Vector database: simple definition and real costs

In short

A vector database is a storage system built for searching data by semantic similarity rather than exact keywords: each document is turned into a numeric vector, and search finds the closest ones. It is the technical building block that makes a RAG chatbot on your documents work.

• Text is split into passages, converted into vectors (embeddings), then indexed.
• At each question, the question is turned into a vector and the closest passages are retrieved.
• Common options: pgvector (in PostgreSQL/Supabase), Cloudflare Vectorize, or dedicated databases.
• For an SME, the cost is often included in your existing database: almost zero extra.

The problem, concretely

A classic keyword search fails as soon as words differ: you search for "notice period" but the contract says "termination compensation". A vector database understands meaning: it finds the passage talking about the same subject, even with different words. That is what makes a chatbot on your documents genuinely useful.

How it works, simply

1. Indexing: your documents are split into passages, each turned into a numeric vector capturing its meaning. 2. Storage: vectors are stored with their metadata (source, page). 3. Search: your question is turned into a vector, and the database returns the closest passages. 4. Answer: the model writes an answer only from these passages, with the source.

Which vector databases for an SME

• pgvector: inside your existing PostgreSQL (Supabase), free if the instance has headroom. The right default.
• Cloudflare Vectorize: at the edge, for large volume or global latency.
• Dedicated databases (Qdrant, Pinecone): relevant at very large scale, rarely needed for an SME.

The choice depends on your existing stack: that is exactly what we settle together at the free audit.

Why it matters to you

A well-chosen vector database is the difference between a chatbot that answers "off the mark" and an assistant that cites the right clause. The complete pipeline (indexing, search, citations) is already built and runs in production on beforbuild.com: I adapt it to your documents, at a fixed price from 2 640 € excl. tax, no subscription. Details on the chatbot on your documents page.