Free NVIDIA-Certified Professional: Agentic AI Sample Questions

Correct answer: B — Injecting the graph schema, including node labels, relationship types, and property keys, directly into the LLM's system prompt.

• A. Providing a complete dump of all graph data is impractical as it would likely exceed the context window and introduce significant noise. Grounding the model in raw instance data is far less effective for query generation than providing an explicit, structured schema.
• B. Injecting the graph schema (node labels, relationship types, and property keys) is the standard best practice for grounding LLMs in Text-to-Cypher tasks. It provides the authoritative set of constraints needed to ensure the model uses existing relationship names and property keys, thereby minimizing hallucinations and ensuring the output aligns with the database's architecture.
• C. Adjusting temperature and top-p can increase the determinism of the output, but these parameters do not provide the factual grounding or structural knowledge required to generate syntactically correct queries against a specific database schema.
• D. Converting a graph database into a relational SQL schema is counterproductive. It removes the graph-specific semantics that Cypher is designed to express and does not address the fundamental requirement of mapping natural language to the existing graph structure.

Correct answers: B, D — Using collapsible UI components (e.g., accordions) that summarize tool calls (e.g., 'Searched the web') but allow expansion for detailed logs.; Providing a visual progress indicator or timeline mapping the agent's completed and pending sub-tasks.

• A. Displaying raw JSON payloads is confusing for non-technical users and causes cognitive overload. It is a technical transparency pattern that fails to communicate the agent's intent or reasoning effectively.
• B. Collapsible UI components follow the principle of progressive disclosure. They provide a concise, human-readable summary of what the agent is doing (e.g., 'Analyzing data') while allowing curious or advanced users to expand and view the underlying execution details.
• C. Full system prompts and token counts are low-level implementation details. They do not help a user understand the logic of a specific task and typically clutter the interface.
• D. A visual progress indicator or timeline maps the agent's multi-step process in an intuitive way. It communicates status, sequence, and progress at a glance, which is highly effective for building user trust without requiring technical knowledge.
• E. While 'Human-in-the-Loop' (HITL) is important for high-risk actions, requiring approval for every individual API call creates excessive friction and makes the interaction cumbersome, which is counter-productive to the goal of a seamless user experience.

Correct answer: B — Agents communicate exclusively through asynchronous message passing and encapsulate their own private state.

• A. Incorrect. The Actor Model explicitly avoids a single globally shared mutable state to prevent contention and race conditions. Instead, state is isolated within individual actors, ensuring data consistency through ownership rather than global access.
• B. Correct. A defining characteristic of the Actor Model is that actors (agents) encapsulate their own private state and interact with others only by sending and receiving asynchronous messages. This design allows for high concurrency and scalability in distributed systems without the overhead of shared memory.
• C. Incorrect. The Actor Model is inherently decentralized and event-driven. It does not require a central orchestrator to impose a strict sequential order; actors process messages as they arrive, enabling concurrent and independent execution.
• D. Incorrect. Distributed locks and semaphores are mechanisms used for managing access to shared memory. Because the Actor Model relies on private state and message passing, it eliminates the need for such synchronization primitives, thereby avoiding common pitfalls like deadlocks.

Correct answer: B — Use a recursive summarization buffer that periodically condenses older turns while keeping the most recent turns verbatim.

• A. A strict LRU eviction policy essentially acts as a sliding window, removing the oldest tokens as new ones arrive. In a multi-day troubleshooting scenario, this would lead to the loss of critical early diagnostic steps, which are necessary for the agent to maintain a coherent understanding of the problem's history.
• B. A recursive summarization buffer (such as ConversationSummaryBufferMemory) allows the agent to maintain the chronological flow by compressing older parts of the conversation into concise summaries while keeping the most recent interactions in their original verbatim form. This effectively manages the token limit while retaining essential context from earlier in the session.
• C. Retrieving nodes from a graph database is a retrieval-based (RAG) approach that focuses on semantic relevance rather than chronological sequence. This strategy is better for long-term knowledge retrieval but does not solve the problem of maintaining a coherent, sequential conversation history within a limited context window.
• D. Increasing the temperature only makes the model's outputs more stochastic or creative; it does not expand the context window or recover lost memory. Hallucination is a failure state in customer service and is not a valid strategy for memory management.

Correct answer: B — gRPC

• A. MQTT is a lightweight, asynchronous publish/subscribe messaging protocol designed for low-bandwidth or unreliable networks (IoT). It is not optimized for strict synchronous request-response negotiation or the sub-10ms latency requirements of high-frequency trading systems.
• B. gRPC is a high-performance RPC framework that uses Protocol Buffers (Protobuf) for binary serialization and strict type-checking. It operates over HTTP/2, offering low-latency, synchronous communication and multiplexing, making it the ideal choice for high-speed, structured agent-to-agent negotiation.
• C. Apache Kafka is a distributed event streaming platform optimized for high-throughput, asynchronous data persistence and replayability. Because it involves writing to a broker/log, it introduces higher latency overhead than direct RPC and is not suited for the synchronous, sub-10ms interaction required here.
• D. HTTP/1.1 REST is synchronous but typically relies on text-based formats like JSON, which lack native strict type-checking and have higher serialization overhead. It generally incurs higher latency and less efficient resource utilization compared to the binary-based HTTP/2 protocol used by gRPC.

Correct answer: A — Implement a state variable to track the iteration count and add a conditional edge from the Reviewer to a terminal node if the threshold is exceeded.

• A. Correct. Tracking the iteration count (recursion limit) in the state and adding a conditional transition to a terminal node is the standard way to prevent unbounded cycles in stateful orchestration. This preserves the multi-step workflow while ensuring the agent terminates after a reasonable number of retries when progress is impossible.
• B. Incorrect. Replacing the cyclic graph with a linear prompt chain would remove the iterative review-and-revise behavior required for multi-step reasoning. While it ensures termination, it compromises the core architectural requirement of high-quality, reviewed output.
• C. Incorrect. Increasing temperature may make the output more varied but does not resolve the logic gap of missing source evidence. This often leads to hallucinations rather than solving the loop, and it does not provide a formal safeguard for termination.
• D. Incorrect. Clearing state memory removes the context needed for the reviewer and summarizer to coordinate and learn from previous steps. This does not stop the loop; it merely makes the agent 'blind' to its previous failures, often leading to repetitive cycles without resolution.

Correct answers: B, C — Implement a two-stage retrieval pipeline: a fast semantic search followed by a cross-encoder reranker to improve the precision of the top 5 tools.; Fine-tune the embedding model specifically on the descriptions and expected arguments of the 500 API tools.

• A. Increasing top-k to 200 would overwhelm the LLM's context window, significantly increasing latency and cost. Moreover, it introduces a 'lost in the middle' effect where the model fails to identify relevant information among too many distractors, which would likely degrade precision further.
• B. Implementing a two-stage retrieval pipeline utilizes a bi-encoder (semantic search) for efficient candidate generation followed by a cross-encoder for high-precision reranking. Cross-encoders score query-tool pairs more accurately than vector similarity alone, ensuring the most relevant tools are placed in the prompt.
• C. Fine-tuning the embedding model on domain-specific documentation (API descriptions and arguments) improves the representation space. This helps the retrieval system better understand the nuances of the 500 tools, leading to higher recall (finding the right tool) and precision (excluding the wrong ones).
• D. Prompting an LLM to guess endpoint URLs from pre-trained knowledge is unreliable and bypasses the retrieval mechanism designed to ground the agent in the actual available toolset. This approach is highly prone to hallucinations, especially for internal enterprise APIs.
• E. Deterministic keyword-based engines are too brittle for a large library of 500 tools. They cannot handle natural language variations, synonyms, or complex user intents effectively compared to learned retrieval and LLM-based reasoning.

Correct answer: A — Add a system prompt instruction to output ONLY valid JSON and implement a framework-level output parser with automatic retry logic that feeds parsing errors back to the LLM.

• A. This approach is the industry standard for robust agentic workflows. It combines prompt engineering (system instructions) with structural guardrails (framework-level parsing) and an error-recovery loop (retry logic). Feeding the specific parsing error back to the LLM allows it to correct the formatting drift in real-time, ensuring Step 2 receives valid input.
• B. Lowering the presence penalty is a sampling parameter adjustment that might slightly reduce token variety, but it is not a reliability mechanism. It does not guarantee strict JSON adherence or prevent the model from generating markdown blocks or conversational preambles.
• C. Smaller models often have weaker instruction-following capabilities compared to larger models. Moving to a smaller model is more likely to increase formatting errors rather than solve them, and it fails to address the lack of a validation or retry mechanism.
• D. Using regular expressions to 'guess' JSON boundaries is highly brittle. It can easily fail with nested objects, trailing whitespace, markdown code blocks (e.g., ```json), or malformed output. It is not a robust architectural solution for high-stakes agentic chains.

Correct answer: A — Implement a client-side fallback to a cached response or a simpler, local heuristic model when the deadline is exceeded.

• A. Correct. A client-side fallback to a cached response or a simpler local heuristic model is the best graceful degradation strategy. This keeps the agent functional even when the server cannot respond before the deadline, and it strictly adheres to the constraint of not modifying the Triton server configuration.
• B. Incorrect. The `max_queue_delay_microseconds` is a server-side parameter defined in the model's `config.pbtxt`. It cannot be dynamically modified via a client-side gRPC header to influence the server's scheduling logic.
• C. Incorrect. Sending requests in a tight loop without backoff (a retry storm) is not a graceful degradation strategy. It can exacerbate server congestion, increase latency for other users, and worsen the overload situation.
• D. Incorrect. Dynamic batching is a model-specific configuration that resides on the server. Modifying this configuration via client APIs (if possible) would violate the constraint of not modifying the Triton server configuration and usually requires a model reload.

Correct answer: A — Implement a circuit breaker and exponential backoff retry logic directly within the custom tool's Python execution code.

• A. Implementing resilience patterns like circuit breakers and exponential backoff retry logic within the tool's execution code is a standard best practice. This programmatic approach ensures the agent handles transient API failures and rate limits gracefully, preventing crashes and respecting the third-party provider's constraints without relying on the LLM's non-deterministic behavior.
• B. Instructing the LLM via prompt or tool description to manage timing is unreliable. LLMs are not built to enforce precise execution delays or track real-time API state, making this approach prone to failure and continued 429 errors.
• C. A semantic cache can reduce the total number of requests by serving cached responses for similar queries, but it does not address the underlying issue of how the system should respond when a 429 error is actually triggered. It is a optimization strategy, not a robust error-handling mechanism.
• D. Switching LLM providers is an inappropriate response because the bottleneck is the third-party API, not the reasoning model. The failure occurs in the tool integration layer, so changing the LLM will not resolve the rate limit issue on the external system.

Correct answer: B — False Positive Rate (FPR) of the topical rail

• A. Topical Jailbreak Success Rate measures how often the agent fails to block restricted topics (false negatives). This is used to assess the security and robustness of the rail against adversarial prompts, rather than quantifying the issue of over-blocking harmless inputs.
• B. The False Positive Rate (FPR) of the topical rail measures how often the guardrail incorrectly identifies and blocks benign, allowed queries. Since the agent is over-triggering and refusing general market trend questions, the FPR is the specific metric needed to quantify and tune the rail's sensitivity.
• C. Context Precision evaluates the relevance of information retrieved from a knowledge base in a RAG pipeline. While important for overall system quality, it does not track or explain the triggering behavior of a topical guardrail.
• D. Faithfulness (or groundedness) measures whether the model's response is supported by the retrieved context to prevent hallucinations. It does not address why an agent might be refusing to generate a response due to excessive guardrail sensitivity.

Correct answer: A — To ensure the automated metrics accurately proxy human preferences, allowing for scalable and reliable offline evaluation.

• A. Calculating the correlation coefficient (Pearson for linear relationship or Spearman for rank order) is the standard method to validate that an automated evaluation metric aligns with human judgment. High correlation indicates the automated judge is a reliable proxy for human preferences, enabling rapid, scalable, and cost-effective offline evaluation during development cycles.
• B. Correlation analysis is a statistical measure of relationship strength between two existing datasets. It does not generate new data. Synthetic data generation requires a separate pipeline involving prompting or data augmentation techniques.
• C. Hyperparameters like learning rate and batch size are optimized through training experiments and empirical testing. They are unrelated to the statistical alignment between feedback scores and evaluation metrics.
• D. While LLM-as-a-judge is naturally faster than human review, the purpose of a correlation coefficient is to assess the quality and accuracy of the judgment (alignment), not to measure or prove the speed of the computation.

Correct answers: B, C — Maximizing Tensor Parallelism (TP=8) minimizes TPOT, assuming high-bandwidth NVLink is available.; Tensor Parallelism reduces latency by splitting matrix operations across multiple GPUs simultaneously.

• A. Maximizing Pipeline Parallelism (PP) usually increases per-token latency because a token must pass through multiple pipeline stages sequentially. While PP helps in fitting large models across GPUs and can improve throughput via pipelining multiple requests, it introduces 'bubbles' and sequential dependencies that increase TPOT.
• B. Tensor Parallelism (TP) is the most effective strategy for reducing latency within a single node. By splitting the computation of individual layers across all 8 GPUs, the time required to compute each token is significantly reduced. This is highly dependent on high-speed interconnects like NVLink to handle the frequent synchronization required between GPUs.
• C. This is the fundamental mechanism of Tensor Parallelism. By partitioning large matrix multiplications (sharding weights), the GPUs work in parallel on the same operation for the same token, directly lowering the compute time per token (TPOT).
• D. Pipeline Parallelism does not process multiple layers simultaneously for a single token; the token must still traverse stages 1 through N in sequence. It allows different tokens or sequences to be processed in a pipeline fashion to improve throughput, but it does not reduce the latency of a single request.
• E. Data Parallelism replicates the model across GPUs and processes different batches or requests independently. While this is excellent for scaling total throughput (Total Queries Per Second), it does not accelerate the processing time for a single individual request.

Correct answer: C — The LLM judge's prompt lacks specific cultural or regional rubrics, leading to a misalignment with human evaluators.

• A. Catastrophic forgetting is a phenomenon observed during model training or fine-tuning where new information overwrites old knowledge. It is not relevant to static evaluation tasks where the model's weights are not being updated.
• B. Position bias refers to a model's tendency to favor options based on their placement in a sequence (e.g., preferring the first item in a list). While a known issue in LLM-as-a-Judge setups, it does not explain a consistent bias against a specific geographic dataset unless that dataset is always evaluated in a specific position, which is not suggested here.
• C. This is the most likely cause. LLMs often have implicit biases based on their training data. If the evaluation prompt does not provide specific cultural or regional rubrics, the judge may apply a generic or North American-centric standard to European interactions, missing nuances that human evaluators correctly identify as high quality.
• D. While LLMs can exhibit length bias, they typically suffer from 'verbosity bias,' where they provide higher scores to longer responses. There is no standard behavior where higher token counts automatically result in lower scores unless the prompt explicitly penalizes length.

Correct answer: B — Increase the similarity threshold (`threshold`) for the canonical user messages in the topical rail.

• A. Decreasing the temperature setting affects the randomness and determinism of the model's text generation. It does not directly influence the matching logic or vector similarity thresholds used by topical rails to classify and block user inputs.
• B. Topical rails often rely on vector similarity to match user inputs to canonical forms. Increasing the similarity threshold makes the matching criteria more stringent, ensuring that only queries with a high degree of similarity to the blocked topic trigger the rail, thereby reducing false positives.
• C. Switching from streaming to batch mode changes the delivery and processing mechanism of the response (latency vs. throughput) but does not impact the classification logic or the decision-making process of the guardrails.
• D. Removing the `generate_user_intent` flow would disable a fundamental component of the guardrail system's ability to classify inputs. This would likely cause the system to fail to identify blocked topics altogether rather than providing a targeted optimization for false positives.

Correct answer: B — Implement a Kubernetes Event-driven Autoscaling (KEDA) ScaledObject configured to monitor the Redis queue length.

• A. Incorrect. While HPA can scale pods based on CPU utilization, it is often suboptimal for event-driven or queue-based agent workloads. Task backlogs can accumulate even when CPU utilization is low (e.g., if agents are waiting on external API responses), leading to reactive rather than proactive scaling.
• B. Correct. KEDA (Kubernetes Event-driven Autoscaling) is designed to scale workloads based on external event sources like Redis. By monitoring the queue length directly, KEDA can scale agent pods horizontally in response to the actual workload backlog, ensuring efficient handling of sudden spikes.
• C. Incorrect. VPA (Vertical Pod Autoscaler) adjusts the resource requests and limits of existing pods rather than the number of pods. It is better suited for right-sizing resources over time and often requires pod restarts, making it ineffective for handling rapid bursts in task volume.
• D. Incorrect. The Cluster Autoscaler adds underlying worker nodes when the cluster lacks the capacity to schedule pending pods. It does not scale pods based on application-level metrics like a Redis queue; it is a complementary tool that provides infrastructure capacity for a pod autoscaler like KEDA to use.

Correct answers: A, C — Applying FP8 or INT8 weight-only quantization to the model.; Utilizing KV Cache Quantization to reduce memory consumed by concurrent requests.

• A. Applying FP8 or INT8 quantization reduces the model's memory footprint and computational requirements. This allows for higher throughput and the use of less expensive hardware (or fewer GPU nodes), directly reducing deployment costs while maintaining the ability to serve requests reliably.
• B. Temperature is a sampling parameter that influences the randomness and creativity of the model's output. It affects the quality and variability of the text but does not reduce memory usage, compute cycles, or infrastructure costs.
• C. KV Cache Quantization reduces the memory consumed by the attention mechanism's key-value caches during inference. This is especially valuable for handling high concurrency or long-context windows, as it allows more requests to fit on the same hardware, thereby optimizing throughput and cost-per-request.
• D. Dynamic batching is a critical feature for cost optimization. It groups multiple inference requests together to maximize GPU utilization and throughput. Disabling it leads to underutilized hardware and higher costs per inference request.
• E. While Pipeline Parallelism enables the execution of models too large for a single GPU, deploying across multiple physical nodes introduces significant communication overhead and network latency. It is typically a requirement for scaling massive models rather than a primary strategy for cost optimization or enhancing high availability.

Correct answer: C — Multi-agent debate with iterative refinement

• A. Actor-Critic Reinforcement Learning is a training paradigm where an actor generates actions and a critic evaluates them using reward signals to improve a policy. While it shares some terminology, the scenario describes a runtime architectural workflow between two agents using natural language feedback and external data retrieval (RAG), rather than a classic reinforcement learning training loop.
• B. Self-consistency prompting is a technique where multiple reasoning paths are sampled from the same model and the most frequent or consistent answer is selected. It does not involve a separate reviewer agent or an iterative feedback loop for revising output based on external evidence.
• C. Multi-agent debate with iterative refinement (often referred to as Reflection or Reviewer-based workflows) involves separate agents taking on distinct roles—one as a generator and another as a critic. The reviewer checks outputs against external sources (like a vector database), provides feedback, and requires the generator to refine its response. This process improves accuracy and reduces hallucinations through iterative cycles.
• D. Direct Preference Optimization (DPO) is an offline alignment and fine-tuning method used to train language models based on preference data (choosing between a 'better' and 'worse' response). It is not a runtime interactive architecture involving multiple agents and natural language feedback loops.

Correct answer: B — Set the preference as a context variable in Colang (e.g., $account_type) so it persists in the session state.

• A. Storing preferences in a local text file is inefficient, non-standard, and introduces unnecessary I/O latency. It does not integrate natively with the Colang session state or flow routing logic provided by NeMo Guardrails.
• B. Setting context variables in Colang (identified by the '$' prefix) is the native and standard mechanism for session-based memory in NeMo Guardrails. These variables persist throughout the session and allow developers to perform conditional routing and logic within the dialogue flows.
• C. Fine-tuning is a process for adapting a model to a domain or style and is not suitable for real-time session state management. Dynamically fine-tuning an LLM during a conversation is computationally prohibitive and would not provide the instantaneous state updates required for flow control.
• D. While vector databases and time-weighted retrievers are used for long-term memory or external knowledge retrieval, they are excessive for simple session-level variables like user preferences. They introduce unnecessary complexity and latency compared to native session state variables.

Correct answer: C — Define a conditional edge in the state graph that catches the 429 error, increments a retry counter in the state, and routes back to the API node with exponential backoff until successful or a limit is reached.

• A. Appending the error to the prompt is insufficient for control-flow management, and asking an LLM to 'wait' during token generation is not a robust or reliable mechanism for handling network rate limits. This approach conflates the reasoning layer with the orchestration layer and does not provide a reliable retry policy.
• B. Halting the graph for manual intervention defeats the purpose of an autonomous stateful orchestrator. This approach fails to provide automated recovery or utilize the state to resume execution, resulting in wasted progress from steps that succeeded previously.
• C. Implementing a conditional edge with a retry counter in the state allows the graph to loop back to the specific failed node rather than restarting from the beginning. Combining this with exponential backoff is a standard resilient design pattern (common in frameworks like LangGraph) that enables automatic recovery while maintaining the agent's progress and state.
• D. While state persistence is valuable for durability, using SQL triggers to handle API retries is an architectural anti-pattern for orchestration. It adds unnecessary complexity and decouples the retry logic from the agent's control flow, making the system significantly harder to maintain and debug.

Correct answer: B — During the Extract or early Transform phase, before the data is chunked and sent to the embedding model.

• A. Implementing PII redaction after the data is written to the vector database is too late. Once the data is loaded, embeddings have already been generated from the sensitive text and persisted, making it difficult to fully remove the risk and potentially exposing the data through similarity searches or direct retrieval.
• B. The most critical time to redact PII is during the Extract or early Transform phase. This 'Shift Left' security approach ensures that sensitive information is never chunked, sent to an embedding model, or indexed in a vector store. This is the safest method for maintaining compliance and preventing data leakage at the source.
• C. System prompts are not a reliable security boundary for PII. Instructing an LLM to ignore sensitive data does not prevent that data from being stored in the embeddings or loaded into the context window, where it could still be accessed or leaked via prompt injection.
• D. Filtering at the retrieval phase can reduce the immediate exposure of PII in a model response, but it does not address the fact that the sensitive data remains stored and indexed within the vector database, which still constitutes a compliance and security risk.

Correct answer: A — Implement a router agent with function calling that dispatches the revenue query to a Text-to-SQL tool and the risk query to a vector search tool, then synthesizes the results.

• A. Correct. This approach utilizes a router agent to decompose the query into structured and unstructured sub-tasks. Text-to-SQL is the standard for high-precision, deterministic retrieval from relational databases like PostgreSQL, while vector search is the industry standard for semantic retrieval from unstructured data like transcripts. The router then synthesizes these separate tool outputs into a unified answer.
• B. Incorrect. Converting structured relational data into unstructured text for vector search is inefficient, lossy, and removes the ability to perform precise aggregations (like summing revenue). It introduces unnecessary latency and undermines the accuracy required for financial reporting.
• C. Incorrect. Text-to-SQL tools cannot process unstructured text stored in a vector database. Furthermore, prompting an LLM to 'infer' risks from revenue fluctuations is unreliable and would likely lead to hallucination, as the query explicitly asks for risks mentioned in the transcript.
• D. Incorrect. While some LLMs have very large context windows, stuffing raw database rows and full transcripts into the context is not scalable, cost-effective, or performant for real-time enterprise queries. It often leads to the 'lost in the middle' retrieval problem and exceeds token limits as the database grows.

Correct answer: B — It allows the index size to exceed available RAM at the cost of increased retrieval latency due to disk I/O.

• A. Memory-mapped files (mmap) do not inherently increase indexing speed or reduce the maximum number of vectors that can be stored. In fact, mmap is often used to support larger datasets than physical RAM can accommodate.
• B. This is the primary trade-off. mmap allows the operating system to map a file into virtual memory, enabling the vector index to be larger than the available physical RAM. However, when the system accesses parts of the index not currently cached in RAM, it triggers page faults that require disk I/O, leading to higher and more unpredictable retrieval latency compared to fully RAM-resident indexes.
• C. mmap is a memory management technique and does not affect the mathematical precision of the search algorithm or the recall accuracy. Furthermore, it is a standard operating system feature that does not require specialized GPU hardware.
• D. mmap maps the existing file on disk into the address space; it does not perform dynamic compression or reduce the physical disk storage requirements. Compression is typically handled by quantization or encoding techniques separate from the memory-mapping mechanism.

Correct answer: B — OpenAPI Specification (OAS)

• A. GraphQL Schema Definition Language (SDL) is used for defining schemas for GraphQL APIs. While it describes API capabilities, it is not the native standard format used by the NeMo Agent Toolkit for automatically generating REST tool metadata for LLMs.
• B. The OpenAPI Specification (OAS) is the industry standard for describing RESTful APIs. The NeMo Agent Toolkit natively supports OAS (formerly Swagger) to automatically derive tool schemas, including endpoints, input parameters, and descriptions, enabling the LLM to interact with external enterprise systems seamlessly.
• C. gRPC Protocol Buffers (.proto) define gRPC service interfaces and message types. While widely used in microservice architectures, they are not the native format for generating REST-based tool descriptions within the NeMo Agent Toolkit context.
• D. Web Services Description Language (WSDL) is an XML-based language used to describe SOAP-based web services. It is considered a legacy standard compared to REST and is not natively used by the NeMo Agent Toolkit for automatic tool schema generation.

Correct answer: B — It provides high-bandwidth, low-latency GPU-to-GPU interconnects, enabling efficient distribution and synchronization of large image and text embeddings across multiple GPUs.

• A. Incorrect. NVLink and NVSwitch are primarily GPU-to-GPU interconnect technologies designed to facilitate high-speed communication between GPUs. The CPU generally accesses GPU memory through the PCIe bus, and tokenization is typically a host-side or specialized preprocessing step not accelerated by inter-GPU links.
• B. Correct. NVLink and NVSwitch provide high-bandwidth, low-latency interconnects between GPUs, which are critical for the synchronization and exchange of massive image and text embeddings in multimodal models. This is especially important for cross-modal attention layers that must aggregate information from different encoders distributed across a multi-GPU system.
• C. Incorrect. NVENC is a dedicated hardware engine used for video encoding and is independent of the NVLink/NVSwitch interconnect fabric. NVLink handles data transport between GPUs but does not perform data compression.
• D. Incorrect. This describes GPUDirect Storage or GPUDirect RDMA (for NICs). While these are NVIDIA technologies, they focus on optimizing I/O paths between storage/network and GPU memory, whereas NVLink/NVSwitch focus specifically on peer-to-peer GPU communication for compute workloads.

Correct answer: C — Faithfulness (Groundedness) score

• A. Incorrect. Time to First Token (TTFT) is a performance and latency metric that measures the time from the initial request to the generation of the first token. It is used for monitoring system efficiency and user experience, not the factual accuracy or truthfulness of the content.
• B. Incorrect. Context Relevance evaluates how well the retrieved documents align with the user's original query. While it is a key part of the RAG triad for assessing retrieval quality, it does not verify if the final generated response is actually derived from that context.
• C. Correct. Faithfulness (or Groundedness) is the specific metric used to measure whether the generated response is supported by the retrieved documents. This is the primary tool for detecting hallucinations or fabricated information in RAG-based systems.
• D. Incorrect. Tool Selection Accuracy measures the agent's ability to correctly identify and call the appropriate tools or APIs for a given sub-task. It monitors the agent's decision-making logic rather than the factual grounding of its final text output.

Correct answer: B — Groundedness (Faithfulness) score calculated via an LLM-as-a-judge evaluator.

• A. Incorrect. Context Relevance evaluates whether the retrieved context is relevant to the user's query. While critical for the retrieval component, it does not verify whether the model's final response is factually supported by that retrieved context or if it introduced hallucinations.
• B. Correct. Groundedness (also known as Faithfulness) is the specific metric used to ensure that the claims made in a generated response are derived exclusively from the retrieved context. Using an LLM-as-a-judge to evaluate this relationship is a standard industry practice for monitoring RAG reliability and preventing ungrounded advice.
• C. Incorrect. Answer Similarity measures how closely the output matches a predefined 'golden' reference answer. While useful for benchmarking during development, it does not dynamically assess if the response is supported by the specific documents retrieved during a live session.
• D. Incorrect. Mean Reciprocal Rank (MRR) is a retrieval metric that measures the effectiveness of the search ranking. It tells you how well the vector database is finding relevant chunks but provides no information regarding the generation phase or the faithfulness of the synthesized advice.

Correct answer: C — Deploy semantic caching for frequent queries and configure an automatic fallback routing mechanism to a secondary inference endpoint hosting an equivalent model.

• A. While a circuit breaker helps prevent repeated failures, routing all traffic to a significantly smaller 8B model regardless of complexity risks losing the nuanced medical reasoning required for triage. This approach solves availability at the cost of clinical accuracy and safety.
• B. Asynchronous queuing and email notifications introduce unacceptable latency for a healthcare triage workflow, where immediate feedback is often critical. This method manages load but fails to meet the requirement for real-time responsiveness in a live deployment.
• C. This is the most robust solution. Semantic caching reduces the total number of calls to the LLM by reusing responses for similar queries, while an automatic fallback to an equivalent secondary model ensures redundancy and high availability when the primary provider reaches its rate limits.
• D. Increasing the temperature parameter controls the randomness and creativity of the model's output; it does not increase the hardware processing speed, bypass rate limits, or resolve server timeouts.

Correct answer: A — RealToxicityPrompts

• A. RealToxicityPrompts is a dataset specifically created to measure toxic degeneration in language models. It contains 100,000 naturally occurring prompts with varying degrees of toxicity (as scored by the Perspective API) to evaluate how likely a model is to produce toxic continuations.
• B. MMLU (Massive Multitask Language Understanding) is a broad benchmark designed to measure a model's general knowledge and reasoning capabilities across 57 subjects. It is not intended to assess the safety, toxicity, or bias of model outputs.
• C. TruthfulQA is designed to evaluate whether models generate truthful answers and avoid common human misconceptions or false beliefs. While it is a safety-related benchmark regarding misinformation, it is not focused on toxicity.
• D. HumanEval is a benchmark used to evaluate the functional correctness of code generated by language models. It tests programming ability and logic rather than the safety or toxicity of natural language generation.

Correct answers: B, D — Implementing a PII redaction guardrail before user prompts are sent to external LLM APIs.; Designing the vector database in the RAG pipeline with document-level access controls and the ability to selectively delete specific user records.

• A. Storing data immutably on a blockchain directly contradicts the GDPR 'right to be forgotten' (Article 17), which requires that individuals have the ability to request the deletion of their personal data.
• B. Implementing a PII (Personally Identifiable Information) redaction guardrail ensures that only the minimum necessary data is processed and shared with third-party APIs, adhering to the principle of data minimization.
• C. Continuously fine-tuning on raw user interactions increases the risk of the model memorizing personal data and makes it difficult to selectively delete that data, which complicates compliance with erasure requests.
• D. The ability to selectively delete specific user records in the vector database is essential for satisfying 'right to be forgotten' requests. Document-level access controls ensure that data is only accessible to authorized users, supporting integrity and confidentiality.
• E. Disabling system prompts is a prompt engineering choice that impacts the agent's performance and safety steering, but it does not provide a mechanism for data minimization or regulatory compliance.

Correct answer: C — Execute the generated code within an ephemeral, unprivileged Docker container with network access disabled and strict CPU/Memory resource quotas.

• A. Filtering specific import statements using an AST is a blacklist-based approach that is easily bypassed through techniques like dynamic execution (e.g., using __import__ or getattr), string obfuscation, or leveraging other standard library modules. Furthermore, executing code natively on the host machine poses a critical risk if the filter is bypassed.
• B. A Python virtual environment (venv) is designed for managing dependencies and preventing version conflicts, not for security isolation. Even with a non-root user, the process still shares the host's kernel and can potentially access sensitive local resources, files, or network services reachable by that user account.
• C. Executing untrusted code in an ephemeral, unprivileged Docker container provides a robust sandbox that isolates the execution environment from the host machine. Disabling network access prevents data exfiltration and command-and-control (C2) communication, while resource quotas prevent Denial-of-Service (DoS) attacks. This multi-layered approach effectively minimizes the 'blast radius' of any malicious code.
• D. While human-in-the-loop (HITL) review adds a layer of governance, it is not a scalable technical control for an automated agent. Manual review is prone to human error, particularly with highly obfuscated code, and introduces significant latency that breaks the utility of a real-time coding assistant.

Correct answer: B — Implement a pairwise comparison interface where experts simply select the better of two generated reasoning traces to train a reward model.

• A. Having experts write full optimal reasoning traces from scratch is highly burdensome and fails to minimize cognitive load. While this supervised fine-tuning (SFT) approach is effective for data quality, it scales poorly for complex medical reasoning tasks where experts' time is limited.
• B. Implementing a pairwise comparison interface reduces expert effort significantly because reviewers only need to select the better of two traces rather than synthesize a new solution. This is a standard and effective method (often used in RLHF) to collect preference data for training a reward model, which then guides iterative model alignment with much lower cognitive overhead.
• C. Unsupervised clustering and self-assigned rewards based on cluster density do not provide reliable human alignment, particularly in high-stakes domains like medical research. Without expert oversight, the agent may converge on plausible-sounding but factually incorrect reasoning patterns.
• D. Requiring a scalar score for every single step in a trajectory is often more cognitively demanding and time-consuming than evaluating the whole output. It imposes a massive annotation burden and provides less natural guidance than simple preference-based comparisons.

Correct answer: A — Implement a Time-To-Live (TTL) on the paused state checkpoint; a background job should transition the graph to a 'timeout/failed' node if no user input is received within the threshold.

• A. Correct. Implementing a Time-To-Live (TTL) on the paused state checkpoint ensures that the system does not remain indefinitely in a suspended state. A background timeout handler provides a clean recovery path by transitioning the graph to a failure or timeout node, releasing resources and avoiding orphaned processes.
• B. Incorrect. Keeping a connection open indefinitely is poor resource management and leads to server leaks. It does not handle the core issue of the workflow state being stuck if the user never returns, potentially leaving execution threads or workflow states hanging.
• C. Incorrect. Having the LLM invent user intent undermines safety, correctness, and user control. Automatically fabricating responses can lead to incorrect database queries or unintended actions, violating fundamental human-in-the-loop design principles.
• D. Incorrect. Immediate deletion is too destructive and results in a poor user experience. It discards valuable state information needed for auditing or potential session resumption. The system should preserve the checkpoint for a specific duration rather than deleting everything on disconnect.

Correct answer: C — Deterministic execution tracing with immutable audit logs

• A. RLHF is a technique used during the alignment and training phase to tune model behavior based on human preferences; however, it does not provide a runtime mechanism for recording or reconstructing the specific actions taken by an agent during execution.
• B. Semantic caching is used to improve latency and reduce costs by storing and reusing responses for similar queries. It is an efficiency mechanism, not an accountability tool, and does not provide the necessary data to reconstruct a sequence of agent behaviors.
• C. Deterministic execution tracing with immutable audit logs ensures that every step an agent takes—including inputs, intermediate reasoning, tool calls, and outputs—is recorded in a tamper-proof manner. This provides a verifiable trail that allows human auditors to reconstruct the exact sequence of events for post-hoc accountability and investigation.
• D. Automated system prompt optimization is used to refine instructions to improve model performance or reliability. While it influences how an agent behaves, it does not record the agent's runtime actions or facilitate post-event analysis of what occurred.

Correct answer: C — ReAct (Reasoning and Acting)

• A. Zero-shot prompting asks the model to perform a task directly without examples or intermediate reasoning steps. Because it skips the internal decision-making process and action-observation loops, it offers very little inherent decision traceability.
• B. Self-Consistency involves sampling multiple reasoning paths and selecting the most consistent result to improve accuracy. While it uses Chain-of-Thought reasoning, it does not inherently interleave those traces with external actions or environmental observations.
• C. ReAct (Reasoning and Acting) explicitly alternates between reasoning steps ('Thoughts'), external actions ('Actions'), and environmental feedback ('Observations'). This interleaving provides a transparent, step-by-step trace of how the agent arrives at a decision based on its interactions with an environment.
• D. Directional Stimulus Prompting uses a separate stimulus (like a hint or cue) to guide the LLM's response in a specific direction. It is a technique for steering output rather than a framework for creating structured, traceable action-observation logs.