What is Vector Embeddings?

Vector embeddings translate human concepts into a mathematical language that machines can reason about. When you embed the phrase "quarterly revenue report" using a model like text-embedding-3-large, you get a 3072-dimensional array of floating-point numbers. Crucially, semantically similar concepts produce vectors that are geometrically close together, "Q3 earnings summary" would be near "quarterly revenue report" in vector space, even though the words are completely different. This enables semantic search: instead of matching keywords, the system finds documents with similar meaning. The technical pipeline involves three steps. First, chunking: large documents are split into semantically coherent segments (using recursive text splitters that respect paragraph and section boundaries). Second, embedding: each chunk is passed through an embedding model that outputs a fixed-dimensional vector. Third, indexing: vectors are stored in a specialized database (pgvector, Pinecone, Milvus) with approximate nearest neighbor (ANN) indexes like HNSW for millisecond-speed retrieval. Advanced implementations use multi-modal embeddings that can represent images, audio, and text in the same vector space, enabling queries like "find all inspection photos that show corrosion" without any manual tagging.