Vector databases in 2026: when to pick which

Quick answer. In 2026, the right vector database depends on three things: (1) scale (single-machine vs. distributed), (2) filtering (do you need to filter on metadata during search?), and (3) stack alignment (do you already have Postgres?). The 90% answer: Qdrant for most new RAG stacks, pgvector when you already have Postgres, Chroma for prototypes. Pinecone is the right pick only when you want a managed service and don't want to operate the database.

The 2026 vector database landscape

The vector database market has consolidated in 2025-2026. Five open-source projects and one managed service cover 90% of the use cases. The GitPop radar tracks all of them; here is the current state:

Project	PopScore (W34)	Best for
Qdrant	67	Most new RAG stacks
Milvus	58	Billion-scale, distributed
Weaviate	54	Modular, with built-in vectorization
Chroma	49	Prototypes, small apps
pgvector	45	Existing Postgres shops
Pinecone	(managed, no GitHub score)	"Don't want to operate a DB"

The decision framework

If you are starting a new RAG stack

Pick Qdrant. Rust-native, single-binary deploy, strong filtering engine, hybrid search built in, and the most momentum on the GitPop radar in 2026 H2. The API is clean, the documentation is good, and the operational overhead is the lowest of the standalone options.

For a deeper look at Qdrant, see Best open source AI tools in 2026 where we cover the AI toolchain end-to-end.

If you already have Postgres

Pick pgvector. Postgres + the pgvector extension is the right pick when you have an existing Postgres deployment and the vector workload is moderate (under ~10M vectors). The advantages:

• No new infrastructure to operate
• ACID transactions across your vector and relational data
• Standard SQL tooling for backup, monitoring, etc.

The disadvantages: at scale (100M+ vectors), pgvector starts to lose to Qdrant and Milvus on raw throughput. For the 90% of teams that don't have a billion vectors to search, pgvector is the right pick.

If you need billion-scale distributed search

Pick Milvus. Distributed architecture, mature, the right pick for very large embedding workloads (search engines, recommendation systems, ad ranking). The operational complexity is higher than Qdrant, but the scale ceiling is much higher.

For the 99% of teams that don't have a billion vectors, Milvus is overkill. But for the 1% that do, Milvus is the right pick.

If you want a modular, "vector DB + plugins" architecture

Pick Weaviate. Weaviate's differentiator is the modular vectorization layer — you can plug in OpenAI, Cohere, Hugging Face, or local models as the embedding backend, all behind a unified API. The right pick when your embedding model is going to change frequently and you don't want to rewrite code every time.

If you are prototyping or building a small app

Pick Chroma. Python-first, single pip install, the lowest-friction option for prototypes and small apps. The right pick when you are still figuring out the schema and don't want to commit to an operational choice.

For production, plan to migrate to Qdrant or pgvector before you ship. Chroma is great for development, and the migration path to a production vector store is well-trodden.

If you don't want to operate a database at all

Pick Pinecone. The managed vector database. The right pick when you want to skip the operational choice entirely. The tradeoffs:

• You don't operate the database
• The bill scales with usage
• You are locked into Pinecone's API surface

In 2026, the cost calculus has shifted. For most teams, the operational overhead of Qdrant is 2-4 hours a month — much less than the Pinecone bill at scale. But if you have a small workload and a tight team, Pinecone is a legitimate choice.

Common mistakes

1. Picking a vector database before you know your data shape. The right pick depends on the number of vectors, the dimensionality, the filter complexity, and the query latency. Don't commit to a vector database in week 1 of the project. Prototype with Chroma or pgvector, then migrate.
2. Picking the wrong embedding model. The embedding model matters more than the vector database. In 2026, the right pick for most teams is text-embedding-3-small (OpenAI), voyage-3 (Voyage), or bge-large-en-v1.5 (local, open-source). See our State of GitPop H2 2026 data report for the broader AI toolchain.
3. Skipping hybrid search. Pure vector search misses exact-match requirements (e.g., "search for project ID 12345"). The right vector database in 2026 has hybrid search built in: vector similarity + metadata filtering. Qdrant, Weaviate, and Milvus all ship this natively.

How to migrate

If you start with Chroma or pgvector and outgrow them, the migration path is well-trodden. Most teams use LangChain or LlamaIndex as the abstraction layer, so the migration is a config change, not a code rewrite. See our LangChain vs LlamaIndex comparison for the broader tooling decision.

For the broader AI toolchain, see Best open source AI tools in 2026. For the methodology behind the data in this post, see What is PopScore?.