NCA-GENM - Software Development (15% of the exam) - Section 4.4
Customise NVIDIA AI Blueprints.
Understand how NVIDIA AI Blueprints provide reference architectures for common generative AI workflows. Customise a Blueprint by swapping models, adjusting pipeline parameters, or integrating domain-specific data to meet a specific application requirement.
NVIDIA AI Blueprints
Practice question for this objective
lock_openFree sampleSoftware Developmentmedium
A team is adapting an NVIDIA AI Blueprint that uses an agent workflow to automate multi-step research tasks. They want the agent to call an internal knowledge-base tool rather than a public search API. Which customisation approach is most consistent with how agent blueprints are designed?
AReplace the underlying large language model NIM with a smaller model that has been fine-tuned on internal documents so tool calls are not needed.
BEnable automatic speech recognition on the agent input layer so that voice queries bypass the public API and route directly to internal systems.
CRegister the internal knowledge-base as a new tool definition within the agent configuration, pointing the tool executor at the internal endpoint instead of the public API.check_circle Correct
DWrap the entire blueprint in a diffusion-based summarisation layer that condenses public API results to match the internal knowledge format.
Describe how to customise the tool configuration of an NVIDIA AI Blueprint agent workflow to use an alternative data source. NVIDIA AI Blueprint agent workflows delegate task steps to tools that are described in a registry available to the orchestrating LLM. Each tool entry specifies a name, a natural-language description the model uses for selection, and an endpoint the executor calls. Swapping a public API for an internal knowledge-base requires updating the tool definition so it points to the new endpoint; the agent reasoning and the rest of the blueprint remain unchanged. This is the composable, low-friction customisation model that blueprints are designed to support.
Why A is wrong: Fine-tuning the base model encodes static knowledge but does not replace the benefit of dynamic tool retrieval; it also discards the agent's ability to call multiple tools and is disproportionate to the goal of swapping one tool endpoint.
Why B is wrong: ASR adds a multimodal input modality and is unrelated to which retrieval tool the agent calls; it does not change how the agent resolves tool endpoints.
Why C is correct: Agent blueprints expose a tool registry where each callable tool is described by a schema and an endpoint. Adding or replacing a tool definition is the intended customisation path: the agent LLM selects tools by name and description, so updating the registry entry redirects calls to the internal service without changing the rest of the workflow.
Why D is wrong: Diffusion models generate images or audio and are not applicable to text summarisation; this option conflates multimodal generation techniques with agent tool customisation.
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-GENM bank for this domain.
The blueprint is a frozen model checkpoint, so the only supported adaptation is fine-tuning its weights on new data before redeployment.
errorWhy it is wrong: Tempting because fine-tuning is a real adaptation technique, but it is wrong here: a blueprint is a multi-component workflow, not a single checkpoint, and most adaptation is reconfiguration rather than weight tuning.
It merges every model in the workflow into one process, so a single restart applies any change uniformly across the whole pipeline.
errorWhy it is wrong: Tempting because a single restart sounds convenient, but it is wrong: a microservice approach separates models into independent services rather than merging them into one process.
Blueprints compile all services into a single monolithic binary, so changing the model endpoint recompiles only the inference path and leaves other code paths verified.
errorWhy it is wrong: Tempting because it sounds like a safe build-time guarantee, but blueprints are composed of independently deployed microservices rather than one monolithic binary, so this description is inaccurate.