Google: Gemini Embedding 2 Preview

Gemini Embedding 2 Preview is Google’s named preview edition of its native multimodal embedding family. It maps text, images, video, audio and PDF into a shared numerical space, with semantic intent represented across more than 100 languages. The output is embeddings: vectors used to compare and retrieve related material. This preview identity remains separate from the stable Gemini Embedding 2 edition and does not imply interchangeable weights or an identical vector space.

The original announcement describes cross-modal search, classification, clustering, sentiment analysis and retrieval-augmented generation as downstream applications. A text query can find a related visual item, while an image or clip can also act as a search source. An application stores the representations, compares them and returns the corresponding records; a separate language model can use the retrieved source material to draft an answer. The embedding itself does not contain a natural-language explanation or transcript.

Its own model page documents an 8,192-token input limit and output dimensions from 128 to 3,072, with 768, 1,536 and 3,072 recommended sizes. The preview announcement gives a default of 3,072 and describes Matryoshka Representation Learning, which keeps useful information in shorter prefixes of the representation. Keep one chosen dimension for vectors that must be compared and retain the model edition with the collection, rather than merging representations simply because their lengths match.

The preview announcement documents up to six PNG or JPEG images per request, up to 120 seconds of MP4 or MOV video and PDF documents up to six pages. It also documents direct audio embedding without an intermediate text transcription. These specifications come from the preview’s own announcement, not from the later stable guide. The announcement does not establish every current audio-duration, codec or frame-sampling limit for this exact preview profile.

Interleaved multimodal input lets the representation reflect relationships between media, such as a picture and its caption. For a product catalogue, supply the photograph and approved description together when they describe one retrievable object. For a document collection, divide the source into focused passages or pages and preserve its title and position. For a media library, select a clip corresponding to the event users should find and keep its original time range with the stored record.

A useful retrieval task makes the intended relationship clear. A question seeking an explanation, a search for visually similar items and clustering by topic are different tasks even when they use the same source collection. Prepare the content consistently, record the edition and dimension and keep every vector linked to its source. Similarity supplies candidates for the application to inspect; it is not proof that two records state the same fact or that a retrieved source fully answers a question.

For a practical media search, represent the chosen image or video excerpt and compare it with the description of the desired content. Present the matching source alongside its caption or time range so a user can review it. For a knowledge assistant, retrieve the relevant page or passage before generating the written response. The Preview model’s role is the multimodal numerical representation that supports those selection steps, while storage, ranking rules and the final answer remain parts of the surrounding workflow.