Vector Database

A store for embeddings that allows semantic search, powering similarity lookups for prompts or recommendations.

A vector database stores embeddings and supports similarity search, enabling semantic retrieval by meaning rather than exact text match.

Teams use it for RAG, recommendations, deduplication, and semantic search. It returns nearest neighbors to a query embedding with filters.

In systems, it pairs with an embedding service and metadata filters. It improves relevance for natural language queries and powers context fetch for LLMs.

Frequently Asked Questions

How do I choose a vector database?

Consider ANN performance, filtering, scaling, latency, cost, and ecosystem. Evaluate on your data and query patterns.

What is ANN search?

Approximate Nearest Neighbor search speeds up similarity queries by trading exactness for fast, scalable lookups.

How do I handle metadata filters?

Use hybrid queries that combine vector similarity with metadata constraints (e.g., type, locale, permissions).

How big should chunks be?

Small enough to be specific (e.g., paragraphs), large enough for context. Test chunk sizes against retrieval quality.

How do I keep the index fresh?

Re-embed changed content and upsert into the index. Handle deletions with tombstones or hard deletes.

How do I evaluate retrieval quality?

Use labeled query-doc pairs, measure recall/precision, and monitor user signals like click-through or success rates.

Can I store sensitive data?

Yes, with access controls, encryption, and segregation. Consider self-hosting for regulated data.

Do I need exact search too?

Often yes—combine vector search with keyword filters or reranking for precision on IDs/names.

What about cost and latency?

Optimize dimensions, use batching, and prune unused vectors. Choose infrastructure that meets your latency budget.

Hourglass background
Ready to move faster

Ship glossary-backed automations

Plan Your First 90 Days