Databricks GenAI Engineer Assembling and Deploying Applications: 212 practice questions

12 of the 212 Assembling and Deploying Applications questions in the Certsqill Databricks GenAI Engineer bank, shown in full below. Each one carries an explanation for every option, not just the correct one — the wrong answers are where the marks go.

Preparing for Databricks GenAI Engineer? Take the free 5-min readiness check →

1. Use on-behalf-of auth with needed scopes and grants: Which implementation meets the requirement?

1. Tell the model to filter records according to each employee's identity.

   Prompt instructions cannot enforce table permissions or turn a shared service identity into the employee's identity.
2. Give the shared service principal broader permissions on every table.

   Broader permissions for one shared identity still cannot apply each employee's individual entitlements and increases exposure.
3. Use on-behalf-of auth with needed scopes and grants. ✓

   Supported on-behalf-of authentication carries the caller identity for downstream authorization when the required scopes and grants are configured.
4. Grant every employee permission to invoke the endpoint.

   Endpoint invocation permission controls who may call the endpoint, not which downstream records each employee may access.

Use supported on-behalf-of authentication with required scopes and grants.

2. The identity lacks required MCP service authorization: What is the most likely diagnosis?

1. The identity lacks required MCP service authorization. ✓

   The evidence directly indicates missing EXECUTE and namespace grants for the managed MCP service.
2. The model needs additional fine-tuning for MCP.

   A 403 with missing MCP grants indicates authorization failure, not insufficient model training.
3. The endpoint invocation permission should automatically authorize every MCP call.

   Endpoint invocation permission does not grant access to downstream MCP services or their governed resources.
4. The managed server must be replaced with a custom MCP App.

   A supported managed service remains suitable; missing authorization does not require rebuilding its hosting.

The agent identity lacks the required MCP service grants.

3. Store checkpoints by user and thread in durable storage: What change provides durable, correctly scoped state?

1. Store checkpoints by user and thread in durable storage. ✓

   Durable user- and thread-scoped storage survives restarts and separates conversations between users and threads.
2. Keep more messages in the dictionary.

   A larger process-local dictionary remains volatile and is not shared reliably across replicas.
3. Put the conversation into the retrieval index.

   A retrieval index stores knowledge representations and is not automatically a transactional session store.
4. Encode each conversation into the model weights for permanent access.

   Model weights are not an appropriate per-user session mechanism for changing conversational state.

Store user- and thread-scoped checkpoints in durable storage.

4. Retrieve authorized context: Which missing step is required?

1. Reserve substantially more output tokens for the response every time.

   Additional output capacity does not provide the missing customer-specific evidence.
2. Generate the answer first, then retrieve documents for citations afterward.

   Post-generation retrieval cannot make an answer grounded in documents the model did not receive.
3. Retrieve authorized context, populate the prompt, then generate. ✓

   The workflow must retrieve authorized context, place it in the prompt placeholder, and generate afterward.
4. Fine-tune the model on current customer documents before each request.

   Fine-tuning does not supply the current customer-specific context required by this request.

Retrieve authorized context, populate the prompt, then generate.

5. Use user authorization with the required scopes and grants: Which authentication approach best fits?

1. Expose personal tokens in browser code.

   Browser-exposed long-lived tokens create unnecessary credential exposure and do not provide safe backend authorization.
2. Use user authorization with the required scopes and grants. ✓

   User authorization can let supported backend operations act with the signed-in user’s identity when the necessary scopes and downstream permissions are configured.
3. Use the App service principal for every tool call.

   A shared service principal authenticates the application but does not automatically enforce each employee’s entitlements.
4. Use endpoint permission and ask the prompt to restrict each employee’s records.

   Endpoint permission does not grant downstream resource access, and prompts cannot enforce authorization.

Use user authorization with required scopes and downstream grants.

6. Pin the exact prompt version and log it with the run: What change resolves the reproducibility defect?

1. Promote the current alias again after comparing the incident output.

   Re-promoting an alias does not restore the historical prompt or identify the version that produced the earlier answer.
2. Pin the exact prompt version and log it with the run. ✓

   Pinning the immutable version and recording it with the run preserves the instruction set used for reproduction.
3. Keep resolving the production alias for each reproduction.

   A mutable alias may resolve to a different prompt version, so repeated reproduction can use changed instructions.
4. Copy the prompt into an untracked notebook before rerunning it.

   An untracked copy may preserve prompt text but loses governed version history and application linkage.

Pin the exact prompt version and record it with the run.

7. Call a Foundation Model API endpoint: Which architecture is the closest fit?

1. Create a Delta Sync AI Search index.

   An AI Search index stores searchable data; it is not the serving architecture for a managed foundation model.
2. Use a prompt registry alias as the inference endpoint.

   Prompt registry aliases manage reusable instructions; they do not themselves provide model inference serving.
3. Register a custom pyfunc model before every request.

   A custom pyfunc deployment adds packaging and lifecycle work that the managed foundation model requirement does not need.
4. Call a Foundation Model API endpoint. ✓

   Foundation Model APIs expose supported managed foundation models through serving endpoints without requiring a custom model package.

Use a Foundation Model API endpoint for managed model inference without custom packaging.

8. Verify the Unity Catalog identity and required privileges: What should it validate next?

1. Move the model alias and retry deployment immediately.

   An alias references a model version but does not grant namespace privileges or automatically resolve deployment authorization.
2. Replace the three-part name with a legacy staging name.

   Unity Catalog uses governed catalog.schema.model identities; legacy stages are not the required namespace validation.
3. Verify the Unity Catalog identity and required privileges. ✓

   Deployment requires the governed model identity and appropriate namespace, creation, and execution permissions.
4. Regenerate the input example with more representative documents.

   Input examples support schema inference and testing but do not resolve a Unity Catalog authorization failure.

Verify the Unity Catalog identity and required privileges.

9. Configure managed embeddings for `body` and `query_text`: Which implementation satisfies these constraints?

1. Compute document and query vectors with a compatible external model.

   This moves embedding generation into application-managed code and violates the stated managed-embedding constraint.
2. Embed `doc_id` and search with equal-dimension vectors.

   An identifier is not the document content intended for semantic embedding, and equal dimensions do not ensure compatible embedding spaces.
3. Create a Direct Vector Access index and upsert application-generated vectors from `body`.

   Direct Vector Access requires application-managed updates and embeddings, which violates the requirement to avoid application-side embedding generation.
4. Configure managed embeddings for `body` and `query_text`. ✓

   The managed embedding configuration can create document embeddings from `body` and query embeddings from `query_text` without application-side generation.

Configure managed embeddings for `body` and `query_text`.

10. Send `question` and `region` using the declared string: Which TWO actions directly diagnose or correct the obs

Select two. More than one option is correct — every correct one is ticked below.

1. Replace the signature with an input example containing only `query`.

   An input example helps infer or test schemas but does not override an explicitly required deployed signature.
2. Move the model to a new endpoint without changing the payload.

   A new endpoint using the same signature will reject the same incompatible request unless its contract changes.
3. Send `question` and `region` using the declared string types. ✓

   A payload matching required names and types addresses the concrete schema mismatch shown in the failure evidence.
4. Inspect the deployed signature and compare it with the request schema. ✓

   Comparing the declared signature with the payload identifies the wrong column name and missing required field causing rejection.
5. Add more policy documents to the search index before changing requests.

   Additional documents may improve retrieval coverage but cannot correct a request whose fields violate the model signature.

Inspect the signature and send the required string fields with matching names.

11. The index is the searchable data structure containing: Which TWO statements correctly interpret the resources?

Select two. More than one option is correct — every correct one is ticked below.

1. The index is the searchable data structure containing document representations. ✓

   The index contains the searchable representations and related search configuration; an endpoint alone does not populate it.
2. The endpoint automatically creates searchable indexes from connected tables.

   An endpoint does not create and populate an index without explicit index and synchronization configuration.
3. The endpoint is the serving boundary used to query configured indexes. ✓

   The AI Search endpoint provides the serving or compute boundary for querying configured indexes.
4. The index is only a permission record and contains no searchable data.

   An index represents searchable data and query configuration, not merely authorization metadata.
5. Endpoint access automatically authorizes access to every index.

   Endpoint and index permissions are distinct, so endpoint access does not grant access to every index.

The endpoint provides search serving; the index provides searchable data and requires configuration.

12. Load reusable resources in `load_context` and have: Which change is best?

1. Create a new serving endpoint for every incoming assistant request.

   Per-request endpoint creation adds deployment overhead and does not provide reusable model-instance initialization.
2. Replace the pyfunc model with a prompt alias without changing serving code.

   A prompt alias manages instructions and does not initialize or cache Python resources required by the serving implementation.
3. Load reusable resources in `load_context` and have `predict` reuse them. ✓

   `load_context` is intended for model-instance initialization, allowing repeated predictions to reuse loaded resources.
4. Move reusable resource initialization into `predict` after validating each request.

   Initializing reusable resources in every prediction preserves repeated startup cost and does not resolve the measured latency problem.

Initialize reusable resources in pyfunc `load_context`, then reuse them from `predict`.

Other Databricks GenAI Engineer domains

• Application Development — 292 questions →
• Design Applications — 136 questions →
• Data Preparation — 133 questions →
• Evaluation and Monitoring — 116 questions →
• Governance — 78 questions →
• All 967 Databricks GenAI Engineer questions →
• Databricks GenAI Engineer certification: requirements, cost and exam format →