Free IBM Certified watsonx Generative AI Engineer - Associate practice test: 30 questions with answers

Defining the business question, users, and tolerance for error shapes every later decision, including model choice, retrieval design, and evaluation. Colour schemes, GPU purchases, and a pre-chosen model are downstream decisions that should follow from those requirements.

Quality, latency, cost, and data residency requirements decide whether a general model or a smaller specialized one fits the use case. Office size, social media popularity, and user age do not determine the technical fit.

A success metric gives an objective basis for evaluating the solution against business goals, which keeps decisions honest. Pricing, programming language, and training counts are important but do not define success.

Grounding answers in retrieved approved passages and showing their sources makes each answer traceable to a verifiable origin. Relying on general knowledge, removing documents, and inventing citations from memory all break the link to approved sources.

Personal data needs controls on what is sent to the model, where it is processed, and how long it is retained. Upload progress, response length, and font choices do not determine the exposure of personal data.

Prompting can often reach the goal with instructions and examples, while customization becomes worthwhile when consistent style or specialized vocabulary is needed. Logo colour, pronunciation, and printed brochures have no bearing on the technical choice.

Realistic test cases and measurement against expert-reviewed reference answers are what make an evaluation plan informative. Marketing names, seating, and printed results do not test whether the solution answers correctly.

Classifying data decides whether it may go to a model and what handling it needs, which constrains the whole design. Fonts, answer length, and naming rights do not depend on how sensitive the data is.

Prompt tuning keeps the base model frozen and learns a small set of prompt parameters, while fine-tuning updates weights of the model itself. Interface changes, replacing the model, and the data requirements in the other options reverse or misstate these differences.

Fine-tuning is worth its cost when consistent behaviour, format, or specialized language cannot be achieved reliably through prompting. Avoiding instructions, one-time use, and fixed short answers do not justify training a model.

Foundation models are broadly trained and adaptable to many tasks, which is why they are the base for customization. A single-task model, a server rack, and a sample dataset do not match the term.

Testing candidates on the team's own representative tasks with quality, latency, and cost measured gives evidence for the choice. Parameter count, social media popularity, and release date are not reliable substitutes for that evidence.

Supervised tuning learns from input and desired output pairs, so the labels are the examples of the behaviour being taught. Hyperparameter storage, access lists, and invoice records are different artifacts.

Hallucination is output that sounds credible but is inaccurate or unsupported, which is why generated answers need grounding and checks. Interface overlap, backups, and retraining schedules are unrelated to this failure mode.

Improvement on held-out data and a reproducible record of the run are the evidence that supports deployment. Testing only on training examples, naming the model after a sponsor, and one colleague's approval are not valid readiness checks.

Instruction tuning uses instruction and response pairs to improve how the model follows directions. User manuals, registry names, and server clocks are unrelated to training behaviour.

RAG retrieves relevant passages at query time and gives them to the model, which grounds the answer in a source. Training from scratch on random data, printing layouts, and image caches are unrelated to this pattern.

Chunking balances precision in retrieval against enough surrounding context for the answer to make sense. Compressing documents into one number, removing metadata, and print speed are not reasons for chunking.

An embedding model maps passages and queries into a shared vector space, which makes semantic search possible. Encryption, answer generation, and credential storage are separate functions in the system.

Chain-of-thought prompting asks the model to show intermediate reasoning, which often helps on multi-step problems. Shuffled, silent, and compressed prompts do not request step-by-step reasoning.

Chunks that split meaning badly or queries that do not match the document wording lead to weak matches. Monitor speed, text colour, and user volume do not change which passages the retriever finds.

Separating sections makes the template maintainable and makes it harder for user text to be mistaken for instructions. Memory, font, and the claim that structure removes evaluation are not accurate.

Re-ranking puts the most relevant passages first, and evaluating retrieval and generation separately reveals which part is failing. Relying on memory, maximum temperature, and removing citations each weaken grounding or reliability.

Ongoing tracking of quality, latency, cost, and feedback, with alerts, shows whether the application is performing as intended in production. A power light, monthly paper logs, and disabled logs each leave problems undetected.

Guardrails check inputs and outputs against policy and block or rewrite content that is harmful or out of scope. Screen brightness, hardware speed, and password vaults do not control what the model says.

A canary or staged rollout exposes a fraction of traffic to the new version and widens it only when results are acceptable. Big-bang replacement, deleting the old version, and laptop-only testing each remove the controlled exposure.

Governance sets the rules for data use, model approval, risk review, and audit records that keep generative AI use accountable. Furniture, poster colours, and holiday planning are not governance concerns.

An audit trail links an output to the prompt and model version that produced it, which is essential for investigating incidents and for accountability. It does not speed the model, reduce image storage, or replace monitoring.

Pausing or rolling back limits further harm, while investigation and stronger guardrails address the cause before the service resumes. Ignoring alerts, deleting logs, and raising temperature each increase harm or hide the incident.

Access controls limit who can call the application and which data it can reach, and request logging with identifiers supports traceability. Public keys, disabled authentication, and shared administrator passwords each remove a control that protects the service.