What is a vector database?

What it does

An embedding model turns a sentence, a photo or a video frame into a vector, a list of a few hundred to a few thousand numbers. Items with similar meaning end up with similar vectors. A vector database stores those vectors, usually next to an ID and some metadata (which file, which timestamp, which speaker), and answers one kind of question very quickly: which stored vectors are nearest to this one?

Nearness is measured with a distance such as cosine similarity or dot product. Comparing a query against every stored vector is exact but gets slow as the collection grows into the millions. So most vector databases build an approximate nearest neighbor (ANN) index, such as a graph that lets the search hop quickly toward the right neighborhood, and check only a small part of the collection.

An example

Imagine 500 hours of transcribed interviews split into 200,000 short passages, each stored as a vector with its file name and start time. You search “the moment someone admits they were wrong.” Your query becomes a vector, the index returns the 50 closest passages in a few milliseconds, and each result carries the file and timestamp needed to jump straight there. A filter on the metadata, say “only 2024 recordings,” can narrow the set before or after the vector lookup.

Limits and trade-offs

• Approximate means approximate. ANN indexes trade a little recall for a lot of speed, so a relevant item can occasionally be missed.
• Vectors are tied to their model. Switch embedding models and every stored vector has to be recomputed; vectors from different models are not comparable.
• Similar is not the same as correct. Nearest neighbors can be related but wrong, which is why many systems add keyword matching (hybrid search) and a reranking pass.
• Vectors are not anonymous. Embeddings can leak information about the text or images they came from, so a hosted vector store deserves the same privacy care as the source files.
• You may not need a separate product. Small and medium collections often work fine with a vector index inside an ordinary database or library.

In MediaFind

MediaFind computes embeddings for your transcripts, frames and photos on your own computer and keeps them in one local index alongside the rest of your library. There is no cloud vector service, account or API key involved, and nothing is uploaded to build or query it. That index is what lets a plain-language description land on an exact timestamp. For how the pieces fit together, see semantic search and retrieval-augmented generation.

Frequently asked questions

Is a vector database the same as a search engine?

It is one part of one. It finds nearest vectors, but a full search engine also handles keyword matching, filters, ranking and showing results.

Do I need a vector database for RAG?

You need some way to find relevant passages by meaning. For a small collection that can be a simple in-memory index; a dedicated vector database helps mostly at larger scale.

Can a vector database run on a laptop?

Yes. Many vector indexes are libraries that run locally, and a personal media library is well within what a laptop can search quickly.