| ”Unlabeled data” + “Domain understanding” | Continued Pre-training | Fine-tuning requires labeled data. Pre-training teaches new concepts/language. |
| ”Block emojis” or “Prevent informal language” | Bedrock Guardrails | Guardrails apply safety/style filters (word filters) on top of models. |
| ”Chain LLM calls” + “Least development effort” | Bedrock Flows | Flows is the “visual/low-code” orchestrator. Step Functions is “code/high-effort." |
| "Long-running process (>15 mins)“ | Step Functions | Lambda times out at 15 mins; API Gateway at 29s. Step Functions runs for up to 1 year. |
| ”Strict RAG Sequence” | Ingest → Embed → Index → Generate | You can’t search what you haven’t indexed, and you can’t index what you haven’t embedded. |
| ”Nested sections” or “Parent-child relationships” | Hierarchical Chunking | Preserves the context of headers (e.g., “Chapter 1” → “Section A”). |
| ”Multimodal” or “A/V recordings” in RAG | Bedrock Data Automation (BDA) | BDA specifically handles parsing video/audio for RAG. |
| ”Runtime updates” + “No redeployment” | AWS AppConfig | AppConfig acts as a dynamic feature toggle for routing logic. |
| ”Reduce token usage” + “Maintain accuracy” | Reranking | Fetch many results (Recall) → Rerank → Send only top 5 (Precision). |
| ”Standardize legacy/messy APIs” for Agents | Model Context Protocol (MCP) | MCP is the standard “driver” layer to abstract API complexity for agents. |
| ”Financial records” + “Vector Search” | Aurora PostgreSQL (pgvector) | Financial data requires ACID compliance (strong consistency), which SQL provides. |
| ”Graph” or “Citation networks” | Neptune Analytics | ”Graph” always points to Neptune. |
| ”Chat History” + “Low latency” | DynamoDB + DAX | DynamoDB handles scale; DAX handles the millisecond caching needed for chat. |