The Vector Database Index
A neutral index of the vector databases teams actually shortlist for RAG and semantic search in 2026 — grouped by what kind of store they are, how they're hosted, and what each is genuinely best at. A vector database stores embeddings and does similarity (nearest-neighbour) search; it's the memory layer underneath most AI agents. We describe each option's type, hosting, and license rather than quoting prices or version-specific benchmarks, because those move fast and the database is rarely the dominant cost. Pair this with the agent-frameworks index — the framework orchestrates, the vector DB remembers.
Last reviewed June 13, 2026 · 12 databases · neutral & vendor-independent
The matrix
| Database | Type | Hosting | License | Best for |
|---|---|---|---|---|
| Pinecone | Dedicated vector DB | Managed | Proprietary | Zero-ops, fully-managed production vector search |
| Weaviate | Dedicated vector DB | Both | Open-source | Open-source vector search with built-in hybrid search and modules |
| Qdrant | Dedicated vector DB | Both | Open-source | High-performance, Rust-based search with heavy metadata filtering |
| Chroma | Embedded | Both | Open-source | Lightweight, developer-friendly prototyping and small RAG apps |
| Milvus | Dedicated vector DB | Both (Zilliz managed) | Open-source | Billion-scale vector search |
| pgvector | Database extension | Both (any Postgres) | Open-source | Adding vectors to an existing Postgres stack without a new database |
| Redis | Document/cache DB + vectors | Both | Source-available | Low-latency vector search alongside caching/session data |
| Elasticsearch / OpenSearch | Search engine + vectors | Both | Open-source (OpenSearch) / mixed | Combining full-text and vector (hybrid) search at scale |
| MongoDB Atlas Vector Search | Document/cache DB + vectors | Managed | Proprietary (Atlas) | Vectors stored next to your operational documents |
| Azure AI Search | Search engine + vectors | Managed | Proprietary | Vector and hybrid search inside the Azure ecosystem |
| Vertex AI Vector Search | Dedicated vector DB | Managed | Proprietary | Google Cloud-scale ANN search (formerly Matching Engine) |
| LanceDB | Embedded | Both | Open-source | Embedded, multimodal, on-disk vector storage |
Store type, hosting, and license — not version-specific benchmarks, which go stale fast and depend on your data and query shape. Benchmark the shortlist on your own workload before committing.
How to read the types
- Dedicated vector DB
- Purpose-built for storing embeddings and running approximate-nearest-neighbour (ANN) search — the most features for pure vector workloads.
- Database extension
- Adds vector columns and indexes to a general database you already run (e.g. Postgres) — fewer moving parts.
- Search engine + vectors
- A full-text search engine that also does vector and hybrid (keyword + semantic) search at scale.
- Embedded
- Runs in-process / on-disk inside your application — no separate server to operate.
- Document/cache DB + vectors
- An operational database (document or in-memory) that adds vector search next to your live data.
Which one should you pick?
- You want zero-ops, fully managed vector searchPinecone
- You already run Postgres and want one databasepgvector
- You want open-source with hybrid searchWeaviate or Qdrant
- You're prototyping a RAG app fastChroma or LanceDB
- You need billion-scale ANNMilvus
- You want vectors next to operational dataMongoDB Atlas or Redis
- You need full-text + vector at scaleElasticsearch / OpenSearch
- You're committed to Azure or Google CloudAzure AI Search / Vertex AI Vector Search
FAQs
What is a vector database and do I actually need one?
A vector database stores embeddings — numerical representations of text, images, or other data — and finds the most similar ones via approximate-nearest-neighbour (ANN) search. It's what powers retrieval-augmented generation (RAG) and semantic search. You don't always need a dedicated one: for many apps, the pgvector extension on a Postgres you already run, or your existing search engine, is enough. Reach for a dedicated vector DB when you need large scale, advanced filtering, or specialized index types.
Which vector database is the best?
There's no single winner — it depends on scale and stack. For zero-ops managed search, Pinecone; for open-source with hybrid search, Weaviate or Qdrant; if you're already on Postgres, pgvector; for fast prototyping, Chroma or LanceDB; for billion-scale, Milvus; for vectors next to operational data, MongoDB Atlas or Redis. Cloud-committed teams often default to Azure AI Search or Vertex AI Vector Search.
Do I need a dedicated vector database, or can I just use Postgres?
For a large share of real applications, pgvector on Postgres is enough and keeps your stack to one database — fewer moving parts, one backup story, familiar tooling. Graduate to a dedicated vector database when you hit very large vector counts, need sophisticated metadata filtering, want specialized ANN index types, or need to decouple vector workloads from your transactional database.
How do vector databases relate to AI agents and frameworks?
They're complementary layers of the same stack. An agent that uses RAG retrieves relevant context from a vector database before answering; the agent framework (LangChain, LlamaIndex, and others) orchestrates the reasoning and tool calls, while the vector database stores and retrieves the embeddings. See the AI Agent Frameworks Index for the orchestration layer that sits on top of these stores.
Are vector databases free, and how current is this index?
Many are open-source: free to self-host, with managed tiers that bill on usage. The database is rarely the dominant cost — embedding generation and model API calls usually are. The ecosystem evolves quickly, so we describe each option's type, hosting, and license rather than version-specific benchmarks; a last-reviewed date is shown at the top, and you should verify each project's current docs before committing.