Databricks Certified Generative AI Engineer Associate Bundle

Not the strongest diagnose choice for this specific question. Inspect adversarial prompt examples, tool permissions and whether refusals are enforced outside the model. This action primarily relates to Unsafe model output and jailbreak response diagnosis: User-provided instructions may attempt to bypass agent policies or cause data leakage; classification and tool boundaries need separate tests. However, the described case — A RAG assistant passes an internal evaluation set but receives complaints about new policy terms after launch. — is testing Offline evaluation versus ongoing production monitoring. It does not directly establish the diagnose outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest diagnose choice for this specific question. Inspect trace context passages, token distribution, coverage of necessary conditions and retrieval rank. This action primarily relates to Chunk quality evaluation and boundary splitting: Retrieval failure can arise from overly small chunks that omit needed context or overly large chunks that dilute relevance. However, the described case — A RAG assistant passes an internal evaluation set but receives complaints about new policy terms after launch. — is testing Offline evaluation versus ongoing production monitoring. It does not directly establish the diagnose outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest diagnose choice for this specific question. Inspect user conversation state, current intent, entity resolution and selected tool trace. This action primarily relates to Conversation context and intent routing: Multi-turn assistants need explicit conversational state while routing each new request to the correct tool or topic. However, the described case — A RAG assistant passes an internal evaluation set but receives complaints about new policy terms after launch. — is testing Offline evaluation versus ongoing production monitoring. It does not directly establish the diagnose outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest diagnose choice for this specific question. Inspect agent routing to Genie, space permissions, metric definitions and tool-returned values. This action primarily relates to Multi-agent Genie Space tool routing and governed answers: Genie Spaces can provide governed natural-language access to structured business metrics; an agent should choose authorized Genie tools over guessing SQL facts. However, the described case — A retrieved wiki page tells the agent to email confidential documents to an external address. — is testing Prompt-injection defenses for retrieved context. It does not directly establish the diagnose outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest diagnose choice for this specific question. Inspect encoder token limits, chunk length percentiles, truncation counters and source suffix omissions. This action primarily relates to Embedding context window and chunk token distribution: The embedding encoder must accept the actual tokenized chunk length; silently truncating chunks can drop relevant suffix facts. However, the described case — A retrieved wiki page tells the agent to email confidential documents to an external address. — is testing Prompt-injection defenses for retrieved context. It does not directly establish the diagnose outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest diagnose choice for this specific question. Inspect runnable input/output schemas, prompt variables and invocation traces around the failing stage. This action primarily relates to LangChain runnable composition and invocation: Composable runnable chains pass typed values between retrieval, prompt and model stages; each binding must preserve expected input keys. However, the described case — A retrieved wiki page tells the agent to email confidential documents to an external address. — is testing Prompt-injection defenses for retrieved context. It does not directly establish the diagnose outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Build a custom code-based scorer checking required currency and amount fields alongside general quality judges. This action primarily relates to MLflow GenAI tracing, built-in and custom scorers: MLflow GenAI traces capture components, and scorers/judges can evaluate a full answer or individual retrieval and tool spans. However, the described case — A helpdesk agent receives frequent negative user feedback but no monitoring process turns it into reproducible examples. — is testing Inference tables and Agent Monitoring feedback loop. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Do not treat request logging or gateway authentication as proof that underlying data access is authorized. This action primarily relates to AI Gateway rate limits, usage tables and inference telemetry: AI Gateway offers controlled access, supported usage tracking and rate limiting to protect shared endpoints and measure consumption. However, the described case — A helpdesk agent receives frequent negative user feedback but no monitoring process turns it into reproducible examples. — is testing Inference tables and Agent Monitoring feedback loop. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest diagnose choice for this specific question. Inspect changed prompt text, outcome metrics, case distribution and any regressed examples. This action primarily relates to Prompt changes evaluated against baseline: A prompt improvement should be measured against a fixed baseline set, not accepted because a few outputs sound better. However, the described case — A public application generates harmful abusive content after an unusual user prompt despite sanitized inputs. — is testing PII masking and output moderation as complementary guardrails. It does not directly establish the diagnose outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest diagnose choice for this specific question. Inspect adversarial prompt examples, tool permissions and whether refusals are enforced outside the model. This action primarily relates to Unsafe model output and jailbreak response diagnosis: User-provided instructions may attempt to bypass agent policies or cause data leakage; classification and tool boundaries need separate tests. However, the described case — A public application generates harmful abusive content after an unusual user prompt despite sanitized inputs. — is testing PII masking and output moderation as complementary guardrails. It does not directly establish the diagnose outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest diagnose choice for this specific question. Inspect runnable input/output schemas, prompt variables and invocation traces around the failing stage. This action primarily relates to LangChain runnable composition and invocation: Composable runnable chains pass typed values between retrieval, prompt and model stages; each binding must preserve expected input keys. However, the described case — A public application generates harmful abusive content after an unusual user prompt despite sanitized inputs. — is testing PII masking and output moderation as complementary guardrails. It does not directly establish the diagnose outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Keep instructions and examples versioned and test new patterns without allowing retrieved data to override policy. This action primarily relates to Instruction hierarchy and bounded prompt examples: Few-shot examples and explicit boundaries can clarify expected outputs, but embedded source documents should not become new system instructions. However, the described case — An order-status assistant needs real-time package state plus the carrier's public shipping policy. — is testing Parallel retrieval and deterministic chain postprocessing. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Retain audit trails and constrain retries, tool permissions and approval for high-impact actions. This action primarily relates to Agent tool schemas, least privilege and retries: Well-described tool schemas constrain parameters, and idempotent retries prevent duplicate side effects. However, the described case — An order-status assistant needs real-time package state plus the carrier's public shipping policy. — is testing Parallel retrieval and deterministic chain postprocessing. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Do not use a document-answering assistant as an autonomous transaction engine without explicit tools and approvals. This action primarily relates to Agent Bricks Knowledge Assistant suitability: A governed knowledge assistant grounds conversational responses in curated documents and access-controlled retrieval. However, the described case — An order-status assistant needs real-time package state plus the carrier's public shipping policy. — is testing Parallel retrieval and deterministic chain postprocessing. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Extract model/country fields and apply them as verified metadata filters to approved retrieval. This action primarily relates to Query-field driven prompt and retrieval filters: User intent and structured fields such as account_id should drive authorized SQL lookups or retrieval filters, not be ignored in free text. However, the described case — Document chunks average 700 tokens but the selected embedding model only accepts 512 tokens. — is testing Embedding context window and chunk token distribution. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Recheck license and data handling when replacing proprietary with open-weight models. This action primarily relates to Open model license versus managed endpoint obligations: Model availability on a hub does not itself confer commercial training, redistribution or hosting rights. However, the described case — Document chunks average 700 tokens but the selected embedding model only accepts 512 tokens. — is testing Embedding context window and chunk token distribution. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest verify choice for this specific question. Invoke the chain with a representative dictionary and check all intermediate variable mappings and final output. This action primarily relates to LangChain runnable composition and invocation: Composable runnable chains pass typed values between retrieval, prompt and model stages; each binding must preserve expected input keys. However, the described case — An invoice workflow needs exact structured amounts and dates but developers choose an unconstrained chat generator only. — is testing Specialized model versus general LLM. It does not directly establish the verify outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest verify choice for this specific question. Verify matched query and document vectors use the approved version and retrieval improves on held-out examples. This action primarily relates to Embedding model swap and reindex compatibility: Changed embedding dimensions or model semantics typically require re-embedding and index consistency before comparing search quality. However, the described case — An invoice workflow needs exact structured amounts and dates but developers choose an unconstrained chat generator only. — is testing Specialized model versus general LLM. It does not directly establish the verify outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest verify choice for this specific question. Run reproducible pairwise experiments with confidence and segment-level result analysis. This action primarily relates to Model evaluation with paired benchmarks: Accuracy, faithfulness, relevance and refusal behavior require task-specific scoring, and a model switch needs common test conditions. However, the described case — An invoice workflow needs exact structured amounts and dates but developers choose an unconstrained chat generator only. — is testing Specialized model versus general LLM. It does not directly establish the verify outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Enable authorized usage/inference tracking and per-principal attribution with clear quota alerts. This action primarily relates to AI Gateway rate limits, usage tables and inference telemetry: AI Gateway offers controlled access, supported usage tracking and rate limiting to protect shared endpoints and measure consumption. However, the described case — Four reviewers rate the same policy answers inconsistently and the team cannot determine whether prompt changes help. — is testing Human SME feedback rubric and reviewer calibration. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Never label judge outputs as factual correctness when the required reference evidence is absent. This action primarily relates to Ground-truth requirement for correctness versus reference-free judges: A correctness comparison often needs an expected answer; other judges such as format or safety can operate without a reference. However, the described case — Four reviewers rate the same policy answers inconsistently and the team cannot determine whether prompt changes help. — is testing Human SME feedback rubric and reviewer calibration. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest verify choice for this specific question. Run side-by-side evaluations for baseline and candidate prompts with format, groundedness and latency metrics. This action primarily relates to Prompt changes evaluated against baseline: A prompt improvement should be measured against a fixed baseline set, not accepted because a few outputs sound better. However, the described case — A support assistant retrieves a full 60-page manual and overwhelms the model prompt with unrelated sections. — is testing Chunk quality evaluation and boundary splitting. It does not directly establish the verify outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest verify choice for this specific question. Inject adversarial retrieved passages and verify they cannot cause unauthorized actions or data disclosure. This action primarily relates to Prompt-injection defenses for retrieved context: Retrieved content is lower-trust data; embedded instructions must not trigger privileged tool calls or override assistant rules. However, the described case — A support assistant retrieves a full 60-page manual and overwhelms the model prompt with unrelated sections. — is testing Chunk quality evaluation and boundary splitting. It does not directly establish the verify outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest verify choice for this specific question. Run equivalent retrieval tests including storage and refresh cost for candidate encoders. This action primarily relates to Embedding dimensions latency and index storage: Embedding width affects index storage and computation, and larger dimensions do not automatically improve task relevance. However, the described case — A support assistant retrieves a full 60-page manual and overwhelms the model prompt with unrelated sections. — is testing Chunk quality evaluation and boundary splitting. It does not directly establish the verify outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Do not present nearest neighbors as verified answers without evidence validation and permission checks. This action primarily relates to Vector Search retrieval workflow and query integration: Semantic retrieval uses an embedding model to query a compatible index, returns ranked passages and metadata for grounded generation. However, the described case — A new Vector Search index is created in staging but production points at its previous source table version. — is testing CI/CD pipelines for retrieval and agent components. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Do not assert every index feature or hybrid/rerank option exists in every SKU, workspace or region. This action primarily relates to Standard versus storage-optimized Vector Search index: Index choice trades capacity, update speed, latency and storage economics; availability and feature support may vary by cloud/region. However, the described case — A new Vector Search index is created in staging but production points at its previous source table version. — is testing CI/CD pipelines for retrieval and agent components. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest safeguard choice for this specific question. Do not equate registering a model with approving its business safety or access to all governed tables. This action primarily relates to Unity Catalog registered model version and privileges: MLflow-logged models can be registered in Unity Catalog, with governed names, permissions and traceable version lineage. However, the described case — A new Vector Search index is created in staging but production points at its previous source table version. — is testing CI/CD pipelines for retrieval and agent components. It does not directly establish the safeguard outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest implement choice for this specific question. Deploy synchronized index definitions and validate source freshness during production cutover. This action primarily relates to CI/CD pipelines for retrieval and agent components: Production promotions should test source index, prompt, model packaging and tool behavior independently before release. However, the described case — An agent produces raw document objects but the serving client expects a JSON list of cited answers. — is testing MLflow pyfunc packaging for pre- and post-processing. It does not directly establish the implement outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest implement choice for this specific question. Use `ai_query()` against a supported endpoint in a controlled batch-inference query with result checks. This action primarily relates to SQL ai_query for governed batch inference: `ai_query()` invokes an available serving endpoint from SQL; suitability depends on supported runtimes, endpoint and workload. However, the described case — An agent produces raw document objects but the serving client expects a JSON list of cited answers. — is testing MLflow pyfunc packaging for pre- and post-processing. It does not directly establish the implement outcome or constraint that the question requires; it may still be useful in a different stage or incident.

Not the strongest implement choice for this specific question. Bundle compatible index/embedding identifiers and validate query-embedding compatibility at initialization. This action primarily relates to RAG production bundle metadata and dependencies: A deployable RAG application needs tracked model code, retriever config, embedding model, dependency versions, signature and example input. However, the described case — An agent produces raw document objects but the serving client expects a JSON list of cited answers. — is testing MLflow pyfunc packaging for pre- and post-processing. It does not directly establish the implement outcome or constraint that the question requires; it may still be useful in a different stage or incident.