Free SAP Certified Associate - Generative AI Developer practice test: 30 questions with answers

Tokens are the units a model reads and generates, and they drive usage and limits. Embeddings represent meaning, hyperparameters are training settings, and checkpoints save training states.

The context window is the amount of text, in tokens, that the model can consider in one call. The other options describe a browser window, a call window and a data collection period, none of which is the context window.

Hallucination is generated content that is plausible but not supported. Embedding, quantisation and tokenisation are technical processes, not the name for this problem.

RAG retrieves relevant passages at answer time and adds them to the prompt, which avoids retraining. Pre-training, temperature and parameter counts do not give access to documents at answer time.

An embedding represents meaning as a vector, so similar texts can be found by distance. Encryption, compression and print formatting are different purposes.

Fine-tuning adapts behaviour or style with training examples. Current documents are better served by retrieval, and fine-tuning does not replace prompts or context windows.

Few-shot prompting includes example input and output pairs before the real request. Zero-shot gives no examples, and random or silent prompting do not provide guidance.

The system prompt sets the role, constraints and format that the model follows. The user message carries the question, the completion is the output, and an embedding is a vector representation.

The generative AI hub in SAP AI Core provides governed, centralised access to foundation models without spreading provider credentials into each application. Storing keys in ABAP tables, browser calls and copying weights into the database all weaken governance.

A deployment configuration names the model, version and resource group used for calls. A tile, a transport request and a user group do not identify the model for API calls.

A reusable prompt template with defined variables keeps prompts consistent and centrally managed. Copies, screenshots and comments drift apart and cannot be governed.

Temperature controls randomness, and lower values make outputs more repeatable. Output length, context window and retrieved documents control different aspects.

Content filtering with category thresholds applied to the call filters harmful content in a governed way. Prompts, temperature and context size do not filter outputs reliably.

Data masking replaces personal identifiers before the prompt is sent, which limits exposure. Trusting the provider, deleting the data and base64 encoding do not protect it, since encoding is not a privacy control.

Pinning production to a tested version and testing new versions separately keeps behaviour stable. Automatic upgrades, untracked redeployments and daily switches make results unpredictable.

Masking sensitive fields before prompts and restricting retrieved content to authorised users both limit leakage. Logging full personal data, sending whole databases and sharing production credentials all increase exposure.

A resource group isolates AI resources and their configurations for a tenant or team. Fiori tiles, source code and working hours are not what resource groups manage.

A deployment is a running instance of a configured model or workflow that exposes an endpoint. Transports, accounts and printed reports are not deployments.

A scenario groups executables and their configuration into a reusable definition. Tiles, tables and notifications are not AI Core workflow definitions.

Service credentials obtained through OAuth client credentials, and rotated, avoid stored passwords. Shared passwords, shared administrator accounts and anonymous endpoints each weaken security.

Testing the deployment with representative data against agreed criteria is essential before promotion. Direct promotion, deleting development versions and copying production credentials are unsafe.

A resource group isolates configurations and a configuration names the scenario and executable. Printed documentation, training attendance and colour themes are not part of the deployment setup.

Declaring the parameter in the scenario template and supplying it through a new configuration keeps the workflow reproducible. Editing weights, tile labels and private notes do not define a workflow input.

Joule is SAP's generative assistant integrated into SAP applications. Spreadsheet macros, batch emails and database indexes are not embedded assistance.

Human review limits the impact of incorrect or unfair outputs on high-impact decisions. The other options are wrong about system capability, speed and regulation.

Showing the basis and source data lets users verify the recommendation. Hiding the basis, presenting it as fact and random confidence numbers all reduce trust and transparency.

Documented testing with approval and auditable logs support acceptance in regulated processes. Releasing to all users at once, unreviewed master data changes and hiding model use from auditors all undermine compliance.

Reviewing and confirming a proposed transactional change keeps the human in control and respects authorisation. Immediate writes, unchecked authority and hidden changes each remove protection.

Applying the user's own authorisations to retrieved data prevents exposure. A shared all-access user, reliance on users and unchecked shared files each expose data.

A risk assessment with mitigations and sign-off comes before launch. Marketing first, skipped testing and post-launch complaint collection are unsafe sequencing.