A vector database is a database optimised for storing and querying high-dimensional vectors via approximate nearest-neighbour (ANN) search. Traditional relational and document databases are built around exact key lookups and range queries; vector databases are built around the question "find the k vectors in this dataset most similar to the query vector" where similarity is cosine, dot product, or Euclidean distance. The workload exploded after 2022 when RAG-based AI architectures made vector similarity search the critical infrastructure layer between embedding models and LLMs.
The 2026 vector-database landscape has three categories. **Dedicated vector databases** — Pinecone, Weaviate, Qdrant, Milvus, Chroma — ship purpose-built for vector workloads with first-class metadata filtering, hybrid search, and multi-tenancy. **Postgres with pgvector** has become a credible alternative for teams already running Postgres at modest scale, with growing feature parity for hybrid search. **Purpose-built search engines** — Elastic, Opensearch, Vespa — have added vector search alongside their existing full-text capabilities, which is often the right answer for organisations that need both lexical and semantic search over the same corpus. **Cloud-native options** — AWS OpenSearch, Google Vertex AI Matching Engine, Azure AI Search — give managed-service convenience with the tradeoff of vendor lock-in.
For APAC mid-market enterprises, the right choice depends on scale and existing infrastructure. Below 10M vectors and Postgres already in the stack, pgvector is usually the simplest answer. Beyond that scale, or when hybrid search quality matters, a dedicated vector DB or search engine is the stronger choice. Data residency constraints in Japan, Korea, Singapore, and Hong Kong often drive self-hosted options (Qdrant, Milvus, Weaviate, pgvector) over SaaS offerings.
The non-obvious operational note: **vector search performance is dominated by the ANN index structure, not the underlying storage engine**. HNSW (Hierarchical Navigable Small World graphs) is the default for most workloads; IVF-PQ (inverted file with product quantisation) wins for very large corpora where memory matters. Understand the recall-latency-memory tradeoff of your chosen index — most silent RAG quality problems trace to recall@k dropping below 80% because an index was configured for latency at the expense of accuracy. Measure recall explicitly on labelled queries before optimising anything else.
Where AIMenta applies this
Service lines where this concept becomes a deliverable for clients.
Beyond this term
Where this concept ships in practice.
Encyclopedia entries name the moving parts. The links below show where AIMenta turns these concepts into engagements — across service pillars, industry verticals, and Asian markets.
Other service pillars
By industry