A vector store is a database that holds embeddings and returns the nearest ones to a query vector, usually alongside the original text and metadata.
The store's job is simple: keep vectors, keep the text and metadata each vector came from, and answer 'which entries are closest to this vector' quickly. Products range from managed cloud services (Pinecone), to extensions of existing databases (Postgres pgvector), to in-memory or file-based stores (a JSONL file with one record per chunk). For a personal or small-team corpus, a flat file searched with cosine similarity is fast enough and has no server to run.
The choice of store decides where your data lives. A hosted store means every chunk of your documents is copied to that provider. A local store keeps them on disk. Metadata matters as much as vectors: a hit is only useful if it comes back with the file path, page number and chunk position so a person can check it. Stores that return only similarity scores and text push that burden back on the reader.