Applications of Foundation Models
The service map
- Amazon Bedrock: a fully managed, serverless API to foundation models from multiple providers plus Amazon models, with Knowledge Bases, Agents, Guardrails, and model evaluation.
- Amazon SageMaker: the full ML platform to build, train, tune, and deploy your own models; JumpStart offers open models and recipes.
- Amazon Q: generative assistants for business users and for developers.
- PartyRock: a no-code playground for experimenting with Bedrock models.
Selecting a model
Compare candidates on task quality, supported modality (text, image, audio, embedding), context length, latency, cost per token, and customization options. Exam items usually state constraints: the lowest-latency option for short classification points to a smaller model; image understanding requires a multimodal model.
Prompt engineering techniques
Zero-shot instructions, few-shot examples inside the prompt, chain-of-thought prompting for multi-step reasoning, role or persona framing, explicit output-format constraints, and prompt versioning and management. Prefer positive instructions over long lists of prohibitions.
RAG and agents
Retrieval-augmented generation embeds documents, stores vectors, retrieves the most relevant chunks for a query, and injects them into the prompt. Bedrock Knowledge Bases is managed RAG. Agents let a model call tools and APIs to complete multi-step tasks and take actions.
Traps
- Fine-tuning is the wrong answer when the requirement is knowledge that changes frequently; use RAG.
- Knowledge Bases does not retrain the model; nothing is baked into the weights.
- Guardrails filter and constrain content; they are not a general accuracy upgrade by themselves, though contextual grounding checks help.
- Bedrock gives managed access to foundation models; SageMaker is where you build custom models end to end.
Worked example
A support team wants answers grounded in 10,000 help articles that change weekly, with citations, and minimal ML operations. The answer is Bedrock Knowledge Bases for managed retrieval plus Guardrails, not fine-tuning, because weekly change makes retraining impractical.