NCP-GENL Prompt Engineering Practice Test 3 — 15 Questions

Question 1

A developer is building a customer support assistant using an NVIDIA NIM microservice for a Llama 3 model. The assistant must always respond in valid JSON with keys 'category' and 'urgency'. The model often returns conversational text instead. Which prompt engineering change most directly enforces the required output format?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Append 'Return only JSON with keys category and urgency' and set the NIM request parameter 'guided_json' to the target schema.. Structured output requires both a clear instruction and a decoding-time constraint. NVIDIA NIM supports guided_json, which restricts token generation to a supplied JSON schema, guaranteeing valid keys and syntax. Pairing that with an explicit prompt instruction aligns the model's behavior with the schema. Few-shot examples or temperature changes alone cannot guarantee strict JSON compliance.

Answer

Add 'Do not hallucinate' to the prompt and lower the max_tokens parameter to 50.

Answer

Add a system prompt that says 'You are a helpful assistant' and increase the temperature to 0.9.

Answer

Use few-shot examples of JSON outputs and set top_p to 0.1 without any schema constraint.

Question 2

A data scientist is using an NVIDIA NeMo LLM to generate Python code from natural language descriptions. The model often produces code that works but does not follow the team's style guide, such as using single quotes instead of double quotes and missing type hints. Which prompt engineering technique should the data scientist use to improve adherence to the style guide?

Accepted Answer

Few-shot prompting with examples that demonstrate the desired coding style.. Few-shot prompting with examples that embody the desired style guide is the most effective way to teach the model the specific formatting rules. By showing the model correct examples, it can mimic the style in new generations. This technique is particularly useful for coding tasks where precise syntax and style matter.

Answer

Chain-of-thought prompting to encourage the model to reason about the style guide before writing code.

Answer

Zero-shot prompting with a detailed instruction describing the style guide rules.

Answer

Increasing the temperature to allow more creative code generation.

Question 3

A team is using an NVIDIA NIM for a Mistral model to classify support tickets into one of five fixed categories. Accuracy is inconsistent, and the model sometimes invents new categories. Which prompt engineering change is most likely to improve reliability without retraining the model?

Accepted Answer

Provide three labeled examples per category in the prompt and instruct the model to output only the category name.. Few-shot examples with explicit labels define the allowed output space, and the instruction to emit only the category name prevents invented labels. This combination directly targets both symptoms, inconsistency and hallucinated categories, and requires no model retraining, unlike sampling or reasoning changes that leave the label set undefined.

Answer

Add a chain-of-thought instruction that asks the model to reason step by step before naming a category.

Answer

Increase the context window by concatenating the entire ticket history for every request.

Answer

Lower the temperature to 0.0 and remove all instructions so the model relies on its pretrained knowledge.

Question 4

An engineer is building a customer-facing FAQ bot using an NVIDIA NIM-hosted Llama 3.1 70B model. The bot must answer ONLY from a supplied product knowledge base and must respond with 'I don't have that information' when the answer is not present. Which prompt engineering approach BEST enforces this constraint?

Accepted Answer

Instruct the model to answer only from the provided knowledge base, explicitly state that it must reply with the refusal phrase when the answer is absent, and include the knowledge base inside clearly delimited sections.. Grounded FAQ bots need an explicit behavioral contract: what source to use, what to do when the source lacks the answer, and clear separation of context from user input. Delimiters prevent prompt injection from blending with instructions, and a fixed refusal phrase makes the fallback deterministic and testable. Sampling parameters alone cannot enforce scope constraints.

Answer

Set the top_p value to 0.1 and rely on the model's pretrained knowledge of the product domain instead of supplying the knowledge base in the prompt.

Answer

Append the entire knowledge base to every user message without any instruction about scope or refusal behavior, letting the model infer the rules from context.

Answer

Increase the temperature to 1.0 and add 'Be creative and helpful' to the system prompt so the model can improvise when the knowledge base is incomplete.

Question 5

A team is using an NVIDIA NeMo-based LLM to answer questions over a product manual. The model sometimes answers using general knowledge instead of the provided manual excerpts. They want to force the model to rely only on the supplied context. Which prompt engineering approach best addresses this?

Accepted Answer

Instruct the model to answer only from the provided context and to reply 'Not in the provided context' when the answer is absent.. Grounding a model in retrieved context requires an explicit instruction that restricts answers to that context and defines what to do when the answer is missing. A fallback phrase such as 'Not in the provided context' prevents the model from filling gaps with pretrained knowledge. Few-shot examples help style but do not enforce the boundary; temperature and context removal work against the goal.

Answer

Increase the model's temperature so it explores more diverse answers from its pretrained knowledge.

Answer

Add more few-shot examples of correct answers without changing the instruction about using the context.

Answer

Shorten the prompt by removing the manual excerpts and rely on the model's product knowledge.

Question 6

An engineer is deploying an NVIDIA NeMo Guardrails system to moderate a chatbot's responses. The chatbot must refuse to answer questions about politics but should answer questions about weather. Which prompt engineering strategy in NeMo Guardrails is most appropriate to enforce this behavior?

Accepted Answer

Define a dialogue flow with a canonical form for political questions that triggers a refusal response, and a separate flow for weather questions that allows the answer.. NeMo Guardrails uses Colang dialogue flows to define bot behavior based on user intent. Creating separate flows for political and weather questions ensures deterministic refusal or answering. This is the core prompt engineering approach within the guardrails framework and provides reliable, auditable control.

Answer

Add a Python function that checks the user input for political keywords and returns a refusal, bypassing the model entirely.

Answer

Use a system prompt that instructs the model to refuse political questions and answer weather questions.

Answer

Fine-tune the underlying LLM to refuse political questions and answer weather questions.

Question 7

A developer is prompting an NVIDIA NIM for a Code Llama model to generate a Python function. The model produces correct logic but frequently omits type hints and docstrings, which the team requires. Which prompting technique best addresses this specific gap?

Accepted Answer

Include a short example of the desired function signature, type hints, and docstring in the prompt, then ask for the new function in the same style.. A concrete example showing the desired function signature, type hints, and docstring teaches the model the exact format by imitation. Vague quality instructions, larger token budgets, and reasoning steps do not specify these structural requirements, so they fail to close the gap. The example-based approach is the most direct and reliable fix.

Answer

Increase max_tokens so the model has more room to include type hints and docstrings.

Answer

Add the instruction "write clean, production-quality code" to the system message.

Answer

Ask the model to first explain its reasoning about the function, then output the code.

Question 8

A developer is building an interactive assistant using NVIDIA NIM microservices. The assistant must answer questions about a specific set of internal policies. The developer wants to ensure the model's responses are grounded in those policies and not in its general pre-training knowledge. Which prompt engineering technique should be applied?

Accepted Answer

Provide the relevant policy excerpts directly within the prompt, instruct the model to answer only from that context, and to state when the answer is not present.. Grounding the model in a specific set of documents requires placing the relevant excerpts in the prompt and explicitly instructing the model to answer only from that context. This constrains the model to the provided facts and reduces hallucination. Other techniques like increasing temperature, chain-of-thought, or fine-tuning do not achieve this grounding for a dynamic, interactive policy assistant.

Answer

Fine-tune the model on the internal policy documents, then use a simple zero-shot prompt to ask questions.

Answer

Use a chain-of-thought prompt that asks the model to reason step by step about the policy before answering.

Answer

Increase the temperature setting to 0.9 to encourage the model to explore a wider range of policy interpretations.

Question 9

A developer is using an NVIDIA NIM for a Llama 3.1 70B model to build a legal document review assistant. The model must answer questions based on a provided contract, but the contracts are often 50,000 tokens long, exceeding the model's 8,000-token context window. Which prompt engineering strategy is most appropriate to handle this constraint?

Accepted Answer

Use a retrieval-augmented generation (RAG) approach to fetch only the relevant sections of the contract and include them in the prompt.. RAG is the most appropriate strategy because it retrieves only the relevant sections of the long contract, fitting them into the model's context window. This allows the model to answer questions accurately without needing to process the entire document. It is a widely adopted prompt engineering pattern for handling context length limitations.

Answer

Split the contract into chunks and ask the model to summarize each chunk sequentially, then combine the summaries.

Answer

Increase the model's context window by fine-tuning it on longer sequences.

Answer

Use a sliding window approach where the model processes the contract in overlapping segments and maintains a memory of previous segments.

Question 10

A team fine-tunes an NVIDIA NeMo model to classify support tickets into five categories. In production, the model sometimes outputs free-form explanations instead of a single category label, breaking the downstream parser. Which prompt engineering change MOST reliably constrains the output format?

Accepted Answer

Add three few-shot examples that each end with the exact category label, and instruct the model to output only the label with no additional text.. Strict output contracts are best enforced by showing the exact desired format through few-shot examples and explicitly prohibiting any additional text. Demonstrations act as in-context conditioning that shapes the model's continuation pattern, while the instruction closes the loophole of adding commentary. Sampling parameters control randomness, not structure.

Answer

Add a polite request asking the model to 'try to keep answers short and to the point' while leaving the output format unspecified.

Answer

Increase max_tokens so the model has more room to explain its reasoning before stating the final category label.

Answer

Lower the temperature to zero and trust that deterministic sampling will force the model to emit only a category label.

Question 11

An engineer is optimizing prompts for an NVIDIA NIM-hosted model used in a multi-turn technical troubleshooting chat. The model forgets earlier constraints, such as the customer's environment and the product version, as the conversation grows. Which prompt engineering technique best preserves these constraints across turns?

Accepted Answer

Maintain a compact running summary of key constraints and inject it into the system prompt each turn.. Multi-turn memory is best preserved by extracting and re-injecting critical constraints rather than replaying all dialogue. A compact running summary placed in the system prompt keeps key facts salient and stable while avoiding context bloat. Temperature and single-turn designs do not address memory, and verbatim history can dilute attention and waste tokens.

Answer

Repeat the full conversation history verbatim in every turn and increase max_tokens.

Answer

Use a single-turn prompt for each message and rely on the model's pretrained knowledge of the product.

Answer

Lower the temperature to 0 and remove the system prompt to avoid conflicting instructions.

Question 12

A team is deploying an NVIDIA NIM for a Llama 3 model as a retrieval-augmented generation (RAG) assistant over internal documentation. Users report that the assistant sometimes answers from its pretrained knowledge instead of the retrieved passages, and occasionally cites a passage that does not support its claim. Which TWO prompt engineering changes best reduce these behaviors? (Choose two.)

Accepted Answer

Instruct the model to answer only from the provided context and to respond with a fixed phrase when the context is insufficient.. Grounding the model with an explicit instruction to use only the provided context, plus a fallback for insufficient information, prevents reliance on pretrained knowledge. Requiring verbatim supporting quotes makes each claim auditable and discourages fabricated citations. Together these changes enforce source adherence and citation accuracy, while sampling changes or removing context would undermine the RAG design.

Answer

Add a chain-of-thought instruction asking the model to reason about why the user's question is interesting before answering.

Answer

Raise the temperature to 0.8 so the model produces more varied answers and avoids memorized responses.

Answer

Remove the retrieved passages from the prompt and rely on the model's pretrained knowledge to keep the prompt short.

Question 13

A team is using an NVIDIA NIM-hosted Llama model to generate product descriptions from a list of technical specifications. The descriptions sometimes omit key specifications or include invented features. The team wants to improve reliability without changing the model. Which prompt engineering change is most likely to reduce these errors?

Accepted Answer

Add a system prompt that instructs the model to act as a technical writer and to include every specification exactly as given, and provide a few-shot example of a correct description.. To reduce omissions and inventions without changing the model, the most effective prompt engineering approach is to provide explicit instructions and few-shot examples. A system prompt that defines the role and constraints, combined with an example that demonstrates including all specifications and avoiding fabrication, guides the model to produce more reliable outputs. Other parameter changes or chain-of-thought do not directly address content fidelity.

Answer

Increase the max_tokens parameter to allow the model to generate longer descriptions, ensuring all specifications are covered.

Answer

Use a chain-of-thought prompt that asks the model to list each specification and then write the description.

Answer

Set the temperature to 0.0 to make the model deterministic and reduce creativity.

Question 14

A developer is writing a system prompt for an NVIDIA NIM-hosted assistant that must always respond in formal English, never use slang, and never reveal internal system instructions. Where should these persistent behavioral rules be placed for the MOST consistent effect?

Accepted Answer

In the system prompt, stated as explicit rules that apply to every turn of the conversation.. Persistent behavioral requirements such as tone, style, and confidentiality belong in the system prompt, which the model treats as standing guidance across the conversation. This placement keeps rules stable regardless of what users type and avoids the fragility of repeating them per message or expecting decoding parameters or base alignment to handle policy.

Answer

Omitted entirely, because well-aligned base models already default to formal English and never disclose their instructions.

Answer

Encoded only in the model's temperature and top_p settings, since decoding parameters control response style.

Answer

Appended to the end of each user message as a reminder, so the rules stay close to the model's most recent input.

Question 15

A developer is using an NVIDIA NIM-hosted model to classify support tickets into a fixed set of categories. The model occasionally invents new category names. The team wants to guarantee that only allowed categories are returned. Which approach is most appropriate?

Accepted Answer

Provide the full category list in the prompt and use the NIM guided_choice parameter with those categories.. When output must be one of a fixed set, enumerating the categories in the prompt and applying a decoding constraint such as guided_choice ensures only valid labels are emitted. Few-shot examples and reasoning improve quality but do not enforce the boundary. High temperature and broad top_k work against the requirement by increasing variability.

Answer

Use few-shot examples of each category and set top_k to 100.

Answer

Ask the model to explain its reasoning before choosing a category.

Answer

Add 'Choose from the list' to the prompt and set temperature to 1.0.