Question

You are building a RAG solution in Azure AI Foundry for a retail company. Users will upload a photo of a product and ask questions, and the system must retrieve matching product records that contain both text descriptions and reference images stored in a vector store. You need to ensure a photo query can be matched against both the stored product images and their text descriptions within a single vector similarity search. Which approach should you use to generate the vectors for indexing and querying?

Accepted Answer

Use a multimodal embedding model that maps images and text into a shared vector space for both indexing and query encoding.

Answer

Generate text embeddings from OCR-extracted captions of the images and embed the user's photo with a separate image classification model.

Answer

Store the images as base64 strings in the index and rely on the LLM to compare them at query time without embeddings.

Answer

Use two independent vector indexes—one text embedding index and one image embedding index—and query only the text index.