# Gemini Embedding 2 - Audio, Text, Images, Docs, Videos Source: https://www.youtube.com/watch?v=zUkKvWBJ_0I Recap page: https://rapidrecap.app/video/zUkKvWBJ_0I Generated: 2026-03-11T13:36:28.522+00:00 --- ## Quick Overview Gemini Embedding 2 introduces a unified multimodal embedding space, allowing users to query text, images, audio, video, and PDFs simultaneously with a single model, overcoming the complexity and inefficiency of previous multi-model RAG pipelines by achieving state-of-the-art performance across various benchmarks, including strong speech capabilities. **Key Points:** - Gemini Embedding 2 maps text, images, audio, video, and PDFs into a single unified embedding space, enabling true multimodal retrieval. - The model supports flexible output dimensions, scaling down from a default of 3072 dimensions to 1536, 768, for balancing performance and storage costs. - It natively ingests and embeds audio data without needing intermediate text transcriptions, supporting up to 120 seconds of video input. - The model directly embeds PDFs up to 6 pages long, processing visual and text content concurrently. - In cross-modal similarity tests, the model correctly matches text descriptions to corresponding images and audio clips, demonstrated by high cosine similarity scores between related inputs (e.g., 'Trees and nature sounds' vs. 'trees.wav'). - Gemini Embedding 2 outperforms previous models like gemini-embedding-001, multimodalembedding-001, Amazon Nova 2 Multimodal Embeddings, and Voyage Multimodal 3.5 on key benchmarks like MTEB (Multilingual/Code) and various cross-modal tasks. - The unified approach simplifies the RAG pipeline, replacing complex, slow, and expensive multi-model/multi-vector store architectures with one model and one unified vector store. ![Screenshot at 1:04: Demonstration of Gemini Embedding 2 unifying inputs \(Text, Image, Audio, Video, PDF\) into a single unified embedding space, contrasting with the complex, tangled 'Old Way' pipeline.](https://ss.rapidrecap.app/screens/zUkKvWBJ_0I/00-01-04.jpg) **Context:** This video introduces and demonstrates Gemini Embedding 2, Google's latest model for creating high-quality multimodal embeddings. The presentation contrasts the new unified approach with the older, fragmented method that required separate embedding models and vector stores for each data type (text, image, audio, etc.). The core benefit highlighted is the ability to perform unified, cross-modal retrieval—like searching for videos using an image query—all processed by a single Gemini model. ## Detailed Analysis The video explains the capabilities and architecture of Gemini Embedding 2, emphasizing its move away from fragmented multimodal RAG pipelines that relied on multiple, separate models and vector stores, which were noted as expensive to maintain, slow to run, and messy to build. Gemini Embedding 2 solves this by mapping text, images, audio, video, and PDFs into a single, unified embedding space using one model (0:00, 1:16, 2:15). This allows for direct cross-modal similarity searches, such as finding an image that best matches an audio clip or a text description. The model supports flexible output dimensions, scaling down from 3072 dimensions to 1536 or 768, offering a balance between performance and storage costs (10:03). Specific capabilities include handling up to 8192 text tokens, processing up to 6 images per request, supporting up to 120 seconds of video input, natively ingesting audio without intermediate transcriptions, and directly embedding PDFs up to 6 pages long (10:09). Benchmark results (0:38) show Gemini Embedding 2 achieving state-of-the-art performance across text-text, text-image, text-document, text-video, and speech tasks, significantly outperforming legacy models. The notebook demonstration confirms this by showing high cosine similarity scores between related modalities (e.g., text descriptions matching their corresponding images and audio clips) and showcases how to aggregate text and image inputs into a single embedding for post-level representations (13:19, 17:47). ### Introduction to Multimodal Search - Concept of an 'All in one search' system - Text, Image, Audio, Video, PDF inputs mapped to a single search bar - Contrast with 'The Old Way' requiring multiple models and vector stores (0:00-0:27) ### Gemini Embedding 2 Core Functionality - Converts any piece of content (text, image, audio, video, PDF) into a vector (list of numbers) in a high-dimensional space (2:15, 2:49) ### Vector Space Semantics - Similar content lives in the same neighborhood/cluster in the high-dimensional space (2:59) ### Key Features and Limits - Supports 8192 text tokens, up to 6 images, up to 120 seconds of video, native audio ingestion, and direct PDF embedding up to 6 pages (10:09-11:18) ### Performance Benchmarks - Achieves state-of-the-art performance, outperforming models like gemini-embedding-001 and Voyage Multimodal 3.5 across various cross-modal tasks (0:38, 9:38) ### Code Demonstration - Shows helper functions for encoding text, images (PNG/JPEG), and audio (WAV/MP3) - Demonstrates text-to-image and text-to-audio similarity via cosine similarity heatmaps - Shows how to aggregate text and image into a single embedding call (11:21-15:36) ### Embedding Aggregation - Explains Single Content Entry (one aggregated embedding) vs. Multiple Entries (separate embeddings) and Post-level representation via averaging (17:29-17:58) ![Screenshot at 0:00: Diagram illustrating the proposed unified search system taking Text, Image, Audio, Video, and PDF inputs into a single search bar, contrasting it with the complex 'Old Way' \(0:15\)](https://ss.rapidrecap.app/screens/zUkKvWBJ_0I/00-00-00.jpg) ![Screenshot at 1:15: Detail of Gemini Embedding 2's input capabilities, including support for up to 120 seconds of video and native audio ingestion without text transcription.](https://ss.rapidrecap.app/screens/zUkKvWBJ_0I/00-01-15.jpg) ![Screenshot at 3:34: Diagram summarizing the core improvement: multiple modalities inputting into Gemini Embedding 2 to produce results from one unified vector space, contrasted with the complex, tangled 'Old Way' pipeline.](https://ss.rapidrecap.app/screens/zUkKvWBJ_0I/00-03-34.jpg) ![Screenshot at 9:38: Benchmark table showing Gemini Embedding 2 achieving superior or state-of-the-art scores across numerous Text-Text, Text-Image, Image-Text, Text-Document, Text-Video, and Speech metrics compared to other leading models.](https://ss.rapidrecap.app/screens/zUkKvWBJ_0I/00-09-38.jpg)