Databricks Certified Generative AI Engineer Associate Bundle

Not the strongest safeguard choice for this specific question. Revisit the choice as workload and evaluation data change instead of permanently trusting benchmark ranking. This action primarily relates to Model selection from quantitative task metrics: Select candidate LLMs by measured task success, latency, safety and token cost, not parameter count or leaderboard rank. However, the described case — A production team logs raw access tokens and complete customer chats when adding observability. — is testing Inference logging and production trace privacy. 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 — A production team logs raw access tokens and complete customer chats when adding observability. — is testing Inference logging and production trace privacy. 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 hide expert disagreement by excluding difficult cases or averaging away severe failures. This action primarily relates to Human SME feedback rubric and reviewer calibration: Domain experts should use a shared rubric, calibrate disagreements and maintain adjudicated examples to improve model evaluation. However, the described case — A production team logs raw access tokens and complete customer chats when adding observability. — is testing Inference logging and production trace privacy. 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. Start a clean container smoke test and exercise both old and new client payloads where compatibility is promised. This action primarily relates to Deployed model dependency drift and environment compatibility: Models logged in experiments can fail in serving if Python dependencies or inference signatures differ between build and target. However, the described case — A maintenance assistant queries a vector index by exact serial number without preserving metadata filters. — is testing Vector Search retrieval workflow and query integration. 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. Test model invocation under realistic throughput and verify output schema, quota and governance constraints. This action primarily relates to Foundation Model API versus custom serving choices: Databricks Foundation Model APIs offer managed access to supported models; custom serving is suitable for specialized models or control needs. However, the described case — A maintenance assistant queries a vector index by exact serial number without preserving metadata filters. — is testing Vector Search retrieval workflow and query integration. 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. Load the logged pyfunc model in an isolated environment and test the declared signature on sample requests. This action primarily relates to MLflow pyfunc packaging for pre- and post-processing: A custom MLflow pyfunc model wraps request parsing, retriever/model invocation and output formatting inside a reproducible predict interface. However, the described case — A maintenance assistant queries a vector index by exact serial number without preserving metadata filters. — is testing Vector Search retrieval workflow and query integration. 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 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 — An agent receives a customer message requesting deletion of every table while only ticket lookup is authorized. — is testing Unsafe model output and jailbreak response diagnosis. 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 nested traces, span order, tool duration and captured payload security. This action primarily relates to MLflow tracing in an Agent Framework workflow: MLflow tracing records agent steps, tool calls and latency; integration must retain child spans and inputs without leaking secrets. However, the described case — An agent receives a customer message requesting deletion of every table while only ticket lookup is authorized. — is testing Unsafe model output and jailbreak response diagnosis. 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 — An agent receives a customer message requesting deletion of every table while only ticket lookup is authorized. — is testing Unsafe model output and jailbreak response diagnosis. 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 implement choice for this specific question. Use server-side ownership checks and transaction approval on the refund tool in addition to input screening. This action primarily relates to Prompt safety filtering and tool allowlists: Content filters and action allowlists are complementary; a moderation score is not a substitute for authorization of external effects. However, the described case — A medical appointment chatbot replaces all dates with redaction tokens, making scheduling answers impossible. — is testing Contextual masking versus destructive data removal. 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. Review license and reuse conditions and use authorized sources or link without copying protected text. This action primarily relates to Training and retrieval corpus license rights: Public accessibility does not automatically authorize copying, embedding or using content in commercial RAG services. However, the described case — A medical appointment chatbot replaces all dates with redaction tokens, making scheduling answers impossible. — is testing Contextual masking versus destructive data removal. 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. Filter toxic or instruction-like content and use reviewed trusted source subsets for customer-facing retrieval. This action primarily relates to Toxic, inaccurate and adversarial corpus cleaning: Unreliable source text can teach unsafe behavior through retrieval; quarantine and source review matter before embeddings. However, the described case — A medical appointment chatbot replaces all dates with redaction tokens, making scheduling answers impossible. — is testing Contextual masking versus destructive data removal. 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. Build a relevance-labeled query set and measure recall@k and rank quality on actual returned chunks. This action primarily relates to Retrieval recall precision and ranking diagnostics: Measure hit recall, precision, MRR or nDCG on labeled query-to-relevant-chunk pairs before tuning retrieval. However, the described case — A manuals corpus produces 2.5 million vectors and exceeds the planned index capacity after 120-token chunks. — is testing Chunk size and overlap tradeoffs. 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 document-aware sections plus parent-child metadata to connect symptom to resolution. This action primarily relates to Semantic boundaries versus fixed-window chunks: Headings, paragraphs and tables can require structure-aware splitting rather than arbitrary token windows. However, the described case — A manuals corpus produces 2.5 million vectors and exceeds the planned index capacity after 120-token chunks. — is testing Chunk size and overlap tradeoffs. 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. Parse and chunk documents, write keyed chunk rows with provenance to a governed Delta table, then index them. This action primarily relates to Delta and Unity Catalog pipeline for chunk tables: A governed retrieval corpus stores chunk text, IDs, source lineage and metadata in managed Delta/Unity Catalog tables before indexing. However, the described case — A manuals corpus produces 2.5 million vectors and exceeds the planned index capacity after 120-token chunks. — is testing Chunk size and overlap tradeoffs. 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. Compare both models on the same held-out support questions, citations, cost and safety cases. 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 order assistant answers shipping questions without using the provided account_id, mixing orders across customers. — is testing Query-field driven prompt and retrieval filters. 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. Remove unused side-effect tools and use a plain summarization chain with bounded output. This action primarily relates to Agent tool wrappers versus direct LLM responses: A tool-using agent needs explicit callable interfaces and typed tool results, whereas plain generation requires no action execution. However, the described case — An order assistant answers shipping questions without using the provided account_id, mixing orders across customers. — is testing Query-field driven prompt and retrieval filters. 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. Design a precise instruction on evidence and uncertainty, then compare against baseline using the same evaluation set. 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 — An order assistant answers shipping questions without using the provided account_id, mixing orders across customers. — is testing Query-field driven prompt and retrieval filters. 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 verify choice for this specific question. Run paired representative requests and record quality versus p95 latency and cost tradeoffs. This action primarily relates to Model selection by quality latency and budget: Select the smallest suitable model that meets a task's accuracy and latency constraints under realistic workload, not marketing rank. However, the described case — A proposed LLM wins on general benchmark score but fails grounded FAQ questions more frequently than incumbent. — is testing Model evaluation with paired benchmarks. 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 — A proposed LLM wins on general benchmark score but fails grounded FAQ questions more frequently than incumbent. — is testing Model evaluation with paired benchmarks. 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. Check embedding dimensions or extracted-field validation against task-specific acceptance tests. This action primarily relates to Specialized model versus general LLM: Embedding, reranking, extraction and generative models solve different stages and should be evaluated for their actual role. However, the described case — A proposed LLM wins on general benchmark score but fails grounded FAQ questions more frequently than incumbent. — is testing Model evaluation with paired benchmarks. 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. Never equate a successful embedding API call with a governed and reproducible source-data pipeline. This action primarily relates to Delta and Unity Catalog pipeline for chunk tables: A governed retrieval corpus stores chunk text, IDs, source lineage and metadata in managed Delta/Unity Catalog tables before indexing. However, the described case — A crawler fetches HTML product manuals but outputs menu text instead of article content. — is testing Format-specific document extraction and OCR needs. 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. Separate missing knowledge from bad generation during incident triage and refresh upstream data responsibly. This action primarily relates to Authoritative source coverage for factual answers: A RAG model cannot answer accurately if the required authoritative record is missing or stale in its index. However, the described case — A crawler fetches HTML product manuals but outputs menu text instead of article content. — is testing Format-specific document extraction and OCR needs. 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. Version preprocessing rules and sample exceptions so cleansing never strips legal requirements. This action primarily relates to Boilerplate and navigation filtering in source corpora: Duplicated headers, disclaimers and navigation menus contaminate embeddings and cause irrelevant hits. However, the described case — A crawler fetches HTML product manuals but outputs menu text instead of article content. — is testing Format-specific document extraction and OCR needs. 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. Split read-only diagnostic tools from authorized destructive operations with separate permissions. 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 — A repairs workflow wants parts suggestions but does not state which serial number, failure code or stock fields matter. — is testing Translate use-case requirements into pipeline inputs and outputs. 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. Revisit cost/latency budgets whenever context window or retrieval size changes significantly. This action primarily relates to Latency, cost and retrieval targets as design requirements: A valid pipeline design has measurable quality, latency, token and throughput constraints before model selection. However, the described case — A repairs workflow wants parts suggestions but does not state which serial number, failure code or stock fields matter. — is testing Translate use-case requirements into pipeline inputs and outputs. 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. Keep masking rules under regression control and treat embeddings and logs as potentially sensitive too. This action primarily relates to Sensitive-field masking before embedding: Masking or tokenization can remove unnecessary PII from chunks before embedding and external model calls while retaining useful meaning. However, the described case — A partner PDF includes text telling a RAG assistant to reveal its secret connection string in the answer. — is testing Input guardrails against prompt-injected source documents. 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. Quarantine untrusted or outdated passages, verify document authorship and reindex approved versions. This action primarily relates to Toxic, inaccurate and adversarial corpus cleaning: Unreliable source text can teach unsafe behavior through retrieval; quarantine and source review matter before embeddings. However, the described case — A partner PDF includes text telling a RAG assistant to reveal its secret connection string in the answer. — is testing Input guardrails against prompt-injected source documents. 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. Join authorized structured values and retrieved policy context through a controlled composition stage. This action primarily relates to Parallel retrieval and deterministic chain postprocessing: Combining structured lookup and semantic retrieval requires explicit routing, result joining, and downstream schema checks. However, the described case — A policy assistant produces unsupported answers because its chain sends only the user question to the LLM. — is testing Retriever-generator chain component boundaries. 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. Define a schema-constrained prompt with field types, required keys and strict structured-output validation at the consumer. This action primarily relates to Structured JSON schema prompt and format compliance: A model can produce plausible prose rather than machine-readable JSON unless the output contract and validation are explicit. However, the described case — A policy assistant produces unsupported answers because its chain sends only the user question to the LLM. — is testing Retriever-generator chain component boundaries. 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. Split read-only diagnostic tools from authorized destructive operations with separate permissions. 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 — A policy assistant produces unsupported answers because its chain sends only the user question to the LLM. — is testing Retriever-generator chain component boundaries. 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.