What is a vector database?
The video is not up yet. The full script is below — it reads fine on its own.
The one-sentence version
A vector database finds the records that are closest in meaning to your query, rather than the ones that contain your exact words.
Why exact matching ran out
A traditional index is built for equality and ranges. Ask it for rows where
city = 'Delhi' and it will answer instantly and correctly. Ask it which
support tickets are about the same problem as this one and it has nothing to
offer, because "the same problem" is not a value you can put in a column.
Embedding models change what you store. Text goes in, and a few hundred or few thousand numbers come out — coordinates in a space where things that mean similar things land near each other. "Refund not received" and "my money never came back" end up as neighbours without sharing a single word.
Then the problem becomes geometry
Once every document is a point, retrieval is just "find the nearest points to this one". With a million documents you could compare against all of them, and with a hundred million you cannot. That is the actual job of a vector database: approximate nearest neighbour search, which trades a small, tunable chance of missing a true neighbour for an enormous speedup.
- HNSW builds a navigable graph and walks it, hopping toward closer neighbours.
- IVF partitions the space into cells and only searches the cells near your query.
- Quantisation compresses each vector so more of the index fits in memory, at some cost in precision.
Where this shows up
Retrieval-augmented generation is the obvious case: a language model has no memory of your documents, so you retrieve the relevant passages and hand them over with the question. The quality of that retrieval sets the ceiling on the answer. No amount of prompt engineering rescues a model that was handed the wrong three paragraphs.
When you do not need one
Under roughly a hundred thousand vectors, a brute-force scan with NumPy — or the vector extension your existing database already ships — is fast enough and one fewer system to operate. Reach for a dedicated vector database when scale, filtered search, or real-time updates actually demand it, not because the architecture diagram looks better with one on it.