Full-Stack GenAI Applications

A production GenAI application is not just a model call. It is an end-to-end system spanning prompt engineering, retrieval infrastructure, backend orchestration, and a user-facing interface. This page provides orientation across the full stack.

The GenAI Application Stack

Key Architecture Decisions

Decision	Options	Key trade-off
Model hosting	Hosted API vs self-hosted open-weight	Cost/latency vs data sovereignty/control
Knowledge integration	RAG vs long context vs fine-tuning	Freshness vs cost vs specialisation
Agentic vs non-agentic	Single-shot vs agent loop	Simplicity vs capability for complex tasks
Streaming	Streaming vs batch	UX responsiveness vs implementation complexity

2025 Defaults

By 2025, the defaults for most production GenAI applications have settled:

RAG as the standard knowledge integration pattern (see data augmentation)
MCP as the standard tool integration protocol
Streaming responses as the default UX pattern
Prompt caching for long, repeated system prompts (80–90% input token cost reduction)
LangGraph or OpenAI Agents SDK as the orchestration layer for agentic workflows

Where to Go Next

Back-end infrastructure — model serving, LLMOps, database choices
Agents — agentic patterns, frameworks, and MCP/A2A protocols
Data augmentation — RAG, fine-tuning, and long-context patterns
Security and governance — compliance, safety, and access control

Commercial Products

Libraries And Tools