Vector Database

No jargon, no math

A vector database stores embeddings, the numeric fingerprints of content, and answers one question fast: which stored fingerprints are closest to this new one? Inside, each record is an array of a few hundred floats plus a pointer back to the original text. To avoid comparing against every record, the database builds an index, most often a graph (HNSW) that links each vector to its near neighbours, so a search walks the graph and reaches the closest matches in milliseconds, even across millions of rows. You insert, you query with a vector and a top-k count, you get back candidates with distance scores. Where it breaks: the index is built for a specific embedding model, so swap models and you re-index; stale embeddings quietly poison retrieval; and a vector match is resemblance, not truth, so production systems add filters and reranking on top.