What is a vector database?
The part of search that matches meaning instead of wording. Less a product than a function, and one your machine performs comfortably.
A vector database stores embeddings, number representations of passages, and finds the ones closest in meaning to a query. Instead of matching exact words, it matches geometric closeness: passages about similar ideas end up near each other in the space, so a question retrieves what it means rather than what it repeats. It is the index behind semantic search, and on a personal scale it is just a file your own machine searches.
Why meaning needs an index.
An embedding turns a passage into numbers that carry its meaning. On its own that is a description; the database is what makes it useful. It files every vector so that a question, itself turned into a vector, can find its nearest neighbors in milliseconds instead of comparing against every passage one by one. This is the second half of the hybrid search that a retrieval loop runs: keywords catch the exact terms, the vector index catches the reworded idea, and each covers what the other misses.
The scale most people actually have.
The vector databases that get written about are server products built for millions of entries and many simultaneous readers. A personal research library is not that. Tens of thousands of passages produce an index that measures in megabytes, fits on a laptop, and searches fast on a CPU with no server process at all. The honest conclusion is that you need similarity search, not necessarily a vector database product: a tool that keeps the vectors beside your documents and searches them directly gives you the same answers with one less moving part and nothing to upload. The collection it searches is the knowledge base, and what a knowledge base is covers what the collection itself contributes.
What the vectors point back to.
Vectors alone are unreadable, a cloud of coordinates that no one can quote. What makes an index worth having is the link from each vector back to the chunk of text it was computed from. That pointer is where citations come from: retrieval surfaces a vector, the vector names its passage, and the passage travels into the prompt so the answer can quote it and you can check it. A local index keeps those pointers on your disk next to the documents themselves, which is why an answer can carry a citation without your library ever leaving home.