources.
Worked example
A support bot must answer from 40,000 policy PDFs that change weekly and must cite sources. Reasoning: weekly change rules out fine-tuning because retraining cost is recurring; citations require retrieval; 40,000 documents cannot fit a prompt. Choose Bedrock Knowledge Bases with tuned chunk size and overlap, an embeddings model, and a generation model instructed to cite chunk identifiers; add guardrails for denied topics.
Traps
- Reaching for fine-tuning when the real gap is missing context.
- Ignoring chunk size and overlap, which drive retrieval quality.
- Treating RAG as a substitute for clear instructions.
- Launching without an evaluation baseline.