Knowledge
Knowledge HubConcepts and strategic researchGlossaryDefinitions across six domainsFrameworksStructured strategy methodsTopicsConnected subject areasLibrary
PlaybooksPractical strategy guidesBooksBooks and long-form researchOperator VaultMember resources and toolkits
Explore Discover, the Knowledge Hub, playbooks, and the Operator Vault.
Glossary
Vector database
A vector database is a data system optimized to store embedding vectors and retrieve nearest neighbors efficiently — often with metadata filters — so applications can find similar items at scale. It is infrastructure for similarity search, not a synonym for embeddings or for RAG itself. Related phrases such as “vector store” appear in product and engineering language but do not always denote the same specialized database class.
Why it matters
Once an organization embeds large corpora, it needs durable storage, indexing, filtering, and operational controls. Specialized vector databases — and vector-capable search engines — provide approximate nearest-neighbor indexes, metadata predicates, and operational features that turn embeddings into a production retrieval service.
Architecture debates should separate representation (embeddings), storage/index (vector database or search engine), and generation (LLM). Substituting a vendor name for the whole RAG stack obscures failure modes such as stale indexes, poor chunking, or incomplete authorization across the retrieval path.
Example (enterprise): An HR knowledge assistant stores embeddings of policies with metadata for region and employment type. The application passes the authenticated user’s identity into retrieval so filters and policy checks limit which vectors are eligible — then passes only authorized passages to an LLM. Metadata fields help; they do not automatically guarantee access control unless identity, authorization, and retrieval are enforced end to end.
Mechanisms
- Approximate nearest-neighbor (ANN) indexes — trade exactness for latency and scale on high-dimensional vectors.
- Metadata filtering — combine similarity with structured constraints (tenant, ACL labels, time, product line) when the application supplies correct identity and policy context.
- Upsert and re-index workflows — keep vectors aligned as source documents change.
Boundaries
- Vector database ≠ embedding model.
- Vector database ≠ “vector store” as a guaranteed exact synonym — vector store may mean a lighter index, library, or feature inside a general search platform.
- Vector database ≠ semantic search UX; it is a retrieval backend.
- Filters do not replace application-layer authorization, document ACLs, or post-retrieval policy checks.
Related topics
None linked yet.
Related concepts & frameworks
Related Discover collections
Further research
No related resources yet.