AI-103 Practice Question 58: Multimodal Image Embeddings
Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · Test 6 · question 8 of 10
A stock-photo marketplace has hundreds of thousands of untagged photos. It wants shoppers to type a free-text description, such as "a red kayak on a calm lake at sunset," and get back matching photos. Manually tagging the catalog isn't feasible, and the set of possible descriptions is open-ended, so a classifier trained on a fixed list of categories won't cover it.
Which Azure AI Vision capability should the search be built on?
Show answer and explanation
Answer:
Multimodal embedding "generates a vector representation of an image that captures its features and characteristics" in a way "compatible with text search over the same vector space." The VectorizeImage and VectorizeText APIs embed photos and free-text queries into that shared space, and results are ranked by vector similarity; Microsoft names e-commerce search by "visual features or descriptions" as a direct application. This is exactly the open-ended, untagged, natural-language scenario described.
- A is wrong because it's traditional keyword search under the hood: it "looks for the exact match of the keywords ... entered by the user" against generated tags, and "relies heavily on the user's ability to use relevant and specific search terms"; a shopper's free phrasing won't reliably match auto-generated tags. - B is wrong because a classifier only recognizes the fixed label set it was trained on. An open-ended natural-language description like the example query isn't a category, and retraining for every new phrase shoppers might type isn't feasible. - D is wrong because OCR extracts text that appears inside an image (signs, labels, printed words). Most stock photos have no visible text at all, so this captures none of the visual content the search needs to match on.
Reference: Multimodal embeddings concepts - Image Analysis 4.0