AI-103 Practice Question 38: Visual Question Answering
Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · Test 4 · question 8 of 10
Claims adjusters at an insurer want to ask open-ended questions about damage photographs: "is the rear bumper cracked or only scuffed?", "is there a second vehicle in the frame?", "does the odometer reading in this photo match the claim?", and get answers grounded in what the picture actually shows. The questions are not known in advance and change with each claim, and there is no labelled training data.
What should the application do?
Show answer and explanation
Answer:
Vision-enabled chat models are large multimodal models that "can analyze images and provide textual responses to questions about them" and "can answer general questions about what's present in the images you upload." You call the Chat Completion API on a deployed vision-capable model with the image alongside the text. One knob worth setting deliberately: the detail parameter on the image_url field: low processes a 512×512 version for speed and fewer tokens, high segments the image for a more detailed read at roughly double the token budget per segment, which is what a question about an odometer reading needs.
- B is wrong because tags and objects come from a fixed taxonomy. They tell you a car and a bumper are present; they cannot distinguish "cracked" from "scuffed" or answer a question nobody anticipated. - C is wrong because it requires labelled examples the insurer doesn't have, and a new model per question type is unworkable when the questions change with every claim. - D is wrong because OCR extracts text. It would read the odometer digits, but it is blind to the damage itself and to everything else in the frame.
Reference: How to use vision-enabled chat models