Skip to main content
AI & Agents

Vector Database

A vector database stores text, images or other content as numerical embeddings and retrieves items by semantic similarity rather than exact keyword match. It is the storage layer most retrieval-augmented systems use to find passages related in meaning to a query, even when they share no words with it.

The practical choice is less exotic than the category name suggests. Pinecone, Weaviate, Qdrant and Azure AI Search are all capable, but for a large share of mid-size corpora pgvector alongside an existing Postgres instance is the correct, boring answer — no new infrastructure to operate, and one less system in the failure path. Be sceptical of any recommendation made before someone has looked at your scale and latency requirements.

A common and costly error is relying on vector similarity alone. Embedding search performs poorly on exact identifiers — part numbers, clause references, error codes — which is precisely what enterprise users search for most. Combining vector search with keyword search and a reranking pass fixes this, and its absence is the single most frequent defect in underperforming retrieval systems.

Codazz builds this in production — RAG Development.

FAQ

Vector Database
FAQ.

Common questions about vector database.

Ask Us Anything

Often not. If you already run Postgres, the pgvector extension handles a large share of real workloads without adding infrastructure. Dedicated vector databases earn their operational cost at high scale, with demanding latency targets, or where you need features like sophisticated metadata filtering. The decision should follow measured requirements, not the assumption that a new category needs a new product.

A traditional search engine matches keywords; a vector database matches meaning, so it can find a relevant passage that shares no words with the query. In production the two are usually combined, because keyword search remains better at exact identifiers while vector search is better at conceptual similarity. Hybrid retrieval with reranking outperforms either alone.