Embeddings are a ubiquitous term in discussions surrounding AI models, particularly in GenAi. While numerous resources explore the mathematical and theoretical underpinnings of embeddings and their significance in training Transformers or ML models in general, there remains a scarcity of material exploring their practical applications. Embeddings serve as data structures containing contextual information essential for executing intelligent tasks.
Among the myriad applications of AI, semantic search stands out as one of the most prominent, directly influenced by the quality and scale of embeddings. The selection of embeddings significantly impacts search capabilities and the supported modalities, ranging from text-only to image search or even a fusion of both, known as Multi-Modality search.
Recognizing embeddings as pivotal data structures, devising efficient storage mechanisms is imperative. Vector databases emerge as specialized solutions optimized for storing and retrieving embeddings.
In this presentation, I will delineate establishing a reference architecture using Vector Database & Embeddings for intelligent search, encompassing options for text, image, and multi-modal searches. Also, I will highlight a few applications leveraging this architecture and outline potential avenues for future exploration.