A vector database stores embeddings (numeric representations of text, images or audio) and finds the nearest neighbours to a query vector, usually with an approximate index such as HNSW. It is the retrieval layer behind most retrieval-augmented generation (RAG), semantic search and recommendation features.
The market splits three ways: managed vector-first services (Pinecone, Zilliz Cloud, Weaviate Cloud, Qdrant Cloud, Chroma Cloud), open-source engines you can self-host (Qdrant, Weaviate, Milvus, Chroma, LanceDB), and vector search added to a database you may already run (pgvector for Postgres, Elasticsearch, OpenSearch, MongoDB Atlas). For many teams under a few million vectors, the last group avoids running another system at all.
Tools, licences and pricing checked against vendor and repository pages on November 8, 2026.
1. Pinecone
CommercialBest for: Teams that want managed vector search with no infrastructure work · Starter free; Builder $20/month; Standard $50/month minimum; Enterprise $500/month minimum
Pinecone is a fully managed vector database. Its pricing page lists a free Starter plan, a Builder plan at $20/month flat, Standard with a $50/month minimum usage and Enterprise with a $500/month minimum.
Strengths
- + No servers or index tuning to manage
- + Free tier for prototypes
Trade-offs
- − Proprietary and cloud-only
- − Usage-based bills need monitoring at scale
2. pgvector
PostgreSQLBest for: Apps already on Postgres that want vectors next to relational data · Free and open source; available on many managed Postgres services
pgvector adds vector similarity search to PostgreSQL with HNSW and IVFFlat indexes, L2, inner product, cosine, L1, Hamming and Jaccard distances, and vector, halfvec, bit and sparsevec types. Indexed vectors are limited to 2,000 dimensions for vector and 4,000 for halfvec.
Strengths
- + Joins, transactions and backups come from Postgres
- + No extra system to operate
Trade-offs
- − Index dimension limits (2,000 for vector)
- − Very large or high-QPS workloads need careful tuning
3. Qdrant
Apache-2.0Best for: Self-hosted or managed vector search with rich filtering · Free 1 GB cluster; Standard usage-based; Premium minimum spend; Hybrid/Private Cloud by quote
Qdrant is an open-source vector search engine written in Rust with payload filtering. Qdrant Cloud has a free-forever single-node cluster (0.5 vCPU, 1 GB RAM, 4 GB disk), usage-based Standard and minimum-spend Premium tiers, plus Hybrid and Private Cloud by quote.
Strengths
- + Permissive licence and a single binary to self-host
- + Free managed tier
Trade-offs
- − Another stateful service to run if self-hosted
- − Advanced cloud tiers are quote-based
4. Weaviate
BSD-3-ClauseBest for: Hybrid search and built-in vectorization modules · Free plan; Flex from $45/month; Premium from $400/month
Weaviate is an open-source, cloud-native vector database written in Go that stores objects and vectors and combines vector and BM25 keyword search in one query. Weaviate Cloud lists a free plan, Flex from $45/month and Premium from $400/month.
Strengths
- + Hybrid BM25 + vector search out of the box
- + Permissive BSD licence
Trade-offs
- − More concepts (collections, modules) to learn than a plain index
- − Memory-heavy at large scale without compression
5. Milvus
Apache-2.0Best for: Very large collections that need horizontal scaling · Open source; Zilliz Cloud free tier, serverless from $0, dedicated from $126/month
Milvus is a cloud-native vector database for large-scale approximate nearest neighbour search. It runs as Milvus Lite (pip-installed, single file), Standalone or Kubernetes-native Distributed, and Zilliz Cloud offers a managed version with a free tier, serverless from $0/month and dedicated clusters from $126/month.
Strengths
- + Scales out to billions of vectors in distributed mode
- + Same API from laptop (Lite) to cluster
Trade-offs
- − Distributed mode has many components to operate
- − Overkill for small datasets
6. Chroma
Apache-2.0Best for: Prototyping RAG apps quickly in Python or JavaScript · Open source; Cloud Starter $0 + usage; Team $250/month + usage; Enterprise custom
Chroma is an open-source retrieval database with Python and JavaScript clients (pip install chromadb, npm install chromadb). Chroma Cloud lists Starter at $0/month plus usage with $5 in credits, Team at $250/month plus usage, and Enterprise custom pricing.
Strengths
- + Very simple developer API
- + Runs in-process for local development
Trade-offs
- − Fewer enterprise features than older engines
- − Cloud pricing is usage-based on top of plan fees
7. LanceDB
Apache-2.0Best for: Embedded vector search and multimodal data on local or object storage · Open source; cloud/enterprise: see vendor pricing
LanceDB is an embedded, open-source retrieval library for multimodal AI built on the Lance columnar format, so it runs inside your application without a server. A managed cloud and enterprise offering exists for production scale.
Strengths
- + No server to run in embedded mode
- + Columnar format suits large multimodal datasets
Trade-offs
- − Embedded model needs extra work for multi-writer services
- − Smaller ecosystem than Postgres or Elasticsearch
8. Elasticsearch
AGPL-3.0-only OR SSPL-1.0 OR Elastic-2.0Best for: Teams already running Elasticsearch for keyword search · Self-managed free tier; Elastic Cloud: see vendor pricing
Elasticsearch supports dense_vector fields with exact and approximate kNN, the managed semantic_text field with automatic embeddings, BBQ/int8/int4 quantization, and hybrid search via retrievers (RRF or linear) or ES|QL FORK and FUSE. Source is triple-licensed under AGPL-3.0, SSPL and the Elastic License 2.0.
Strengths
- + Mature keyword + vector hybrid search
- + Quantization options cut memory use
Trade-offs
- − Heavy to operate for vector-only use
- − Licensing is more complex than permissive alternatives
9. OpenSearch
Apache-2.0Best for: Apache-licensed search clusters that need vector and keyword search · Free and open source; managed services priced by the provider
OpenSearch, governed by the OpenSearch Foundation under the Linux Foundation, offers a k-NN vector field and query, neural and neural-sparse queries and hybrid search alongside full-text search, under the Apache 2.0 licence.
Strengths
- + Permissive licence with foundation governance
- + Combines logs, full-text and vector search
Trade-offs
- − Cluster operations are non-trivial
- − Vector features spread across plugins and query types
10. MongoDB Atlas Vector Search
CommercialBest for: Apps whose operational data already lives in MongoDB · See vendor pricing (part of MongoDB Atlas)
MongoDB Vector Search indexes embeddings alongside documents with ANN (HNSW) and exact (ENN) search, pre-filtering on indexed fields, hybrid search with full-text, and embeddings up to 8,192 dimensions. It runs on Atlas, on self-managed deployments and on local Atlas CLI deployments, and can auto-generate embeddings with Voyage AI models.
Strengths
- + Vectors and documents in one query and one database
- + Optional automatic embedding generation
Trade-offs
- − Best experience is on Atlas
- − Dedicated search nodes recommended for workload isolation add cost
