A legal-summary assistant built in Foundry occasionally states facts that are not actually present in the internal contract documents it is meant to summarize. The team wants to reduce these unsupported statements. (Select TWO.) Which two changes would directly address this?
Select all correct answers, then click Submit.
Short Explanation
The problem here is that the assistant is confidently saying things that just are not in the actual contracts, so the fix has to attack that from two directions at once: first, make sure it is actually pulling real passages from the real documents before it answers, so its statements are tied to something true, and second, explicitly tell it in its instructions to stick to what that retrieved material actually says and to admit when the answer just is not there rather than filling the gap with something invented. Turning up the randomness setting pushes in exactly the wrong direction, since it makes the wording more varied and unpredictable, which tends to invite more invented detail, not less. And simply moving to a bigger, more powerful model without changing anything else does not fix this either, because a bigger model with no grounding and no instruction to admit uncertainty will still confidently make things up when it does not actually know.
Full Explanation
The correct answers are A and B. Grounding the assistant's responses in a retrieval index built from the actual contract documents forces it to generate answers from real, retrieved text rather than from patterns memorized during training, which directly reduces statements that are not backed by the source material. Adding an explicit system-message instruction telling the model to answer only from the provided context, and to admit when the context does not contain the answer, reinforces that same boundary at the instruction level, giving the model an explicit permission to say it does not know instead of filling gaps with invented detail. Option C is incorrect because raising the temperature setting increases the variety and creativity of wording, which tends to make ungrounded, invented statements more likely rather than less, working directly against the goal. Option D is incorrect because switching to a model with a larger parameter count, with no other change, does not address the underlying cause, since even large models will generate unsupported statements when they are not grounded in the actual source documents and are not instructed to acknowledge uncertainty.