Answers are grounded in this paper and cite their sources.
Purpose: Vector databases have become the default persistence layer for machine-learned embeddings, and with it they have inherited the full problem set of database management: storage and indexing, query processing under predicates, updates, distribution, access control, and evaluation. This survey examines the vector database as a data management system, rather than as a search library or a retrieval component. Methods: The survey organizes a corpus of 77 primary-verified sources — 59 archival papers, 13 preprints, three documentation records, and two benchmark project sources, each verified against Crossref, the arXiv record, or the publisher's page — into five branches: index structures, system architectures, query processing, operations, and evaluation. Each branch is synthesized comparatively, and the evidentiary status of claims is marked at the point of reliance. Results: The index literature is rich and self-correcting, and filtered search has, since 2023, built a design literature. The systems literature is concentrated in a few research groups and falls silent where deployment is loudest, because the most widely deployed native systems publish no architecture papers. The properties that make a store a database — consistency semantics, tenant isolation, optimizer behavior — are those on which the literature says least: access control rests on two 2026 papers, and embedding-model upgrade operations on none. Conclusion: A nine-item gap register, each item grounded in documented searches of the primary record, closes the survey as a research agenda ordered by consequence for production use. The vector database's unfinished work is the database half of its name.
No takes yet. Share an insight, caveat, or question.
Vinuraj Devaraj (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: