direction right

// Client & Security Layer
ClientLayer [label: "Client & Presentation Layer", color: yellow] {
  WebUI [label: "Apple visionOS Frosted Glass UI\n(Live Telemetry & Quality Chip)", icon: browser]
  RabbitVault [label: "3D Interactive Rabbit Vault\n(Three.js + WebCrypto AES-256-GCM)", icon: key]
  RESTClient [label: "REST / SSE Gateway API\n(/v1/rag/query & /v1/rag/stream)", icon: terminal]
}

// Layer 1: Edge Security & Guardrails (<0.5ms)
GuardrailLayer [label: "1. Pre-Inference Guardrails (<0.5ms)", color: red] {
  RateLimiter [label: "Redis Lua Token-Bucket\n(60 / 600 / 3,000 RPM)", icon: shield]
  InjectionGuard [label: "Prompt Injection Blocker\n(100% Interception | 0% FP)", icon: lock]
  PIIRedactor [label: "Luhn + Regex PII Redactor\n(94.7us p95 | ISO/IEC 7812)", icon: eye-off]
}

// Layer 2: Vectorization & C++17 AVX2 Semantic Cache
CacheLayer [label: "2. Vectorization & SIMD Semantic Cache", color: purple] {
  ONNXEmbedder [label: "all-MiniLM-L6-v2 Embedder\n384-Dim L2 Unit Vector (vq)", icon: cpu]
  SIMDCache [label: "C++17 AVX2 Semantic Cache\n(Cosine >= 0.86 | 8.40us p50)", icon: zap]
}

// Layer 3: Advanced 4-Pillar Production RAG Engine
ProductionRAG [label: "3. Advanced 4-Pillar Production RAG Engine", color: green] {
  QueryRewriter [label: "Pillar 1: Query Rewriter\n(Domain Expansion & Policy Acronyms)", icon: edit-3]
  BM25Index [label: "Pillar 2a: Okapi BM25 Index\n(1,467 Terms | k1=1.5, b=0.75)", icon: search]
  AVX2Retriever [label: "Pillar 2b: C++17 AVX2 SIMD\n(80 Dense Chunks | 9.77us)", icon: database]
  RRFFusion [label: "Pillar 2c: RRF Fusion (k=60)\n(0.60 Dense + 0.40 Sparse)", icon: git-merge]
  NeuralReranker [label: "Pillar 3: Cross-Attention Reranker\n(Code Alignment + Window Density)", icon: sliders]
  QualityGate [label: "Pillar 4: Retrieval Quality Gate\n(Sufficiency Score >= 0.35 | Grounding)", icon: award]
}

// Layer 4: Resilient Upstream LLM Router & Storage
UpstreamLayer [label: "4. Multi-Provider Router & Output Guard", color: blue] {
  CircuitBreaker [label: "Sliding-Window Circuit Breakers\n(Trip > 2000ms or > 5% Err)", icon: activity]
  LLMRouting [label: "Dynamic BYOK LLM Router\n(OpenAI / Groq / Anthropic / Gemini)", icon: cloud]
  LocalSim [label: "Deterministic Offline Fallback\n(Zero External Outages)", icon: server]
  OutputGuard [label: "Output Sanitizer &\nNegative-Cache Filter", icon: check-circle]
}

// Connections & Data Flow
ClientLayer.WebUI -> ClientLayer.RabbitVault: "Taps 3D Rabbit (BYOK)"
ClientLayer.RabbitVault -> ClientLayer.WebUI: "Decrypted Client Keys"
ClientLayer.WebUI -> GuardrailLayer.RateLimiter: "Raw Prompt + Headers"
ClientLayer.RESTClient -> GuardrailLayer.RateLimiter: "API Request"

GuardrailLayer.RateLimiter -> GuardrailLayer.InjectionGuard: "Token Allowed"
GuardrailLayer.InjectionGuard -> GuardrailLayer.PIIRedactor: "Clean Intent"
GuardrailLayer.PIIRedactor -> CacheLayer.ONNXEmbedder: "Sanitized Text\n([CARD/SSN_REDACTED])"

CacheLayer.ONNXEmbedder -> CacheLayer.SIMDCache: "Lookup Vector vq"
CacheLayer.SIMDCache -> ClientLayer.WebUI: "CACHE HIT (24.2ms | $0.00)"

// Cache Miss Path -> Production RAG
CacheLayer.SIMDCache -> ProductionRAG.QueryRewriter: "CACHE MISS"
ProductionRAG.QueryRewriter -> ProductionRAG.BM25Index: "Rewritten Query Text"
CacheLayer.ONNXEmbedder -> ProductionRAG.AVX2Retriever: "Reused Vector vq (Zero Double-Emb)"

ProductionRAG.BM25Index -> ProductionRAG.RRFFusion: "Top-20 Sparse Candidates"
ProductionRAG.AVX2Retriever -> ProductionRAG.RRFFusion: "Top-20 Dense Candidates"
ProductionRAG.RRFFusion -> ProductionRAG.NeuralReranker: "Merged Rank List"
ProductionRAG.NeuralReranker -> ProductionRAG.QualityGate: "Top-K High-Precision Chunks"

ProductionRAG.QualityGate -> UpstreamLayer.CircuitBreaker: "Verified Context Chunks\n(Quality Score Evaluated)"
UpstreamLayer.CircuitBreaker -> UpstreamLayer.LLMRouting: "Healthy Circuit (CLOSED)"
UpstreamLayer.CircuitBreaker -> UpstreamLayer.LocalSim: "Tripped Circuit (OPEN)"

UpstreamLayer.LLMRouting -> UpstreamLayer.OutputGuard: "LLM Stream / Completion"
UpstreamLayer.LocalSim -> UpstreamLayer.OutputGuard: "Deterministic Response"

UpstreamLayer.OutputGuard -> CacheLayer.SIMDCache: "Cache Store (If Grounded)"
UpstreamLayer.OutputGuard -> ClientLayer.WebUI: "Stream Response + Quality Telemetry"