GenAI Architecture: Architecture for Generative AI Systems

Generic, brand-neutral engineering capability; deep, sourced playbooks live in references/ and knowledge/. [DOC]

Generic, brand-neutral engineering capability; sourced playbooks in references//knowledge/. [DOC]

TL;DR

GenAI architecture defines how LLM-powered systems retrieve knowledge, orchestrate models, execute agent workflows, and ensure quality. This skill produces comprehensive architecture documentation covering RAG design, LLM orchestration with multi-model tiering, agent workflows, vector database selection, knowledge connector integration, and quality assurance for generative AI systems [EXPLICIT]

When to Use

• Designing RAG architecture for enterprise knowledge systems
• Selecting and configuring vector databases for embedding storage and retrieval
• Designing LLM orchestration with multi-model tiering and routing
• Architecting agent workflows with tool use, memory, and guardrails
• Integrating structured knowledge connectors (CRM, ERP, ITSM)
• Planning quality assurance for GenAI (hallucination reduction, grounding, continuous improvement)
• Evaluating CAG vs. RAG vs. hybrid approaches

When NOT to Use

• Traditional ML model architecture (non-generative) -> ai-software-architecture
• CONOPS and operational concept -> ai-conops
• Data pipelines and ML CI/CD -> ai-pipeline-architecture
• General AI design patterns (Feature Store, Drift Detection) -> ai-design-patterns
• Testing strategy for AI systems -> ai-testing-strategy
• Infrastructure provisioning for GenAI -> infrastructure-architecture

---

Sub-capabilities (resource map)

Deep, evidence-tagged playbooks — open the one the task needs (ICM Layer 3, on-demand). [INFERENCE]

Procedure

1. Resolve the sub-capability; open the matching references/ playbook. [EXPLICIT]
2. Apply its decision tables; pick the strategy explicitly. [EXPLICIT]
3. Validate against the Quality Criteria and tag every claim. [EXPLICIT]

Quality Criteria

• Sub-capability resolved to one playbook. [INFERENCE]
• Claims evidence-tagged. [EXPLICIT]

Contract

• Aceptación: capability resolved to its reference playbook, applied, validated, evidence-tagged. [EXPLICIT]
• Límites: · Focuses on GenAI architecture, not general ML architecture (see ai-software-architecture) · Does not design traditional ML pipelines (see ai-pipeline-architecture) -. [EXPLICIT]
• Casos borde: Enterprise with Strict Data Residency: Cloud-managed vector DBs may not meet data residency requirements. Self-hosted vector DB (Qdrant, Milvus) with region-specific deployment. [EXPLICIT]
• Supuestos: · Use case benefits from generative AI (not all AI problems need LLMs) · Knowledge sources are identified and accessible · LLM API access is available (cloud provider or self-hoste. [SUPUESTO]
• Trade-off: Decision Enables Constrains When to Use --- --- --- --- Advanced RAG High retrieval quality, production-grade Complex pipeline, more infrastructure Enterpri. [EXPLICIT]

Packet

Capas del packet, cargables bajo demanda (disciplina ICM: una capa por vez, nunca todas juntas): references/ guías de profundidad (cargar UNA por etapa) · knowledge/ cuerpo de conocimiento · prompts/ prompts listos · examples/ salida de ejemplo · agents/ subagentes del packet · assets/ recursos estáticos.