Articles · Gemini AI Tools
Compare supported items only, and fail closed when the embedding model cannot
A multimodal compare box is easy to treat as a universal “same thing?” meter for photos, clips, and PDFs. It is not. Multimodal Similarity only ranks combinations the configured embedding model actually supports. On DevNestro the default model is gemini-embedding-2, which this host treats as text embeddings. Text ↔ text and document-as-text ↔ query can run. Image, audio, and video combinations fail closed instead of inventing a score. When a comparison is allowed, the items are sent to Google for embeddings. The cosine score is still not proof of the same meaning.
Why this is worth doing in the browser
People want image ↔ text search because a caption and a photo feel comparable. Shipping a fake score when the model cannot embed the image would be worse than a refusal. This page follows the fail-closed rule used elsewhere on Gemini tools: unsupported modality combinations are not dressed up as success. Related tools include text similarity for two passages and the image analyzer when you need a caption rather than a vector. Audio and video stay on their own analyzer pages.
How to use the tool
Choose text or a text document for each side. Confirm cloud embeddings. Compare. If you pick an image on gemini-embedding-2, the request fails closed with a clear message. Clear drops the session files and text.
- Do not upload confidential files unless you accept Google processing.
- Do not expect image, audio, or video scores on the default embedding model.
- A refused request is the honest result when the model cannot embed that type.
- Use text similarity when both sides are passages.
Privacy
Supported items are sent to Google Gemini for embeddings. Unsupported types are rejected before a pretend score. Vectors are not written to disk. This is not processed only in your browser. Free-tier training language follows the configured tier.
Have two supported items you are willing to send? Open Multimodal Similarity and treat an unsupported combination as a closed door, not a bug.