Which Vector Database Should You Use?
For most new RAG projects: start with pgvector (free, inside your existing Postgres) or Chroma (free, local dev). At 1M+ vectors or production scale, move to Pinecone Serverless or Qdrant Cloud. Weaviate is best when you need hybrid search (keyword + vector) or multi-tenancy.
How many vectors do you expect to store?
FAQ
Is pgvector production-ready?+
Yes, for most use cases. pgvector with HNSW indexes handles millions of vectors in production at companies like Supabase, Vercel, and many YC startups. The main limitations are at extreme scale (100M+ vectors) where dedicated vector databases have better memory efficiency. For under 5M vectors on a properly sized Postgres instance, pgvector is an excellent production choice.
What is the difference between HNSW and IVF-Flat indexes?+
HNSW (Hierarchical Navigable Small World) trades memory for query speed — it's 2–5x faster at query time but uses significantly more RAM. IVF-Flat uses less memory but requires a separate quantization step and has slightly lower recall. For most production RAG workloads, HNSW with default settings is the right choice. Use IVF-PQ when you need to store 100M+ vectors on a memory-constrained server.
How much does it cost to store 1M vectors?+
Approximately: Pinecone Serverless ~$0/mo base + $0.04/1M queries; Qdrant Cloud from ~$25/mo for 1M 1536-dim vectors; Weaviate Cloud from ~$25/mo; self-hosted Qdrant on a 16GB RAM VM ~$50–100/mo. At 1M vectors, managed services are almost always cheaper than self-hosted when you factor in engineering time.
Should I use the same embedding model as my vector database's native embedder?+
Not necessarily. Use the embedding model that performs best on your domain, regardless of your vector database. All major vector databases accept pre-computed embeddings from any model. OpenAI text-embedding-3-large (3072 dims) and Cohere embed-v3 lead the MTEB leaderboard in 2026. Match embedding dimensions between your indexing pipeline and query pipeline — they must be identical.