What is a Vector DB?
Vector databases, told as a story about a shelf where books are placed in order of their coordinate tags.
A million books, every one tagged. But measuring against each one by one takes a day.The coordinate tags are there. The problem is how slowly you find things.
An ordinary shelf is sorted by name or number. There is no meaning order.Kitty and meow-cat mean the same yet sit far apart. So to find near tags, you must search everything.
A vector DB is a shelf organized by tag coordinates. Near tags sit in near cells.If the question tag is (3, 9), only the animals cell needs checking — not all million books. What a tag is: the coordinate tag story.
Pull just a few books in closest order from nearby cells. Then filter once more by label.Ask for five and you get five — that is top-5. A near tag with the wrong label gets skipped.
Checking only nearby cells sometimes misses a better book. And the shelf finds — it never answers.It is fast but approximate. Handing the found book to the parrot so it can answer is the page the librarian brought.
Vector DB = a shelf organized by tag coordinates. Near tags sit in near cells, so you search only nearby cells, then filter by each book's label. Fast but approximate — and it finds, never answers.
A store that keeps embedding vectors in an index (HNSW and others) and uses approximate nearest-neighbor search to return the closest top-k quickly. Supports metadata filters, chunk-level storage and hybrid keyword+vector search; changing the embedding model means re-indexing everything. It is the retrieval layer of RAG.
When grown-ups say it
- Vector database
- The shelf organized by tags. Pinecone, Milvus, pgvector, Chroma are this shelf.
- Index (HNSW)
- How the cells are laid out. Bridges between near tags let you reach the neighborhood in a few hops.
- Approximate nearest neighbor
- Searching nearby cells only. Fast because it skips most books — but sometimes misses a better one.
- top-k
- The k closest books. Ask for 5, get 5. Pull too many and the tray overflows.
- Metadata filter
- Filtering by the book's label. Rules like language, year, team — used alongside tag distance.
- Chunk
- Shelving a book as pages. Tags go on a few paragraphs at a time, not a whole book — so they fit on the tray.
- Hybrid search
- Letters + tags. Exact things like names and numbers by letters, meaning by tags — both at once.
- Embedding
- The coordinate tag. The tag that goes on the shelf. → a coordinate tag on every word