Vector databases

A vector database stores embeddings, the numeric vectors that AI models produce from text, images or code, and finds the nearest ones to a query, usually alongside keyword search and metadata filters. Teams use them for retrieval-augmented generation (RAG), semantic and hybrid search, recommendations and agent memory. The managed services differ most in how they bill: per read and write request, per vector dimension stored, per provisioned node, or per byte stored and scanned. They also differ in which plan carries backups, single sign-on and a HIPAA agreement, and in whether an open-source engine with the same API lets you leave and run it yourself.

We compare them on the same six indicators, read from each vendor's pricing pages and calculators, trust, legal and documentation pages:

• Pricing clarity: whether the plan you need has published unit prices, can be bought online and month to month, whether a minimum spend is shown on the pricing page, and whether the pages leave prices out or contradict each other.
• Free tier and trial: a permanent free tier, how much it holds, whether it includes hybrid search, pauses when unused or needs a card; otherwise the trial or credits.
• Price for the reference scenario: the monthly cost of 1M vectors of 1,536 dimensions with about 1 KB of metadata each, 1M queries and 1M upserts a month in one US region, on the cheapest self-serve plan that can run it, with the arithmetic shown; 0 when the pages cannot price it.
• Search and data features: hybrid search, metadata filtering, multi-tenancy, backups, built-in embedding or reranking and bulk import from object storage, and on which plan each one sits.
• Compliance and deployment: SOC 2 Type II, ISO 27001, a HIPAA agreement, an EU region, deployment in your own cloud, customer-managed keys and private networking.
• Access control and exit: single sign-on, roles or scoped API keys, an audit log, a documented full export and an open-source engine you can run yourself.

Query speed, recall and uptime depend on the workload and how it is set up, so we do not score them. Each review says which pages it was written from and when. Where we have run a product hands-on, the evidence is on its page.