AI Engineering

Vector Databases Explained: 15 Options and How to Choose

Pinecone, Weaviate, Milvus, Qdrant, Chroma, FAISS, pgvector — the landscape is wide. Here is what each is optimized for and how to pick for your use case.

3 min read
On this page

Summary

A vector database stores embeddings and finds the nearest ones to a query — the retrieval engine behind RAG. The market is crowded (fifteen-plus credible options), but they fall into a few clear categories: in-process libraries, Postgres extensions, self-hosted open-source engines, and fully managed cloud services. You do not need to evaluate all fifteen. Pick by where you are — a student prototyping, a team already on Postgres, or a production service at scale — and then benchmark the two or three candidates on the things that actually matter for your workload.

What vector databases do

Embeddings turn text, images, or other data into high-dimensional vectors where "similar things are close together." A vector database stores those vectors and answers one core question fast: which stored vectors are nearest to this query vector? Because exact nearest-neighbor search is too slow at scale, these systems use approximate nearest-neighbor (ANN) indexes to trade a sliver of accuracy for large speedups. This similarity search is what powers retrieval in a RAG pipeline.

The categories

The wide field organizes into a handful of types:

  • In-process library — FAISS. Meta's library (not a server) for finding the closest matches among dense vectors. Runs inside your app, zero infrastructure, free. Ideal for prototypes and learning.
  • Postgres extension — pgvector. Lets you store vectors in ordinary Postgres tables and query them by distance. No new service to operate if you already run Postgres.
  • Self-hosted open-source engines — Milvus, Qdrant, Weaviate, Chroma. Milvus is built for collections with billions of vectors; Qdrant is strong at combining vector search with metadata filters on each stored record; Weaviate mixes keyword and vector search and has plug-in modules; Chroma is small and easy to run on your own machine, so many people use it for first RAG experiments.
  • Managed cloud services — Pinecone, and managed tiers of the above. Pinecone runs the whole vector store for you as a hosted service, aimed at production apps that need quick lookups. You trade some control for someone else operating it.
  • Vector features bolted onto existing systems — Redis, Elasticsearch/OpenSearch, MongoDB Atlas, and others. If you already run one of these, its vector capability may be "good enough" without adding a new dependency.

A selection guide

  • Student or prototype: FAISS or Chroma — zero infra, free, fast to start.
  • Already on Postgres: pgvector — keep your data in one place, no new service to run.
  • Self-hosted production: Milvus or Weaviate (or Qdrant) — open-source engines built to scale.
  • Production on the cloud, minimal ops: Pinecone or a managed Qdrant tier — managed and scalable.

The cleanest decision rule: start from your existing stack and operational appetite, not from a benchmark leaderboard.

What to actually benchmark

Once you have two or three candidates, measure the things that decide real-world fit:

  • Query latency at your scale — performance on a toy dataset tells you nothing about millions of vectors.
  • Filtering support — can it filter by metadata (tenant, date, category) efficiently alongside the vector search?
  • Hybrid search — does it combine keyword and semantic search, which often beats either alone?
  • Cost — total cost of ownership, including operations for self-hosted options, not just sticker price.

Pick the simplest option that clears your bar on those four, and resist switching until a concrete limit forces you to.

Key takeaways

  • A vector database stores embeddings and runs approximate nearest-neighbor (ANN) search — the retrieval engine behind RAG.
  • The field sorts into categories: in-process library (FAISS), Postgres extension (pgvector), self-hosted engines (Milvus, Qdrant, Weaviate, Chroma), and managed cloud (Pinecone).
  • Choose by your situation: FAISS/Chroma for prototypes, pgvector if you are on Postgres, Milvus/Weaviate for self-hosted scale, Pinecone for managed cloud.
  • Benchmark what matters — query latency at your real scale, metadata filtering, hybrid search, and total cost — not leaderboard numbers.
  • Start from your existing stack and ops appetite; pick the simplest option that clears your bar and switch only when forced.

What to do next

Based on a LinkedIn infographic by Alok Sharan (2026). The ideas are theirs; the wording and the extra detail here are ours.