Design Semantic Search - System Design | WinJob

Problem Statement: Meaning-Based Search at Scale

Design enterprise semantic search (Google-scale meaning-based retrieval, Elastic hybrid kNN, Algolia NeuralSearch-class) for Google/Elastic/Algolia-class workloads. This section covers problem statement: meaning-based search at scale in the understanding phase.

Why interviewers probe here

Frame the product as search infrastructure, not a single FAISS notebook demo.

Operational detail

Anchor on 80M documents, 768-d embeddings, 12k peak QPS, p95 query <120ms including embed step.

Failure and edge cases

Cold-start catalog, cross-lingual query, adversarial long queries blowing embed GPU queue.

Interview checkpoints

Checkpoint 1 (understanding): State cosine metric, L2-normalized vectors, and top-50 ANN before rerank. — unique to sec-01.

javaOne Dark Pro

1 public record AnnQuery(float[] embedding, int topK, String tenantId) {
2   public AnnQuery {
3     if (embedding.length != 768) throw new IllegalArgumentException("dim mismatch");
4   }
5 }

pythonOne Dark Pro

1 def reciprocal_rank_fusion(rank_lists: list[list[str]], k: int = 60) -> dict[str, float]:
2     scores: dict[str, float] = {}
3     for ranks in rank_lists:
4         for i, doc_id in enumerate(ranks):
5             scores[doc_id] = scores.get(doc_id, 0.0) + 1.0 / (k + i + 1)
6     return scores

typescriptOne Dark Pro

1 export function l2Normalize(vec: number[]): number[] {
2   const norm = Math.sqrt(vec.reduce((s, v) => s + v * v, 0)) || 1;
3   return vec.map((v) => v / norm);
4 }

Why interviewers care

Semantic Search interviews reward crisp scope, explicit trade-offs, and failure stories—not generic microservice diagrams.

Interview checkpoint

Name one failure story for Problem Statement: Meaning-Based Search at Scale that proves you understand real outages, not happy-path diagrams.

Key Highlights

•Users search by intent, not exact keywords—'laptop for video editing' must match 'mobile workstation GPU'.
•Dual path: offline indexing pipeline + online query orchestration with strict p95 budgets
•Multi-tenant SaaS: 20k catalogs, isolated namespaces, per-tenant embedding model pins
•Interviewers expect hybrid retrieval (BM25 + dense ANN) and measurable recall@k / MRR

1	public record AnnQuery(float[] embedding, int topK, String tenantId) {
2	public AnnQuery {
3	if (embedding.length != 768) throw new IllegalArgumentException("dim mismatch");
4	}
5	}

1	def reciprocal_rank_fusion(rank_lists: list[list[str]], k: int = 60) -> dict[str, float]:
2	scores: dict[str, float] = {}
3	for ranks in rank_lists:
4	for i, doc_id in enumerate(ranks):
5	scores[doc_id] = scores.get(doc_id, 0.0) + 1.0 / (k + i + 1)
6	return scores

1	export function l2Normalize(vec: number[]): number[] {
2	const norm = Math.sqrt(vec.reduce((s, v) => s + v * v, 0)) \|\| 1;
3	return vec.map((v) => v / norm);
4	}