Imagine searching not just with words, but with images, or even a combination of both. That's the promise of Universal Multimodal Retrieval (UMR), a cutting-edge technology that aims to revolutionize how we search for information. Current search engines mostly focus on text, but the world is increasingly multimodal, filled with images, videos, and audio. UMR aims to bridge this gap, allowing users to query with any modality and retrieve results from diverse sources. One of the most promising approaches to UMR leverages the power of Multimodal Large Language Models (MLLMs). These AI powerhouses, trained on massive amounts of text and image data, possess a remarkable ability to understand and connect information across different modalities. Researchers have developed an innovative framework called General Multimodal Embedder (GME), which uses MLLMs to create a shared embedding space for different modalities. This means that text, images, and image-text combinations can be represented as vectors in the same space, making it possible to directly compare and search across them. But there's a catch: training effective MLLMs for UMR requires vast amounts of diverse, high-quality multimodal data, which is currently scarce, especially for complex queries involving combined image and text elements. To address this data scarcity, researchers have devised an ingenious solution: an automated data synthesis pipeline. This pipeline leverages the generative capabilities of LLMs and MLLMs to create synthetic multimodal training examples, effectively amplifying the limited real-world data available. The pipeline generates queries from existing text, identifies key entities within these queries, and then either retrieves or generates related images to create fused image-text queries. This not only expands the training data but also diversifies it, covering a wider range of topics and scenarios. Early results are promising. GME, trained with this synthesized data, achieves state-of-the-art performance on a newly compiled benchmark called UMRB, outperforming existing methods across various single-modal, cross-modal, and fused-modal retrieval tasks. UMR, powered by MLLMs and innovative data synthesis techniques, holds tremendous potential to transform our search experience, opening up exciting new ways to discover and access information in an increasingly multimodal world. However, challenges remain, such as scaling the training of MLLMs and ensuring that synthetic data truly represents the complexity of real-world scenarios. The journey towards truly universal multimodal search continues, and the breakthroughs outlined in this research pave the way for a future where search is as intuitive and flexible as our own perception.