Short answer: for e-commerce product search, use a staged retrieval architecture with explicit collections, bounded queries, and source context that can be traced back to a product record.
At 3 a.m., the useful question is not “is the dashboard green?” It is “which page fired, and can I explain the result a shopper is seeing?” Product search makes that question concrete: price, inventory, and compliance fields change on different clocks, so one global freshness window quietly turns either into stale results or needless re-indexing.
The incident lesson: freshness is a retrieval contract
The production scenario is bounded rather than heroic. A catalog importer writes product documents, an embedding job chunks descriptions, and a vector index serves semantic candidates before a lexical or business-rules stage narrows them. When a price edit arrived during an embedding run, the system had two conflicting truths: the text chunk was recent, while the offer metadata was old.…