How to conduct a complete system design interview for a multi-tenant enterprise RAG search platform?

Short answer: Conduct an enterprise RAG system design interview by outlining an end-to-end multi-tenant architecture: API Gateway with RBAC pre-filtering, document ingestion pipeline with vision layout parsing, hybrid vector/sparse search with cross-encoder reranking, vLLM serving with PagedAttention and prefix caching, and continuous Ragas evaluation telemetry.

Detail

This advanced scenario tests a candidate's ability to synthesize all major sub-systems of enterprise AI engineering into a single coherent blueprint.

[Enterprise Web Clients]
           │
           ▼
[API Gateway] ──► RBAC Tenant Token Validation (OAuth 2.0)
           │
           ├──► [Redis Semantic Cache] (Hit -> Return Instant Response)
           │
           ▼ (Cache Miss)
[Hybrid Retrieval Engine]
 ├── Sparse Search (BM25 for exact SKU/part codes)
 ├── Dense Vector Search (Qdrant with RBAC Pre-Filtering)
 └── Cross-Encoder Reranker (Cohere Rerank v3)
           │
           ▼
[vLLM Inference Cluster] (PagedAttention + Chunked Prefill + Prefix Cache)
           │
           ▼
[Output Guardrail & Telemetry] ──► DeBERTa NLI Entailment Check + OpenTelemetry Spans

Complete Sub-System Checklist

1. Ingestion Pipeline: Parsing PDFs/DOCX using layout parsers, embedding via Matryoshka models, attaching RBAC tenant metadata.
2. Retrieval Pipeline: Hybrid BM25 + HNSW vector search merged via Reciprocal Rank Fusion (RRF), passed to cross-encoder reranker.
3. Inference Pipeline: Disaggregated prefill/decoding clusters running vLLM with PagedAttention and prefix caching.
4. Governance & Observability: Reversible PII redactor, OpenTelemetry tracing, and Ragas evaluation benchmarks.

Example

Whiteboard architecture blueprint summary:

1. Security Layer: PII Redaction Proxy + Tenant Isolation RBAC Pre-Filter.
2. Search Layer: Hybrid BM25 + Dense Vectors merged via RRF (k=60) + Cross-Encoder Rerank.
3. Generation Layer: vLLM Inference Engine with PagedAttention & Speculative Decoding.
4. Quality Layer: NLI Entailment Hallucination Guardrail + Continuous Ragas Evaluation.

Interview tips

• Cover every phase of the architecture systematically without skipping security or observability.
• Highlight scaling bottlenecks (e.g. KV cache VRAM pressure under peak concurrent user load) and how PagedAttention solves them.

Related Concepts

• [[What is document metadata filtering in vector databases?]] (#123): What is document metadata filtering in vector databases?
• [[How does Reciprocal Rank Fusion (RRF) combine scores from sparse and dense retrievers?]] (#127): How does Reciprocal Rank Fusion (RRF) combine scores from sparse and dense retrievers?
• [[How does vLLM PagedAttention manage Virtual Memory Pages to eliminate KV cache fragmentation?]] (#159): How does vLLM PagedAttention manage Virtual Memory Pages to eliminate KV cache fragmentation?

⬅ Back to Interview Experience · All topics