How embeddings work
An embedding model converts text into a dense vector — typically a list of 1,536 or 3,072 floating-point numbers. The key property: texts with similar meaning end up close together in this number space. "How do I return a product?" and "What's your refund process?" produce nearly identical embeddings, even though the words differ.
This mathematical similarity is what powers semantic search, which is far more useful than keyword matching for real-world queries.
Practical applications
Semantic search: Search your help docs, product catalogue, or knowledge base by meaning, not just keywords. Essential for RAG systems.
Clustering: Group customer feedback, support tickets, or survey responses by theme automatically. Discover patterns that manual categorisation misses.
Recommendations: "Customers who viewed this service also enquired about..." — powered by embedding similarity between service descriptions and browsing behaviour.
Embedding providers
OpenAI, Cohere, and Voyage AI offer embedding APIs. For most business use cases, OpenAI's text-embedding-3-small model provides excellent quality at very low cost. Embeddings are stored in vector databases for efficient retrieval.