Databricks Certified Generative AI Engineer Associate

Not this time. The 2 questions need different tools. A figure such as last quarter's EMEA returns is an aggregate that has to be computed with SQL over governed tables, which is what a Genie Agent does: it answers natural-language data questions grounded in Unity Catalog data. Policy wording is unstructured text, which is what chunked retrieval is for. Embedding table rows makes the model guess a sum from a handful of similar-looking rows. Fine-tuning goes stale between runs and does not respect table permissions. A whole table in the prompt is costly, hits the context limit and is out of date the next day.

Not this time. On a standard endpoint, a Delta Sync Index follows its source through the table's change data feed, so the source table must have it (tables with row tracking get it automatically). Turning it on with the table property and creating the index again fixes the error. A Direct Vector Access index would avoid the requirement only by giving up syncing: you would then push every change through the REST API yourself. OPTIMIZE compacts files and has nothing to do with it. AI Search needs a Unity Catalog workspace, and indexes are governed by Unity Catalog, so the Hive metastore is no way round.

Not this time. AI Search finds nearest neighbours with HNSW and measures similarity by L2 (Euclidean) distance. Databricks' documentation says that to get cosine similarity you normalise the embeddings first: for unit-length vectors, ranking by L2 distance gives the same order as ranking by cosine similarity. There is no cosine setting to choose, and AI Search does not inspect your model. Hybrid search adds a keyword score, which changes the ranking for another reason and does not turn L2 into cosine.

Not this time. Part numbers are tokens an embedding treats as near-noise: XJ-4411 and XJ-4417 look almost the same to a similarity search. Hybrid search combines the vector search with a BM25 keyword search and merges the 2 rankings with reciprocal rank fusion, so a document that contains the exact string rises. Databricks names SKUs and identifiers as the case it is for. A bigger model still compares meaning, not exact strings. More results costs tokens and still may not include the part. Shorter chunks lose the description that the rest of the query needs.

Not this time. Structured outputs with a JSON schema make the response follow that schema: the fields you require and the values you allow. json_object only asks for some JSON object, of any shape, so a missing field still gets through. Examples and a temperature of 0 make mistakes rarer but guarantee nothing over thousands of rows. Databricks recommends structured outputs for exactly this kind of batch extraction and classification. One catch: Claude 5.5 models on Databricks take the schema through output_config.format in the Messages API, not response_format.

Not this time. A Delta Sync Index with self-managed embeddings reads vectors you have already computed and stays in sync as the source table changes, which is exactly the requirement. A Direct Vector Access index also takes your vectors, but you must update it through the REST API whenever the data changes. Letting Databricks compute embeddings would embed the text again with another model, wasting the team's work, and a self-managed index cannot later be converted to a managed one or the other way. A full-text index answers keyword searches only and throws away the semantic search the embeddings are for.

Not this time. A standard endpoint holds about 160 million embeddings at 1,536 dimensions and about 320 million at 768; capacity scales roughly with the inverse of the dimension. A 768-dimension model fits 200 million vectors, after a test that retrieval quality holds. More overlap and smaller chunks both create more chunks, not fewer. A Direct Vector Access index is far smaller on a standard endpoint, about 2 million vectors at 768 dimensions. A storage-optimized endpoint would hold over a billion vectors, but it does not offer high QPS and adds about 250 ms to each query.

Not this time. AI Functions parallelise, retry and scale a batch themselves, so Databricks recommends submitting the full dataset in one query, and splitting it by hand reduces throughput. With failOnError false, a failed row returns its error message instead of failing the job, so you keep every good result and reprocess only the failures. Parallel notebooks repeat the hand-splitting at a larger scale. AI Functions do not use the compute provisioned to an endpoint; they manage batch scaling themselves. A row-by-row UDF rebuilds badly what ai_query already does.

Not this time. In the MLflow prompt registry, versions are immutable snapshots and aliases are mutable pointers to a version. The agent loads whatever the production alias points to, so moving the alias back to version 7 rolls back every agent that loads it, with no deploy, and keeps version 8 in the history for comparison. A version cannot be edited. Registering a version 9 changes nothing until an alias points to it. Writing a version into the code works, but it is the redeploy the question rules out and gives up the alias workflow.

Not this time. By default an app authenticates with app authorization: a service principal created with the app, whose permissions every user shares. The row filter is evaluated for that service principal, not for the manager. With user authorization the agent acts with the caller's identity, so row filters and column masks are enforced for each user. Put the check in to confirm it: if the forwarded user token is missing, the client falls back to the service principal without an error. A rule in the prompt is not access control and can be talked around. A column mask hides columns, not rows. Granting EMEA only would show every manager EMEA.

Not this time. Correctness compares the response with ground truth you provide in expectations, so without expected answers or facts it has nothing to compare against; RetrievalSufficiency and ToolCallCorrectness need ground truth too. RelevanceToQuery checks the response against the question, RetrievalGroundedness checks it against the retrieved context, and Safety checks the content itself, so all 3 run on inputs and outputs alone. A common plan is to start with those, then have experts write expected facts for a subset and add Correctness.