Google’s EmbeddingGemma 2 brings AI search across text, images and audio onto devices

Google released EmbeddingGemma 2 on 6 October, extending its lightweight embedding model to images, audio and video as well as text. The aim is to make different kinds of information searchable together on consumer hardware.[1]

Imagine remembering a discussion about a delayed shipment but forgetting whether the useful detail was in a voice note, a screenshot or a meeting document. The natural question is about the shipment. A filing system asks you first to remember where the evidence lives. A shared search representation can reduce that mismatch between human memory and the way files are stored.

The search layer behind an assistant

An embedding represents an item as a list of numbers. A search system can compare that list with a representation of a query and rank likely matches. Google’s model maps its supported media into the same vector space, allowing a text query and an image, for example, to be compared through that shared representation.[3]

This addresses a different part of the workflow from composing an answer. Retrieval finds potentially relevant material. A reader or a separate generative system must still determine what that material establishes. Finding an invoice similar to the requested one does not establish that it is the correct invoice; matching a discussion about a decision does not establish that the decision was approved.

The modular design is useful because a document search tool need not carry every media component. For a builder, the sensible starting point is the material the product actually needs to retrieve. Supporting more formats has value when it solves a real search problem; adding them simply because the model can process them creates more indexing and evaluation work.

The storage saving has a measurable cost

The model supports vectors with 768, 512, 256 or 128 dimensions. Google’s model card reports lower scores on the Massive Multimodal Embedding Benchmark (MMEB) as these representations become smaller. All benchmark results in that card use the full-precision checkpoint.[2]

In this example, moving from 768 to 256 dimensions cuts raw vector storage by two-thirds. Moving to 128 dimensions reduces it by five-sixths. Those savings are reproducible arithmetic; whether the smaller representation retrieves the right files often enough is a separate product decision.

An archive of 100,000 documents will not necessarily produce 100,000 vectors. A long document or recording may be divided into many searchable sections. The relevant sizing question is how many indexed items the application creates, and what else it stores with each item. Counting original files alone can understate the eventual footprint.

Google also specifies that shortened vectors must be normalized again and that queries and stored items must use matching dimensions. These are implementation requirements that affect ranking. A system returning plausible results is not, by itself, evidence that its numerical configuration is correct.[2]

What can be tried now

Google’s AI Edge Gallery project now lists Instant Media Search and Video Moment Finder for exploring local retrieval. Its published minimums are Android 12 and iOS 17, and it describes the app as an experimental beta. These are documented entry points, not a guarantee of identical performance on every supported phone.[5]

Google’s developer announcement also introduces Foresight, an experimental Mac application combining local retrieval with a meeting companion. The same announcement places the Android ML Kit service integration in the coming weeks. The available demonstrations and the planned platform integration are different stages of delivery.[3]

For an initial evaluation, create a small collection with known matches and deliberate near misses. Include two similar receipts with different dates, a recording containing a relevant phrase in the wrong context, and a screenshot whose filename gives no clue to its content. Check whether the exact evidence appears among the first results. This tests a useful property that a polished demonstration may not expose.

Local retrieval changes the data path

Google describes its Gallery demonstrations as generating and searching embeddings on the device without a cloud round trip. That can remove an external transfer from the retrieval stage. Its developer tools are intended to support this local architecture across media types.[3]

The privacy implication depends on the whole application. If a later step sends the retrieved document to a cloud assistant, the document still leaves the device. If an index is copied into a shared backup, local computation has not made that copy private. Developers should explain where source files, derived representations and final answers are processed and stored.

The significance of this release lies in making a useful component easier to build into everyday software: search that follows meaning across different media. Success should be visible in ordinary tasks—finding the right passage, recording or image, then opening the evidence in context. That is a concrete benefit an application can demonstrate without claiming to understand everything in a person’s digital life.

Report an error or suggest a correction