1.5. Reflection Checkpoint

Key Takeaways

• Foundation models predict probable token sequences — they don't recall facts. Every architecture that requires factual accuracy must add a grounding mechanism (RAG) or output validation.
• Amazon Bedrock = managed FM API invocation. Amazon SageMaker = custom model training, fine-tuning, and hosting. They are complementary, not competing.
• RAG fixes the knowledge problem (domain-specific or recent data). Fine-tuning fixes the behavior problem (style, format, tone). Agents fix the action problem (taking steps in external systems).
• The six Well-Architected pillars apply to GenAI but with different risk profiles — especially security (prompt injection), reliability (non-determinism), and cost (token economics).
• Every PoC-to-production transition requires adding: error handling, guardrails, monitoring, prompt governance, and cost controls.

Connecting Forward

Phase 2 moves from conceptual models to concrete implementation — how to select, configure, and deploy foundation models for specific business requirements. You'll apply the mental models from Phase 1 to real architectural decisions: which Bedrock model to choose, when fine-tuning makes sense, and how to build resilient FM invocation patterns.

Self-Check Questions

• A company wants to build a support bot that accurately answers questions about its internal product documentation (updated weekly) and also can raise support tickets in Jira. Which architectural patterns does this require, and what AWS services implement each?
• Without looking, explain the difference between temperature and top-p sampling. For a legal document summarization use case, what values would you choose for each, and why?