STUB. This is retrieval infrastructure, not a governance control. The governance-critical part of retrieval is entitlement-aware-rag — read that page. A vector store with no permission enforcement is exactly how a RAG system surfaces documents a user was never allowed to see.

Business objective

The AI’s card catalog: turns documents into searchable “meaning” (embeddings) so the model finds the right snippet to ground its answer. Semantic search and retrieval infrastructure — the index layer between your content-sources and the enterprise-ai-assistant.

When you need it

Day-1 if you are doing RAG; otherwise N/A. It is a build choice (which vector database / search service), not a security control. Many assistants ship their own retrieval and you never touch a standalone vector store. The governance question is not “which vector DB” but “does retrieval respect the user’s permissions” — and that lives in entitlement-aware-rag.

Security role

A sensitive-data store, not a control. A vector index is a copy of your sensitive content, often stripped of the source system’s ACLs — which is the classic RAG failure mode: the permissions lived on the SharePoint folder, not on the embedding. Whether the index re-checks entitlements at query time is the whole game. Lives in the yellow/green zone.

Vendors

Infrastructure choices, not a competitive survey shortlist. Pages exist for four:

• azure-ai-search — Azure’s managed search/vector service; default for Microsoft-stack RAG, supports security-trimming via document-level ACLs.
• pinecone — managed vector database; popular standalone choice for custom RAG builds.
• weaviate — open-source / managed vector database.
• opensearch — open-source search engine with vector support; the self-hosted option.

Also in this space (named for completeness; no vendor pages — they add nothing to the governance story the four above do not already carry):

• Qdrant — Apache-2.0 Rust engine; the filtering/performance pick. Hybrid Cloud keeps the data plane in your infrastructure. $50M Series B, 2026-03.
• Milvus / Zilliz — Apache-2.0 (LF AI & Data); the billion-scale option, heaviest to operate.
• Chroma — embedded-first, Apache 2.0; the prototyping default. Chroma Cloud GA 2025-08.
• LanceDB — embedded + cloud over the Lance columnar format; the multimodal angle, and a common agent-memory substrate.
• pgvector, MongoDB Atlas Vector Search, Redis, Vespa, Amazon S3 Vectors, turbopuffer — the “use what you already have” and object-storage-native options. For most funds pgvector on existing Postgres is the right default.

Full comparison, market shapes, and consolidation dynamics: vector-database-landscape.

Consolidation / M&A dynamics

Not a seed-flagged M&A category, but the standalone vector DB is under pressure — bundled from below (every OLTP engine and search engine added vectors) and undercut from above by object-storage-native indexes. pinecone is reportedly exploring a sale (unconfirmed, no deal announced as of 2026-08-24), while the Apache-2.0 engines keep raising. Detail and sourcing: vector-database-landscape.

Adjacent categories

• entitlement-aware-rag — the layer that makes retrieval permission-aware; the part that matters for governance.
• content-sources — the systems of record this index is built from.
• enterprise-ai-assistant — the consumer of retrieved snippets.

Survey

No standalone survey question. Vector database choice is an engineering decision, not a CTO governance survey item. Where it matters, it is captured under entitlement-aware-rag and enterprise-ai-assistant.

Open taxonomy questions

• Like content-sources, this is infra rather than a buying-decision category. Kept as a stub to preserve the layer-cake row; the security-relevant content all routes to entitlement-aware-rag.
• [2026-08-24] Confirmed the stub stance rather than minting stub pages for Qdrant/Milvus/Chroma/LanceDB — coverage discharged by vector-database-landscape. Open: whether the RAG platforms (Vectara, Contextual AI) earn pages under enterprise-ai-assistant / entitlement-aware-rag.