Summary

  • 3 minutes

You set out to give the Contoso product knowledge base a search box that answers both the shopper who types helmet and the shopper who types something to protect my head, without adding a second data store to the architecture. Azure Cosmos DB for NoSQL does both natively, against the container that already holds the catalog.

What you learned

  • Full-text search needs a container full-text policy naming the searchable paths and a matching full-text index, and it ranks results with BM25 through FullTextScore in an ORDER BY RANK clause.
  • A container vector policy fixes the vector path, data type, dimension count, and distance function at creation time, so those four decisions belong in the design rather than in a later migration.
  • Vector index type follows the size of the search scope and the dimension count, with flat capped at 505 dimensions, quantizedFlat suiting scopes up to roughly 50,000 vectors, and diskANN scaling beyond that scope.
  • VectorDistance ranks items by similarity, and a TOP N clause, a projection of only the properties you need, and a partition key filter are what keep the request charge proportionate.
  • The change feed refreshes embeddings when source text changes, and a stored content hash both prevents needless model calls and stops the consumer from repeatedly processing its own writes.

Learn more

Clean up resources

If you completed the exercise and no longer need the resources, delete the resource group it created to stop all charges.