Question

You are building a visual question-answering feature for an insurance claims app. Adjusters upload photos of damaged vehicles and ask natural-language questions such as 'Is the rear bumper cracked?'. Compliance requires that the model must never guess: if the uploaded image does not contain evidence to answer, the response must clearly state that the answer cannot be determined from the image rather than fabricating a plausible-sounding answer. Which approach best satisfies this requirement?

Accepted Answer

Instruct the multimodal model via system prompt to answer strictly from visible content in the image and to respond that it cannot determine the answer when the required region is not visible, then evaluate with a groundedness metric.

Answer

Increase the temperature parameter so the multimodal model produces more varied answers and is less likely to repeat an incorrect claim.

Answer

Post-process every model answer with an OCR step to extract any text in the image and return that text as the answer.

Answer

Route all questions to a text-only large language model, passing only the adjuster's question and no image, to reduce hallucination risk.