A vector database is a specialized storage and indexing system optimized for high‑dimensional embedding vectors and similarity search. It supports approximate or exact k‑nearest neighbor (kNN) queries over vectors, typically with metadata filters, hybrid keyword+vector retrieval, sharding/replication, and consistency guarantees tailored for Retrieval‑Augmented Generation (RAG) and semantic applications.
What is Vector Database?
Vector databases manage the full lifecycle of embeddings: ingestion (upserts, TTL), indexing (HNSW, IVF/IVF‑PQ, ScaNN/ANN variants), distance metrics (cosine, dot, Euclidean), and query‑time filters over structured metadata. They support batching, multi‑tenancy, namespaces, and streaming updates; some add reranking hooks and cross‑modal fields (image/audio/text). Hybrid search pairs sparse (BM25) with dense retrieval to balance recall and precision. Operationally, they expose APIs/SDKs for CRUD, search, and collection/schema management, and integrate with embedding models and RAG orchestrators.
Why it matters and where it’s used
Embeddings unlock semantic search and grounding for LLMs. Vector databases provide low‑latency nearest‑neighbor lookups at scale, enabling RAG, personalization, recommendation, deduplication, anomaly detection, and multimodal search. They improve factuality (by retrieving evidence), relevance (semantic recall), and freshness (near‑real‑time updates) while supporting access controls and tenancy.
Examples
- Hybrid enterprise search: keyword prefilter + vector kNN + cross‑encoder rerank for cited answers.
- Product recommendations: similar‑item retrieval using behavior/text/image embeddings with category filters.
- Dedup and clustering: near‑duplicate detection via cosine thresholds; topic clustering for analytics.
- Multimodal search: unify text queries with image/audio embeddings for cross‑modal retrieval.
FAQs
- How is it different from a traditional DB? It optimizes for vector similarity indexes and ANN latency/recall trade‑offs, not only relational scans.
- Approximate vs exact? ANN (e.g., HNSW, IVF) yields sub‑linear latency with tunable recall; exact search is costlier but precise.
- Which metric should I use? Cosine or dot for normalized embeddings; Euclidean for some models—match training objective.
- Do I store raw docs? Commonly: store text in an object store/DB, keep references and metadata with the vectors.
- How do I scale? Partition/shard by collection/space, replicate for HA; monitor memory, graph/build times, and recall.