Source: Google Developers Blog — 2026-07-13
Summary
Google shipped general availability of Gemini Embedding 2, its first embedding model natively trained to place text, images, video, audio, and PDFs into a single shared embedding space across 100+ languages, with task-specific prefixes and Matryoshka Representation Learning that lets vectors be truncated to smaller sizes without retraining. Google highlights a real customer case: fashion rental platform Nuuly's visual search Match@20 accuracy rose from 60% to 87%, with product-identification rate climbing from 74% to 90%.
Key Takeaways
- One embedding space now covers text, image, video, audio, and PDF inputs together — previously multimodal RAG often meant stitching together separate embedding models per modality and reconciling their spaces.
- Matryoshka Representation Learning lets a single 3,072-dimension vector be truncated down to 1,536 or 768 dimensions on demand, trading storage/speed for accuracy without training a separate smaller model.
- Max input is 8,192 tokens of text plus up to 6 images per call, and Batch API pricing runs at 50% off standard rates for non-real-time indexing workloads.
- Nuuly's real production numbers — Match@20 accuracy 60%→87%, product ID rate 74%→90% — are a concrete before/after on a genuine visual-search use case, not a synthetic benchmark.
Reel Script
Hook (17s / 38 words) A fashion rental company just took their visual search accuracy from 60% to 87% by swapping one embedding model. That's not a typo — and it's not even the most interesting part of what shipped.
Core Concept (65s / 140 words) An embedding model turns any piece of content — text, an image, a video clip, even a PDF — into a list of numbers, a vector, that captures what that content actually means. Think of it as a mathematical fingerprint: things that mean similar things end up with similar fingerprints, which is how search-by-meaning and recommendation systems work under the hood. Until now, if you wanted to search across text and images together, you typically needed separate embedding models for each type, then had to awkwardly reconcile two different fingerprint systems. Gemini Embedding 2 is trained from the ground up to put text, images, video, audio, and PDFs all into the exact same fingerprint space — so a text query can directly match against an image, video frame, or document without a translation step in between.
Hands-On (60s / 130 words) Two technical details worth knowing if you're actually building with it. First, it supports Matryoshka Representation Learning — the model outputs one big 3,072-dimension vector, but you can truncate it down to 1,536 or even 768 dimensions and it still works well, because the most important information is front-loaded in the vector, like a nested doll where the smaller version inside is still a complete doll. That means you can shrink your storage and search costs without retraining anything. Second, the real number: Nuuly, a clothing rental platform, used it for visual search and their accuracy metric — Match@20 — jumped from 60% to 87%, while correctly identifying the actual product went from 74% to 90%. That's a production deployment, not a lab benchmark.
Takeaway (23s / 50 words) If you're running separate embedding pipelines for text and images today, that's now solved architecture, not a research problem. Go check whether consolidating into one multimodal embedding space would simplify your RAG stack and improve your match rates the way it did for Nuuly.