Embeddings & Retrieval · 3 to 4 hours
Design hybrid retrieval within a latency budget
Design hybrid retrieval with reranking for a multi-tenant help center, choose a configuration from measured quality and latency gates, and keep access rules intact through timeouts and model migrations.
Hard
System Design
4 h
Premium
1
Step 1: The question
Design retrieval for a multi-tenant help center with exact error codes and natural-language questions. Meet the scenario's quality and latency gates. Explain candidate sizes, fusion, reranking, authorization, timeout behavior, evaluation, and a future embedding migration.
Constraints
- Several organizations share the help center, and every result must be authorized for the requesting user
- Queries mix exact error codes with natural-language questions
- Provisional gates: candidate recall at least 0.90, nDCG@5 at least 0.83, and end-to-end p95 latency at most 250 ms
- Choose among four measured configurations that share one workload and relevance set: a dense baseline, hybrid fusion, hybrid plus reranker, and a wider hybrid plus reranker
- Access rules must hold even when a component fails or times out
Deliverables
- Pipeline with the three sizes named: results per retriever, candidates sent to the reranker, and results returned
- Fusion, deduplication, and reranking choices, and the configuration you select from the measurements
- Authorization placement, cache keys, and timeout behavior for each stage
- Evaluation plan with stage metrics, query groups, and the evidence still missing
- Embedding migration plan, and your decision if the latency gate tightens to 180 ms
Hard questions are part of Premium: the study guide, reference answer, rubric, and exercises unlock with Premium.