What tutorials show

  • Store embeddings in a vector database, do similarity search, done (7:51).

What production actually needs

Vector similarity alone can't do these (7:51, 8:10):

  • Filter by date — so you only get the latest version of a policy document.
  • Filter by department or document type.
  • Join information across multiple chunks belonging to the same document.

All of that is relational data, and doing it with just vector similarity doesn't work well (8:10).

The conclusion

  • Production systems often use a database that handles both: embeddings for semantic search, relational data for filtering and joining (8:10).
  • Note the direct link back to the parental-leave failure in 02: the stale-version problem is a filtering problem, not a similarity problem.