Gemini Embedding 2 collapses five modalities into one vector space

One embedding space instead of five pipelines

The core claim in this release is architectural, not incremental. Gemini Embedding 2 maps text, images, video, audio and documents into what Google calls a single, unified embedding space. That is different from stitching together separate encoders per modality and hoping their vectors align.

For teams building retrieval systems, the practical effect is that a query in one modality can retrieve content in another without a bridging step. Google frames this as simplifying complex pipelines, and that is the honest read: fewer separate models to host, version, and keep in sync.

Gemini Embedding 2 maps text, images, videos, audio and documents into a single, unified embedding space, and captures semantic intent across over 100 languages.Montana Labs

The 100-plus language coverage is stated but not quantified beyond that. Treat it as a capability claim to validate against your own corpus rather than a benchmark result.

The input limits define what you can actually build

The specifics matter more than the headline. Text supports up to 8192 tokens. Images allow up to 6 per request in PNG or JPEG. Video accepts up to 120 seconds in MP4 or MOV. Documents can be PDFs up to 6 pages. These are the boundaries you design around.

A 120-second video ceiling and a 6-page PDF limit mean longer media needs to be chunked before embedding. That chunking strategy — where you split a video, how you window a long document — becomes an engineering decision the model pushes back onto you.

The audio detail is the one that changes workflows most. Google says the model natively ingests and embeds audio without needing intermediate text transcriptions. Removing the speech-to-text step removes a source of error and latency that most audio search pipelines currently carry.

Interleaved input and Matryoshka dimensions

Beyond single-modality processing, the model accepts interleaved input — an image plus text in one request — so it can capture relationships between media types rather than embedding each in isolation. That is the difference between indexing a diagram and its caption separately versus embedding them as one unit of meaning.

On storage, Gemini Embedding 2 keeps Matryoshka Representation Learning from the earlier text models. The default output is 3072 dimensions, and developers can scale down to 1536 or 768 while Google recommends those three sizes for highest quality. That gives a direct lever to trade vector storage cost against retrieval quality without retraining anything.

Available where retrieval stacks already live

The model ships in public preview through the Gemini API and Vertex AI, with interactive Colab notebooks for both. More telling for adoption is the integration list: LangChain, LlamaIndex, Haystack, Weaviate, Qdrant, ChromaDB and Vector Search.

Those are the frameworks and vector databases most RAG systems are already built on. Meeting teams inside their existing orchestration and storage tools lowers the switching cost of testing a new embedding model against whatever they run today.

What the single-space design asks you to reconsider

The specific implication of this release is that teams maintaining separate embedding pipelines per modality now have a concrete reason to consolidate. If audio, images, video and documents can share one vector space, cross-modal retrieval stops being a custom project and becomes a query.

The work that remains is unglamorous but real: benchmark the claimed multimodal quality on your own data, design chunking for content that exceeds the 120-second and 6-page limits, and choose an MRL dimension that fits your storage budget. This is a public preview, so those decisions should be validated before anything depends on them.