Yes, embed-english-v3.0 can handle cross-modal image-text search when you build your pipeline so that text inputs and image inputs are embedded into a comparable vector space and then searched consistently. In developer terms, cross-modal search means workflows like “text-to-image retrieval” (a user types a phrase and you return relevant images) and “image-to-text retrieval” (a user provides an image and you return related text passages or captions). The model’s role is to produce vectors that make those comparisons meaningful; your system’s role is to store, index, and filter those vectors correctly.
A practical implementation starts with clear ingestion rules for each modality. For images, you embed the image input (or an image representation that your embedding endpoint accepts) and store the resulting 1024-dimensional vectors with metadata such as asset_id, url, tags, created_at, and modality="image". For text, you embed captions, alt text, product descriptions, or documentation chunks and store vectors with metadata like doc_id, section, source_url, and modality="text". These vectors typically live in a vector database such as Milvus or Zilliz Cloud. You can store everything in one collection with a modality field for filtering, or split by modality into separate collections if that simplifies indexing and access patterns.
Query-time behavior depends on your UX. For text-to-image, embed the user’s text query and search only modality="image" vectors (and optionally also caption vectors if you store both). For image-to-text, embed the user’s image query and search modality="text" vectors if you want to retrieve text passages related to the visual content, or search modality="image" vectors if you want similar images first. In both cases, good metadata and post-processing matter: you may need to merge results (image + caption), apply business filters (only show images from a specific catalog), and format outputs for the UI. Cross-modal systems are especially sensitive to inconsistent preprocessing, so keep ingestion and query paths symmetrical and version your pipeline.
For more resources, click here: https://zilliz.com/ai-models/embed-english-v3.0