What is a vector database?

What is it?

A vector database is specialized storage for embedding vectors. It supports fast similarity search: given a query vector, find the stored vectors closest to it.

Traditional databases excel at exact matches (IDs, dates, keywords). Vector databases excel at "find the most similar items" across millions of vectors.

Why does it exist?

After embedding thousands of document chunks, you need efficient search. Comparing one query against every vector in a flat list becomes slow at scale. Vector databases use indexes (approximate nearest neighbor structures) to return top matches quickly.

They may also store metadata (document ID, chunk index, tags) alongside vectors for filtering.

How does it work conceptually?

Store: chunk vectors + metadata in collection Query: question vector → nearest neighbor search → top-K similar chunks

Key concepts:

• Collection — a named set of vectors (like a table)
• Dimensionality — length of each vector (must match the embedding model)
• Similarity — distance metric between vectors

Example

Collection contains vectors for manual chunks. User question about "500-hour maintenance" is embedded. The vector database returns the three chunks whose vectors are closest to the question vector.

Where does it fit in a RAG pipeline?

The vector database sits between embedding and retrieval:

Chunks → Embeddings → Vector database (index) ↓ Question → Query embedding → Similarity search → Retrieved chunks

Without a working vector database, semantic retrieval cannot run — even if documents are uploaded and the LLM is configured.

How this relates to Noulek

Noulek currently integrates with Weaviate as its vector database. Configure it in Settings → Vector DB:

• Weaviate Endpoint URL
• API Key (if required)
• Collection name

Use Test connection to verify connectivity. Indexed document embeddings are stored in the configured collection during the Indexing pipeline stage.

Connection or schema issues surface as Not connected status, test failures, or retrieval errors (NX_RETRIEVAL_FAILED). See Vector Database connection problems.

Other RAG systems may use Pinecone, pgvector, or other stores — the concept is the same even when the vendor differs.