nexus

AI Engineering Patterns

Overview

NEXUS — A nexus is the central point where all connections converge. When invoked: assesses system type (RAG / agent / prompt / evaluation), loads the relevant pattern file, and applies AI-specific engineering discipline — hallucination guards, context budgets, injection defenses, cost tracking.

Core principle: LLM applications have unique failure modes — hallucination, prompt injection, context overflow, cost explosion. Engineer systems, not just prompts.

Announce at start: "Running NEXUS for AI application patterns."

---

Entry Point — First 5 Minutes

SYSTEM TYPE ASSESSMENT:

"What are you building/debugging?"
A) RAG / knowledge retrieval system
B) Autonomous agent / tool-using agent
C) Prompt engineering / LLM integration
D) LLM evaluation / benchmarking
E) Multi-agent system
F) Debugging a hallucination / quality problem
G) Cost/latency optimization

Type → Section mapping:

• A → RAG Architecture (see patterns/rag-architecture.md)
• B → Agent Design Patterns: ReAct or Plan-Execute (see patterns/agent-patterns.md)
• C → Prompt Engineering Workflow (see patterns/prompt-engineering.md)
• D → LLM Evaluation Framework (see patterns/llm-evaluation.md)
• E → Agent Patterns: Multi-Agent Debate or Tool Routing
• F → Hallucination Detection + run hunter
• G → Latency/Cost Tracking + vector skill

After identifying type, ask: "What model are you using and what's the context window limit?"

---

RAG Architecture

Load patterns: patterns/rag-architecture.md

Key decisions in order:

1. Chunking strategy — fixed / semantic / recursive / code-aware
2. Embedding model — MiniLM (fast) vs bge-large (quality) vs text-embedding-3-large (enterprise)
3. Retrieval method — lexical (BM25) / semantic / hybrid (RRF) / MMR
4. Re-ranking — cross-encoder reranker from top-20 → top-5
5. Context budget — allocate: system 500 + query 100 + context 6000 + output 1592

Rule: Test retrieval quality (precision/recall) before testing generation quality.

---

Agent Design Patterns

Load patterns: patterns/agent-patterns.md

Pattern	Best For	Iteration Limit
ReAct	Factual QA, tool use	10
Plan-Execute	Multi-step tasks	5 plans
Reflection	Quality-critical output	3 cycles
Multi-Agent Debate	High-stakes decisions	3 rounds
Tool Routing	Multiple specialized tools	N/A

Always set max iteration limits. Agents without limits will loop indefinitely on failure.

---

Prompt Engineering Workflow

Load patterns: patterns/prompt-engineering.md

Process:

1. Write initial prompt
2. Test on 10 diverse examples — categorize failure modes
3. Add constraints / few-shot examples to address failures
4. Re-test until >90% success rate
5. Add output format schema + retry logic (max 3 retries)
6. Add injection defenses (keyword filter + context isolation)

Rule: Never deploy a prompt tested on fewer than 10 diverse examples.

---

LLM Evaluation Framework

Load patterns: patterns/llm-evaluation.md

Metric type	Method	Use when
Exact match	string equality	Factual QA, code gen
F1 score	token overlap	Extractive QA
Semantic similarity	cosine >0.8	Open-ended QA
Rubric-based	LLM grades	Complex tasks
Hallucination	fact verification + self-consistency	High-stakes output
Human preference	blind A/B, win rate >0.55	Model comparison

Cost budgets: p99 latency <5s, cost per 1k requests <$10.

---

Red Flags

Never:

• Deploy RAG without testing retrieval quality
• Use agents without max iteration limits
• Skip prompt injection testing
• Deploy without hallucination monitoring
• Ignore cost monitoring (can explode quickly)

Always:

• Test prompts on diverse examples before deploying
• Include few-shot examples for complex tasks
• Validate output format with schema + retry
• Log all prompts and responses for debugging
• Set up cost alerts before production deployment

---

Integration with Superpowers

Skill	Integration
forge	Write eval tests before model/prompt changes
hunter	Debug hallucination, retrieval failures
sentinel	Verify eval metrics before claiming success
chronicle	Store prompt patterns that worked
vector	Route queries to appropriate model tier

---

Final Checklist

• RAG retrieval quality tested (precision/recall)
• Agent has max iteration limits
• Prompt injection testing passed
• Hallucination detection configured
• Cost monitoring and alerts set up
• Latency budgets defined and met
• Output format validation in place
• Human evaluation plan defined
• Rollback plan if model degrades