Guide

The best open source vector databases in 2026, and how we chose

Six open source options for storing and searching embeddings, compared on licence, index types, filtering, hybrid search and how you run them, from each project's own pages. We say which one we run and where that biases us.

  • Checked 17 Nov 2026
  • OSStend runs pgvector, Qdrant and OpenSearch

The best open source vector database depends on where your data already lives. Checked 17 Nov 2026: pgvector keeps vectors in PostgreSQL, OpenSearch adds vectors to keyword search, and Qdrant, Milvus, Weaviate and Chroma are dedicated engines. Four of the six use the Apache 2.0 licence.

Which open source vector databases are worth shortlisting?

Six, grouped by how they run. pgvector and OpenSearch add vector search to a database you may already run; Qdrant, Milvus, Weaviate and Chroma are engines built for vectors. None is best everywhere, and the table says what each page states.

Illustrative: similar documents sit close together, so retrieval is a nearest-neighbour search.
OptionWhat it isLicenceIndex types and search the docs stateLatest release
pgvectorA PostgreSQL extension; vectors sit beside relational dataPostgreSQL licence (LICENSE)HNSW and IVFFlat; exact and approximate search; L2, inner product, cosine, L1, Hamming and Jaccard distances; filtering with iterative index scans from 0.8.0 (README)Tag v0.8.7
QdrantA vector search engine written in Rust, with a payload attached to each vectorApache 2.0 (GitHub)Filterable HNSW with payload indexes, quantization, and a Query API for hybrid and multi-stage queries (indexing, hybrid)v1.20.2, 16 Nov 2026
MilvusA distributed vector database that separates compute from storageApache 2.0 (GitHub)HNSW, IVF, FLAT, SCANN and DiskANN, metadata filtering, BM25 full-text and sparse vectors for hybrid search; Milvus Lite for local use (README)v3.0.6, 11 Nov 2026
WeaviateA vector database that can vectorize data on importBSD-3-Clause outside its "wl" directory; that directory is under a separate enterprise licence (LICENSE)HNSW, flat and dynamic indexes, plus a cluster-based HFresh index; hybrid search with BM25 (vector index docs, README)v1.40.5, 10 Nov 2026
OpenSearchA search and analytics suite with k-NN vector searchApache 2.0 (GitHub)Approximate and exact k-NN, neural sparse search, and hybrid search through a search pipeline (vector search techniques)3.9.0, 29 Sep 2026
ChromaAn embedded or client-server vector store with a small APIApache 2.0 (GitHub)Runs in memory or persisted in your process, or as a server your clients connect to (client-server docs)1.5.9, 5 May 2026

How we compared: each project's README, licence file, documentation and GitHub releases page, read on 17 Nov 2026, with the link in each cell. Chosen because each is open source under a licence the project states, can be self-hosted, has a vector index in the project itself and published a release in the last year. Vespa, LanceDB and Valkey Search also meet those tests and are not covered here. We ran no benchmark and quote none: vendor benchmarks differ in hardware, index settings and recall targets, so test with your own vectors. OSStend runs PostgreSQL with pgvector, Qdrant and OpenSearch and does not run Milvus, Weaviate or Chroma, so weigh our views with that interest in mind.

Which open source vector database fits which job?

Match the option to where the data lives and what the query needs. Data already in PostgreSQL points to pgvector, keyword plus vector search points to OpenSearch, and a vector-only workload at large scale points to a dedicated engine.

If you needLook atBecause the project's page says
Vectors next to the rows they describe, with joins and transactionsIt stores vectors with the rest of your data and keeps ACID compliance and point-in-time recovery (README)
Keyword and vector results in one queryOpenSearch, Weaviate, Qdrant or MilvusHybrid search is documented by each: OpenSearch, Weaviate, Qdrant, Milvus
Heavy payload filtering alongside vector searchQdrantIt is described as tailored for extended filtering, with payload indexes for filter fields (README, indexing)
A cluster that scales compute and storage separatelyMilvusIt describes a distributed architecture that separates compute and storage (README)
A quick prototype on a laptop, then a serverChromaThe same client runs in process or against a server (docs)
Several index types and built-in vectorization on importWeaviateFlat, HNSW and dynamic indexes and integrated model providers (docs, README)

Sources linked in each row, checked 17 Nov 2026. A reading of project pages, not a benchmark.

From our on-call engineers: pgvector settings that decide recall and build time

With pgvector the defaults are safe, not fast, and filters can silently reduce results. These are the settings and checks we apply to any HNSW index before calling a vector search production-ready.

  • HNSW builds are much faster when the graph fits in maintenance_work_mem, and Postgres logs a notice when it no longer does, so we raise it for the build only and watch server memory.
  • max_parallel_maintenance_workers defaults to two workers plus the leader, and raising it speeds up large builds if max_parallel_workers allows.
  • hnsw.ef_search defaults to 40, and higher values improve recall at the cost of speed, so we measure recall against exact search on a sample before choosing.
  • With approximate indexes the WHERE filter runs after the index scan, so a selective filter can return fewer rows than requested; pgvector 0.8.0 added hnsw.iterative_scan (strict_order or relaxed_order) to fix this.
  • Vacuuming HNSW indexes can be slow, and pgvector's README suggests REINDEX INDEX CONCURRENTLY before VACUUM; we also check index size with pg_relation_size, since performance is usually better when the index fits in memory.

What limits does each option state?

Each has a catch. pgvector indexes vectors up to 2,000 dimensions, Weaviate carries an enterprise licence for part of its code, and OpenSearch is a search engine to size and run, not just an index.

pgvector.
An index on the vector type covers up to 2,000 dimensions, halfvec up to 4,000, with filtering applied after the index scan unless iterative scans are on (README, checked 17 Nov 2026). Embedding models commonly produce 384 to 1,024 dimensions per Ollama's embeddings docs.
Weaviate.
Code in its "wl" directory is available only under a separate enterprise licence, enabled by a key (LICENSE, checked 17 Nov 2026).
Qdrant.
Payload indexes should be created before data is loaded to get the filter-aware graph, and they cost memory and disk (indexing docs, checked 17 Nov 2026).
Milvus.
A distributed, Kubernetes-native design is more to run than a single node; Milvus Lite and a standalone mode exist for small use (README, checked 17 Nov 2026).
Chroma.
Its documentation describes embedded and client-server modes; check the docs for how it scales before you plan for a large cluster (docs, checked 17 Nov 2026).
OpenSearch.
It lists approximate search as the best option for most uses, with exact search for small or filtered sets (docs, checked 17 Nov 2026).

Where OSStend fits, and where it does not

OSStend runs PostgreSQL with pgvector, a dedicated Qdrant and OpenSearch, so it is an option for three of the six. It does not run Milvus, Weaviate or Chroma, and for a vector-only workload at large scale a dedicated engine can be the better tool.

Contact us if

  • The vectors belong beside relational data, and you want the database designed and run 24/7: managed PostgreSQL and pgvector.
  • You need keyword and vector search in one engine: managed OpenSearch.
  • The workload is vector-first at scale and you want Qdrant run for you: Qdrant Cloud alternatives.
  • The corpus must stay in your cloud account or data center.
Talk to an engineer

Not a fit if

  • You want Milvus, Weaviate or Chroma run for you: use the project's own install or its cloud.
  • You are deciding between pgvector and Qdrant: see pgvector vs Qdrant.
  • You are prototyping: a library or a free plan is quicker.

For the retrieval pattern behind these choices, see vector database for RAG and the vector database and RAG guide. For running the whole stack next to a private model, see self-hosted AI.

FAQ

What is the best open source vector database?

There is no single best one. pgvector fits data already in PostgreSQL, OpenSearch fits keyword plus vector search, and Qdrant, Milvus, Weaviate and Chroma are dedicated engines. The table compares them from each project's own pages, checked 17 Nov 2026.

Is pgvector a real vector database?

It is a PostgreSQL extension that adds vector types and HNSW and IVFFlat indexes, with exact and approximate search (README, checked 17 Nov 2026). It gives you vector search inside a relational database rather than a separate engine.

Which open source vector database is truly free?

Qdrant, Milvus, OpenSearch and Chroma state Apache 2.0, pgvector uses the PostgreSQL licence, and Weaviate is BSD-3-Clause apart from an enterprise-licensed directory (checked 17 Nov 2026). Running any of them still costs infrastructure and operations time.

Which vector database is best for RAG?

The one that holds your corpus and filters where you already work. See vector database for RAG for how pgvector and OpenSearch compare with dedicated engines.

Does OSStend run Qdrant, Milvus or Weaviate?

Qdrant, yes: OSStend runs a dedicated Qdrant in your environment (see Qdrant Cloud alternatives), alongside PostgreSQL with pgvector and OpenSearch. Milvus and Weaviate, no: use the project's own install or its cloud.

Talk to an engineer

Take the vector decision to an engineer

45 minutes: your corpus, filters and scale are reviewed and an engineer says where the vectors should live.

  • A design sketch for your workload, not a plan menu
  • A zero-downtime cutover path for what you run today
  • A price for that design, after the review
See pricing →

Your request goes to the OSStend team. No newsletter sign-up.

Get a quoteTalk to an engineer
2Engines
3Where should it run?
4What do you need?

Your request goes to the OSStend team. No newsletter sign-up.