3.4. Reflection Checkpoint

Key Takeaways

• Vector store selection is a multi-dimensional decision: query volume, metadata filtering needs, update latency requirements, and team expertise all drive the choice between OpenSearch, Aurora pgvector, and Bedrock Knowledge Bases.
• Bedrock Knowledge Bases is not real-time — sync jobs complete in minutes to hours. Architectures requiring immediate document availability need event-driven sync or custom OpenSearch ingestion.
• Chunking is a precision/recall trade-off; optimal strategy depends on document structure and query pattern. Semantic chunking costs extra FM invocations at ingestion time.
• Changing your embedding model requires re-indexing the entire corpus — it's not a zero-downtime operation without a blue/green index strategy.
• Hybrid search outperforms pure semantic for entity-specific queries; pure semantic search is optimal for purely conceptual queries. Add a reranker for highest-quality retrieval.
• Prompts are governed artifacts: they need versioning (Bedrock Prompt Management), testing against golden datasets, approval workflows, and CloudTrail audit logging.
• IAM controls who can call Bedrock. Guardrails controls what content flows through Bedrock. Both are required independently.

Connecting Forward

Phase 4 covers Domain 2: the implementation and integration patterns that connect FM capabilities to real-world applications. This includes the agent architectures (Bedrock Agents, Strands, MCP), enterprise integration patterns (CI/CD, GenAI gateways), and the FM API patterns (streaming, resilience, intelligent routing) that the exam tests at Professional difficulty.

Self-Check Questions

• A company indexes 100,000 documents using Titan Embeddings v1 (768 dimensions). AWS announces Titan Embeddings v2 (1024 dimensions) with significantly better retrieval benchmarks. Walk through the complete migration plan, including how you'd validate quality improvement before switching production traffic.
• You're reviewing a RAG system's retrieval logs and notice that queries about "employee termination procedures" consistently retrieve chunks about "contract termination clauses" instead. What is the likely cause, and what are three possible fixes at different levels of the retrieval stack?