Software Systems Atlas

Metadata Card

Prerequisites: Chapter 17 (Event-Driven Architecture); Basic LLM concepts
Estimated Time: 45 minutes
Core Difficulty: Advanced
Completion Milestone: Design RAG systems and basic Agent patterns

Your Progress

You've learned all the classic crafts in the City of Artisans: from static inspection to layered assembly, from design patterns to engine splitting.

But in 2026, the City of Artisans has a new craft: intelligent forges — they don't just forge from blueprints, they learn and adjust parameters during forging.

Your engines need new intelligent features: auto-match opponents by skill level, answer player questions in natural language. Impossible with traditional mechanical logic alone. You need to integrate AI into the City of Artisans production line. Your Task

AI systems differ from traditional software: traditional is deterministic (same input → same output), AI is probabilistic (same input → multiple possible outputs). This changes testing, error handling, deployment.

Required reading: RAG architecture, Evaluation patterns Selective: Agent patterns (ReAct) Advanced: Cache-Augmented Generation

First Layer: RAG — Retrieval-Augmented Generation

Don't stuff data into the model — stuff it into a database. Retrieve relevant documents, combine with question, send to LLM.

User question → Retriever → Vector DB
         ↓                     ↑
         └── Retrieved docs ───┘
         ↓
    Combine into Prompt → LLM → Answer

Second Layer: Agent Pattern — LLM That Can Act

ReAct (Reasoning + Acting): LLM can call tools (functions/APIs). Thought → Action → Observation cycle.

Key challenge: LLM may loop infinitely. Set max iteration limit.

Third Layer: Cache-Augmented Generation

Cache LLM outputs for similar questions. Use semantic similarity for cache key.

Fourth Layer: Evaluation Patterns

AI testing can't use assertEquals. Use LLM-as-a-judge: another LLM evaluates output quality. Plus production monitoring: collect user ratings.

java

public String safeAnswer(String question) {
    try {
        return ragService.answer(question);
    } catch (LLMException e) {
        return defaultAnswers.getOrDefault(question, "Please try again later");
    }
}

Common Pitfalls: RAG data freshness, Agent given too many tools, cached results become stale, no fallback when LLM API fails.

Traveler's Notes

AI brings software engineering to a probabilistic world — RAG gives LLM external knowledge, Agents let LLM call tools, CAG uses cache for latency, evaluation makes non-deterministic outputs measurable. In 2026's City of Artisans, AI is becoming a standard module in every system.

Next: API Design & Contract Governance (Chapter 19).