A search index answers queries that the source database can't serve cheaply, such as full-text relevance with filters. It's a derived view.
A product edit can commit before the index consumes the update, so a search result may temporarily show an old title or a deleted product. Decide what delay the interface allows.
Store stable document IDs and versions in the index. Replaying an update should replace the same document, and a delayed event must not resurrect a deleted record. For sensitive details, fetch the current authoritative record before acting on a search hit. You'll build the indexing path, measure update backlog, and plan a new index version that can be tested before an alias changes over.