Case study

Contoso Retail — semantic product search

Contoso Retail is replacing keyword product search with a semantic search service for a catalogue of 4 million SKUs.

The ingestion pipeline embeds each product description with a 1536-dimension model and writes the vector alongside the product record. Catalogue updates arrive continuously from three regional merchandising teams, and roughly 40,000 products change on a normal weekday.

The storefront queries the service on every keystroke after a 200 ms debounce. Peak traffic is 12,000 queries per second on promotional days, and 95% of those queries hit the same 2,000 popular products. Product records are read far more often than they are written.

The search API must return results in under 150 ms at the 99th percentile. Search is regarded as revenue-critical: a regional outage must not take the storefront down.

Requirements

Vector similarity search over 4 million embeddings
Sub-150 ms p99 read latency for the storefront
Continuous ingestion without blocking reads
Survive the loss of a single Azure region
Control cost — the team flagged RU consumption as a concern

Q45 Cosmos DB Domain 2 · 25–30% of exam Priority · Real Exam Choose 2

Which TWO changes most directly reduce RU consumption for this workload?

A Exclude the 1536-dimension embedding array from the indexing policy (correct answer)
B Enable the integrated cache with a dedicated gateway for repeated reads (correct answer)
C Increase provisioned throughput
D Switch every read to strong consistency
E Add a second write region

Shortcut: Indexing a 1536-element array on every write is a large and pointless RU cost, since the vector is searched by the vector index, not the property index. With 95% of traffic hitting 2,000 products, the integrated cache serves repeats at 0 RU. Raising throughput increases cost; strong consistency doubles read RU; a second write region adds cost and conflict risk.

Which TWO changes most directly reduce RU consumption for this workload?

More Cosmos DB questions