AI apps need two things from a database: store traditional relational data (users, sessions, events) and store vector embeddings for similarity search (RAG, semantic search, recommendation). The 2026 reality: Postgres with pgvector handles both for most apps. Dedicated vector databases (Pinecone, Qdrant, Weaviate) are right answers at scale but overkill for the typical sub-1M-vector application.
The decision tree
| Vector count |
Query QPS |
Pick |
| <100k vectors |
<10 QPS |
Postgres + pgvector (Supabase, Neon) |
| 100k–10M |
10–100 QPS |
Postgres + pgvector still, with proper indexing |
| 10M+ |
100+ QPS |
Qdrant or Pinecone |
| Embedded / desktop |
Any |
SQLite + sqlite-vec |
| Edge / serverless |
Any |
Cloudflare Vectorize |
Best default — Postgres + pgvector
The default answer for almost everything. Supabase free tier ($0) gives you Postgres with pgvector pre-installed; scales to several million vectors before you need a dedicated solution. Neon also includes pgvector.
Why this beats dedicated vector DBs for most apps:
- One database to manage instead of two
- Joins between vectors + relational data are first-class
- Mature tooling, ORMs, backups
- Familiar SQL
When you need to upgrade: when query latency at >10M vectors becomes your bottleneck, or you need 100+ QPS with strict <100ms p95 latency.
When dedicated vector DBs make sense
Qdrant — open-source, self-hostable, serverless cloud option. Strong filtering performance.
Pinecone — managed only, fastest serverless cold start, good for spiky workloads.
Weaviate — full-featured, hybrid search (vector + keyword) built-in.
For most AI apps, you'll never need any of these.
What's NOT worth your money
- Dedicated vector DBs for sub-1M vector apps when pgvector handles it
- Multiple vector DBs for "redundancy" — pick one and trust it
- Premium "AI database" SaaS charging $200+/mo for what Supabase does free
- Custom embedding storage in plain JSON files when SQLite + sqlite-vec exists
FAQ
Should I use Postgres or a dedicated vector DB for a new RAG app?
Postgres + pgvector almost always. Migrate later only if vector queries become a real bottleneck.
What about MongoDB Atlas Vector Search?
Works fine if you're already on MongoDB. For new projects: Postgres + pgvector is simpler.
Best embedding model to use with these DBs?
text-embedding-3-small (OpenAI, $0.02/M tokens) for general use. Voyage-3 or Cohere embed-english-v4 for higher quality. We covered API economics in How to use AI APIs without going broke 2026.
Can I use SQLite for production RAG?
For desktop apps, embedded scenarios, low-traffic web apps: yes, sqlite-vec extension works well. For multi-user web apps with concurrent writes: Postgres is safer.
What's Cloudflare Vectorize?
Cloudflare's edge-deployed vector DB, integrates with their Workers + AI products. Worth considering for edge-first apps.
Best free option overall?
Supabase (Postgres + pgvector + auth + storage, free tier covers a lot) or Neon (Postgres only, generous free tier).
Related reading