Embedding

Definition

A numerical vector representation of text, images, or other data that captures semantic meaning. Embeddings power search, recommendations, and RAG systems by letting you find conceptually similar content.

Why it matters

Embeddings are the plumbing of modern AI applications. Every time you search semantically, every time a RAG system retrieves relevant documents, every time a recommendation engine suggests similar items, embeddings are doing the work. They convert human-readable content into math that computers can compare, cluster, and retrieve at massive scale. For builders, choosing the right embedding model (dimension size, domain specialization, multilingual support) directly impacts the quality of every downstream feature. Embeddings are also where data privacy concerns concentrate: a well-crafted embedding can sometimes be reversed to reconstruct the original content.

In practice

OpenAI's text-embedding-3-large produces 3,072-dimensional vectors and is the most widely used commercial embedding model. Cohere's embed-v3 and Google's Gecko compete on quality and pricing. In the open-source world, BAAI's bge-large and GTE models match or exceed commercial quality. A typical enterprise RAG pipeline embeds documents at ingestion time, stores vectors in a database like Pinecone or Weaviate, and retrieves the top-k most similar documents for each query. Modern embedding models support 8,192+ tokens, enabling chunk-free embedding of long documents. The market has commoditized rapidly: embedding costs dropped 95% between 2023 and 2025.