Task 4.2: Create and query a Vector Search index; Explain the key concepts and components of Mosaic AI Vector Search; Configure vector search for a particular solution based on number of embeddings, update frequency, latency, and cost requirements

• AA standard endpoint with a continuous Delta Sync index so the newest catalog data is always searchable

  Incorrect

  600 million 768-dimension vectors exceed what one standard endpoint holds, and continuous sync adds streaming cost that nightly data does not need.

• BA storage-optimized endpoint with a continuous Delta Sync index

  Incorrect

  Continuous sync is not supported on storage-optimized endpoints.

• CA standard endpoint with a target QPS set for high throughput

  Incorrect

  High QPS raises throughput and cost; it does not lift the standard endpoint's capacity limit.

• DA storage-optimized endpoint with a triggered Delta Sync index, synced after each nightly load

  Correct

  Storage-optimized endpoints hold over a billion 768-dimension vectors at lower cost per vector with a few hundred milliseconds of latency, and they support only triggered sync.

Choose standard endpoints when latency is critical and the index fits well within their capacity, and storage-optimized endpoints for 10 million or more vectors when some extra latency is acceptable and cost per vector matters.

• AA Direct Vector Access index is created and updated only through the REST API or the SDK

  Correct

  Direct Vector Access indexes cannot be created in the UI; you upsert and delete vectors yourself through the API or SDK.

• BA Direct Vector Access index refreshes automatically when its Delta table changes

  Incorrect

  Automatic syncing from a source table is the Delta Sync index; direct access indexes are updated by your own calls.

• CA self-managed embedding index can later be switched to managed embeddings

  Incorrect

  Converting a self-managed embedding index to a Databricks-managed one is not possible; a new index is needed.

• DAdding a column to the source table is picked up by the index without a rebuild

  Incorrect

  The index schema is fixed at creation, so schema changes require creating a new index.

• EA self-managed embedding index needs an array<float> vector column and its dimension when created

  Correct

  With self-managed embeddings you name the precomputed embedding column, which must be array[float], and give the embedding dimension.

Delta Sync indexes sync incrementally from a source table; Direct Vector Access indexes are written through the API. Embedding options and the index schema are fixed when the index is created.

• AThey take SQL-like filter strings and filter an overfetched candidate set

  Correct

  Storage-optimized endpoints use a SQL-like filter string, and because results are overfetched and then filtered, matches that are not among the top candidates can be missing.

• BFilters are not supported there, and rare values need full-text queries

  Incorrect

  Filtering is supported on both endpoint types; only the filter syntax differs.

• CDictionary filters need an extra index rebuild, and rare values need reranking

  Incorrect

  No rebuild changes the filter syntax, and a reranker reorders results rather than recovering filtered-out matches.

• DHybrid queries ignore filters, and the product line column was not synced

  Incorrect

  Hybrid queries accept filters, and an unsynced column would fail every filtered query, not only rare ones.

Standard endpoints use a filter dictionary; storage-optimized endpoints use a SQL-like filter string. On storage-optimized endpoints filters are applied to an overfetched set, so a query can return nothing when matching documents score below the top.

• ASwitch each endpoint's indexes to continuous sync mode

  Incorrect

  Continuous sync adds streaming cost; it does not remove base costs.

• BSet a target QPS on every endpoint

  Incorrect

  High QPS provisions extra capacity and raises cost.

• CMove each index to Direct Vector Access

  Incorrect

  The index type does not change the per-endpoint base price.

• DServe all five indexes from one endpoint

  Correct

  One endpoint can serve up to 50 indexes, and combining low-QPS workloads avoids paying several base endpoint costs.

Each endpoint has a base price and scales with the indexes it serves. Combining smaller workloads on one endpoint lowers total cost when QPS is low.

• AImmediately when the index is deleted

  Incorrect

  The documented delay is 24 hours.

• BNever, until the endpoint itself is deleted

  Incorrect

  An empty endpoint stops incurring charges after the delay.

• C24 hours after the last index was deleted

  Correct

  Endpoints are charged only once an index exists, and 24 hours after the last index is deleted the endpoint no longer incurs charges.

• DAt the end of the billing month

  Incorrect

  Billing stops on the documented 24-hour timeline, not monthly.

AI Search bills for indexes and endpoints; an endpoint starts charging once it hosts an index.

• AIn <index name>_writeback_table, to read, not modify

  Correct

  The saved table is named after the index with _writeback_table appended; you can query it but should not drop or modify it.

• BIn the source table, as a new embedding column

  Incorrect

  Saved embeddings go to a separate table, not back into the source.

• CIn a volume, as Parquet files the team may edit

  Incorrect

  The embeddings are saved to a Unity Catalog table, not editable files.

• DNowhere; managed embeddings are never saved

  Incorrect

  Saving generated embeddings is an index option.

The writeback table is created in the same schema as the index, linked from the index page, and deleted automatically when the index is deleted.

• A10,000

  Incorrect

  10,000 is the limit for ANN and full-text queries.

• B200

  Correct

  Hybrid keyword-similarity search returns at most 200 results per query.

• C1,000

  Incorrect

  1,000 is the page size used for pagination.

• D500

  Incorrect

  500 exceeds the hybrid cap of 200 results.

AI Search query limits include 10,000 results for ANN, 200 for hybrid, 1,024 words of hybrid query text and 1,024 elements per filter clause.