Vector Database
Last updated:
What a vector database is
A vector database stores data as vectors: long lists of numbers produced by an embedding model. Texts with similar meaning get vectors that sit close together, so the database can answer “which stored passages are closest in meaning to this one?” rather than “which rows contain this exact word?”.
The term covers both dedicated products built only for vectors and general-purpose databases that have added vector columns and vector indexes. What they share is the ability to run similarity search over large collections.
How it works
When content is added, each piece of text is turned into an embedding and stored together with the original text and metadata such as the source, title or date. At query time, the question is embedded with the same model, and the database returns the stored vectors nearest to it according to a distance measure such as cosine similarity.
Comparing a query with every stored vector is exact but slow for large collections, so most vector databases build an approximate nearest neighbor (ANN) index. These indexes trade a small amount of accuracy for much faster lookups.
- Store: the text, its embedding and its metadata are saved together.
- Index: an ANN index organizes vectors so close neighbors can be found quickly.
- Query: the question is embedded and the closest vectors are returned.
- Filter: metadata filters, such as one customer's data only, narrow the results.
An example
A bike shop has a few hundred help articles. A visitor asks “Can I bring my e-bike back if I change my mind?”. No article contains those words, but the returns policy talks about “refunds after purchase”. Because both texts are about returning a product, their embeddings are close, and the vector database returns the returns policy as the top match. The chatbot then answers from that page.
Why it matters for business chatbots
A vector database is the retrieval layer of most retrieval-augmented generation (RAG) systems. It decides which passages the language model sees before it answers, so its results directly shape answer quality. If the right passage is not retrieved, the model cannot use it.
Vector search alone has a known weakness: exact identifiers such as product codes, model numbers or rare names are not always matched reliably by meaning. That is why many chatbot systems combine vector search with keyword search in a hybrid setup.
How intoCHAT handles it
With intoCHAT you do not set up or run a vector database yourself. When you add a website, a file, a text snippet or a Q&A pair, intoCHAT splits the content into chunks, creates an embedding for each chunk, and stores and indexes them for you. When a visitor asks something, intoCHAT runs a vector similarity search together with a keyword search and merges both rankings before the agent answers.