← Back to Registry Hub/Site Search/L3 · capstone_project
⚡Flashcard Drill➔
L3S2 · Data & RetrievalS5 · Eval & LLMOpsS6 · Security & FinOpsVerified: 2026-11-17

Capstone: Multi-Tenant RBAC, Incremental CDC & Semantic Cache RAG Hub

Portfolio-grade enterprise RAG capstone solving the three hardest production failures probed in senior interviews: zero-leak ACL pre-filtering in HNSW, sub-60s CDC deletion propagation, and tenant-isolated semantic caching.

Best For

Backend, Data, and AI Engineers proving production readiness across multi-tenant security, hybrid search precision, and token FinOps.

Human-AI Boundary

Identity groups originate solely from upstream IdP claims; retrieval filters enforce strict read-only intersections.

Anti-Patterns to Avoid

  • Caching embeddings globally by cosine similarity alone without prefixing `tenant_id` and `rbac_group_hash`
  • Applying RBAC permission checks post-top-K retrieval, leaving restricted minority teams with zero results

Latency/Cost Profile & Eval Metrics

Semantic cache hit < 38ms P95; Cache-miss pre-filtered hybrid RAG P95 < 1.1s · 48%–62% token reduction via tenant-partitioned semantic & radix prefix caches

  • Zero Cross-Tenant Leakage & 1% Selectivity Recall@10: 0 unauthorized chunks across 500 RBAC probes; Recall@10 >= 0.91 on 1% tenant slices (Automated CI suite runs cross-tenant isolation probes on every merge)

💰 Business ROI Model & Evidence Level

Serves 2M chunks across 50 enterprise workspaces—cutting monthly embedding and inference bills from ~$6,200 to ~$1,450 while passing zero-leak ACL penetration tests.

🏗️ End-to-End Architecture Pipeline

Step 1: Content-Hash CDC & Signed ACL Indexing Pipeline

Deduplicate chunks via SHA-256 content manifest and push atomic ACL revocations into Qdrant within 60s.

Step 2: Tenant-Scoped Semantic Cache + Pre-Filtered Hybrid RRF + Quota Breaker

Check tenant+ACL-keyed semantic cache first; on miss, run pre-filtered dense+sparse RRF under Redis TPM quota enforcement.

📊 Benchmark & Alternative Comparison

  • Pre-Filtered HNSW vs Post-Filtered HNSW at 1% tenant selectivity: Recall@10 improves from 0.24 to 0.93 without increasing p95 latency beyond 35ms.

📦 Project Deliverables & Portfolio Artifacts

  • End-to-End FastAPI + Qdrant + Redis Lua TPM Limiter + CDC Sync Worker Reference Codebase
  • 500-case Cross-Tenant RBAC Penetration Test & Recall@10 Regression CI Benchmark Script

🎤 Interview Defense & System Design Deep-Dive Q&A

  1. Why does post-filtering break completely when an enterprise tenant holds only 0.5% of documents, and how does filterable HNSW maintain graph connectivity?
  2. If two employees in different security clearances ask the exact same question, why must the semantic cache still miss between them?
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.
🎯 Drill This Card Now➔

Bidirectional Taxonomy Links

Internal Tools & Primary Sources