Google drops Gemini Embedding 2 for unified multimodal search

Building a multimodal search pipeline used to feel like duct-taping five different systems together. You ran images through CLIP, processed audio through a transcription model, chopped up videos, and prayed the resulting vector spaces played nicely together.

Google is taking a direct swing at that mess with Gemini Embedding 2. Built on the Gemini architecture, it is Google's first natively multimodal embedding model, now live in public preview on the Gemini API and Vertex AI.

Basically, it maps text, images, video, audio, and documents into a single, unified vector space across more than 100 languages.

Native audio and mixed inputs in one shot

The big shift here is that everything is truly native rather than chained together on the backend.

On the text side, the model takes up to 8,192 tokens. But the media specs are where things get interesting for developers.

A single request can handle up to 6 images in PNG or JPEG format, or up to 120 seconds of video in MP4 and MOV. It can also ingest PDFs directly up to 6 pages long.

Audio is where Google made the smartest cut. Instead of transcribing speech to text and losing nuance along the way, Gemini Embedding 2 natively absorbs raw audio without any intermediate transcription step.

It also supports interleaved inputs. That means you can send an image and a text prompt in the same payload, allowing the model to capture the exact relationship between the media types instead of treating them as isolated blobs.

To keep storage bills under control, Google baked in Matryoshka Representation Learning. While the default embedding size is 3,072 dimensions, you can cleanly scale down to 1,536 or 768 dimensions without starting from scratch. Integrations are already live across LangChain, LlamaIndex, Weaviate, Qdrant, and ChromaDB.

Why it matters

We think this release solves one of the biggest architectural headaches in modern context engineering.

For builders working on multimodal RAG, customer service audio logs, or video search, the value is immediate. You can throw away the brittle preprocessing pipelines that converted voice to text or extracted video frames separately. One single endpoint turns raw, mixed-format data into searchable math.

Across the industry, this sets a higher baseline for what an embedding model needs to be. For years, vector search has mostly been a text-first game with awkward image plug-ins. By extending Gemini's multimodal roots straight into vector embeddings, Google is pushing competitors to stop treating audio and video as downstream text conversions.

There are still obvious friction points, though. A 120-second video cap and a 6-page PDF limit mean enterprise-grade documents and long-form recordings will still require chunking and orchestration on your end. The grunt work has moved, not vanished entirely.

Still, if your application touches multiple media types, we think it is well worth spinning up a test run today. The operational simplicity alone makes it hard to ignore.