Applications of foundation models

A retail company wants to use a foundation model to generate product descriptions. Legal requires that the model never produce outputs containing competitor brand names. Which mechanism is the most direct way to enforce this constraint at the application layer?

• AReduce the model's temperature to zero so it always picks the highest-probability token, making competitor names statistically unlikely.
• BFine-tune the foundation model on internal product data only, so the model's weights no longer associate competitor brand names with product description tasks.
• CUse a larger foundation model with higher reasoning capability, because larger models are better at following nuanced instructions to avoid competitor references.
• DApply output guardrails that scan generated text and block or redact any response containing prohibited brand names before it reaches the user.check_circle Correct

Why A is wrong: Greedy decoding (temperature 0) selects the most probable token at each step but cannot guarantee exclusion of specific words. Competitor names may still appear if they are highly probable in context.

Why B is wrong: Fine-tuning can shift output distributions but does not reliably eliminate specific outputs. The model retains knowledge of competitor names from pre-training and may still generate them in novel contexts, so this approach does not provide the deterministic enforcement legal requires.

Why C is wrong: Model size improves general instruction-following but is probabilistic. A larger model may comply more often, but compliance is not guaranteed on every call. Guardrails provide the deterministic enforcement the question calls for.

Why D is correct: Guardrails act as a post-generation filter at the application layer, providing a deterministic enforcement mechanism independent of model probabilities. This directly satisfies the legal requirement by intercepting non-compliant outputs before delivery.

A company is building a retrieval-augmented generation (RAG) application that retrieves passages from a knowledge base before generating answers. As the knowledge base grows to millions of documents, the team observes that the retrieved passages are occasionally irrelevant to the user's question. Which design consideration should they address to improve retrieval relevance?

• AIncrease the foundation model's context window so it can ingest more retrieved passages per call, improving the chance that a relevant passage is included.
• BApply output guardrails to filter model responses that appear to contradict the retrieved passages, so users only see answers grounded in retrieved content.
• CImprove the embedding model and similarity search configuration used in the retrieval step, so semantically relevant passages are ranked higher for each query.check_circle Correct
• DReplace the foundation model with a domain-specific model pre-trained on the same corpus as the knowledge base, so the model can answer questions from memory rather than relying on retrieval.

Why A is wrong: A larger context window allows more passages to be sent to the model per call, but it does not improve the quality of retrieval ranking. If the retrieval step consistently returns irrelevant passages, adding more of them to a larger context window does not solve the underlying problem.

Why B is wrong: Guardrails check outputs after generation. They can catch responses that contradict retrieved passages but cannot compensate for retrieving irrelevant passages in the first place. The root cause is the retrieval step, not the generation step.

Why C is correct: In a RAG architecture, retrieval quality is determined by how well the embedding model captures semantic meaning and how the vector similarity search is configured. Improving the embedding model or tuning retrieval parameters (such as similarity threshold or top-k) directly addresses the root cause of irrelevant passage retrieval.

Why D is wrong: Pre-training on the corpus would embed static knowledge into model weights, not provide access to a live, growing knowledge base. This approach removes the scalability and update benefits of RAG and does not address retrieval relevance as a design consideration.

A developer wants to give a foundation model access to the company's internal policy documents, which are updated monthly. The model should cite the exact policy clause in each answer. Which architecture satisfies both requirements?

• AProvide the entire policy library as a static system prompt at model initialisation.
• BFine-tune the model on a labelled dataset of policy clauses and expected answers, then re-fine-tune each month.
• CUse prompt chaining where the first model call summarises all policies and the second call answers from that summary.
• DBuild a RAG pipeline that indexes policy documents in a vector store, retrieves matching clauses at query time, and injects them with source metadata into the model's prompt.check_circle Correct

Why A is wrong: Tempting for small document sets, but a full policy library will likely exceed context limits and cannot be refreshed without redeploying the service. Exact clause citation also requires structured retrieval, not relying on the model's recall from a long prompt.

Why B is wrong: Tempting because fine-tuning teaches domain knowledge, but monthly retraining cycles are expensive and the model cannot reliably cite specific clause identifiers - it paraphrases rather than quotes stored source text.

Why C is wrong: Tempting because multi-step prompting can improve answer quality, but summarisation loses the precise clause text needed for citation, and the summary itself would be stale unless regenerated every month.

Why D is correct: RAG retrieves the actual clause text along with its identifier at inference time and places it in the prompt, enabling the model to quote it directly. Monthly updates require only re-indexing, not retraining.

A customer-service assistant holds long multi-turn conversations. The team notices that in lengthy sessions the assistant starts to ignore a constraint the customer stated near the very beginning, such as a preferred delivery address, while still handling recent turns well. Each request resends the whole conversation transcript to the model. Which design consideration best explains this behaviour and points to the fix?

• AThe model's temperature is set too high, so it samples less likely tokens and drifts away from the early constraint as the conversation grows
• BThe assistant has begun hallucinating the delivery address, so grounding each response in a retrieved knowledge base will restore the early constraint
• CThe growing transcript exceeds the model's context window, so the earliest turns fall outside the tokens the model can attend to, and the fix is to summarise or retain key earlier facts within the windowcheck_circle Correct
• DThe output token limit is being reached, so the model truncates the part of its answer that would have applied the early constraint

Why A is wrong: High temperature increases randomness in wording, which is a believable cause of drift, but it would degrade recent turns too rather than specifically dropping the earliest stated constraint as the transcript lengthens.

Why B is wrong: Hallucination involves fabricating unsupported facts, which is a plausible guess, but here the address was genuinely supplied earlier and simply stops being honoured, which is a context-capacity issue rather than invented content.

Why C is correct: Once the resent transcript grows past the context window, the oldest tokens are no longer available to the model, which is why early constraints are lost while recent turns work; carrying a running summary or pinned facts keeps the important earlier information inside the window.

Why D is wrong: An output limit can cut off the end of a response, which sounds related, but it would shorten answers rather than selectively forget an instruction given many turns earlier in the input.

During a hallucination audit, evaluators need to measure how often a foundation model generates factually incorrect claims that are not supported by the provided source documents. Which evaluation approach is most appropriate for this measurement?

• AMeasure perplexity on a held-out factual corpus to identify confidence drops that correlate with incorrect generation.
• BRun the model on a multiple-choice benchmark that tests factual knowledge recall without reference documents.
• CCalculate BLEU score between the generated output and a gold-standard reference answer to measure lexical overlap.
• DApply human raters or an NLI-based automatic checker to assess whether each model claim is entailed by the source document.check_circle Correct

Why A is wrong: Perplexity reflects language model fit to a distribution but does not directly identify whether individual claims are supported by source documents. A model can be confident (low perplexity) while still hallucinating facts not in the source.

Why B is wrong: A knowledge-recall benchmark tests whether the model knows facts from pretraining, not whether it generates claims that are grounded in provided source documents. This does not address the hallucination scenario described.

Why C is wrong: BLEU measures n-gram overlap between a candidate and a reference translation or answer. It cannot distinguish a factually correct but paraphrased response from a hallucinated response that shares surface words with the reference.

Why D is correct: Grounded faithfulness evaluation uses Natural Language Inference (entailment) or human verification to determine whether each generated claim is supported by the source. This directly measures unsupported hallucination in a retrieval-augmented or document-grounded context.