This appendix summarizes the AI architecture patterns from Chapters 11, 12, 13, 15, and 16.
Generative AI and Agentic Patterns
| Pattern | Description | Use when | Essential controls |
|---|---|---|---|
| Model-only generation | Application sends a request to a model and returns the result | Task needs no enterprise-specific knowledge | Data minimization, output review where needed |
| Retrieval-augmented generation | Application retrieves governed enterprise content as context | Answers must be grounded in current enterprise knowledge | Permission-aware retrieval, source authority, citations, evaluation |
| Structured retrieval | APIs or queries obtain precise current values | Questions need balances, statuses, or counts | Authorization at the service, validated queries |
| Graph-assisted retrieval | Knowledge graph relationships guide retrieval | Answers depend on multi-hop relationships | Entity resolution, relationship governance, provenance |
| Tool-enabled assistant | Model requests approved tools through an orchestration layer | Task requires interaction with enterprise services | Narrow tools, schema validation, read/write separation |
| Workflow-embedded AI | AI is embedded within an existing process | AI should operate within established controls | Process ownership, audit, exception handling |
| Human-in-the-loop | AI drafts or proposes; a person approves | Judgment or authority must remain human | Meaningful review with evidence and time |
| Bounded agent | Agent plans and executes multi-step tasks within limits | Multi-step tasks with clear scope and authority | Agent identity, step and time limits, stopping conditions, idempotency, approval for high-impact actions |
| Decision service | Reusable service combines rules and models | Many applications need consistent decision logic | Versioned rules and models, decision logs, outcome monitoring |
RAG Design Checklist
- Sources are approved, owned, and classified for the intended use.
- Extraction preserves tables, headings, and other meaningful structure.
- Chunking preserves the context needed to interpret each passage.
- Indexed content carries source identity, version, status, effective dates, and access-control metadata.
- Retrieval combines vector, keyword, and structured filtering as the questions require.
- User permissions are enforced before content reaches the model.
- Retrieved content is treated as untrusted evidence, never as instruction or authority.
- Context construction preserves source identity and qualifications, and avoids redundant passages.
- The application acknowledges when evidence is insufficient or conflicting.
- Retrieval quality, answer grounding, and security are evaluated separately, before release and in production.
- Updates, deletions, and index rebuilds are managed as lifecycle operations.
- Traces connect each response to its query, retrieved sources, prompt version, and model configuration, with protected logging.