Which vector search setup should I use for a small production app?

For a small production app, the best vector search setup is usually:

Recommended default

PostgreSQL + pgvector

Why this is a good choice

• Simple ops: one database for both app data and vectors
• Production-ready: reliable, mature, easy to back up and monitor
• Good enough performance for small-to-medium scale
• Flexible: supports metadata filtering, joins, transactions
• Cost-effective: no extra vector database to run

Best if your app has:

• up to a few hundred thousand vectors, sometimes more depending on latency needs
• moderate query volume
• need for metadata filtering
• a small team or solo developer
• an existing PostgreSQL stack

---

When to use a dedicated vector database instead

Consider a vector DB like Qdrant, Pinecone, Weaviate, or Milvus if:

• you expect millions+ vectors
• you need low-latency high-QPS search
• you want hybrid search and advanced retrieval features
• you don’t want to manage tuning Postgres indexes for vector workloads

For a small production app, though, these are often more complexity than you need.

---

Practical setup I’d choose

If starting today:

• Database: PostgreSQL 15+ with pgvector
• Index: HNSW if your pgvector version supports it and you want fast approximate search
• Fallback: IVFFLAT if HNSW isn’t available or if your workload fits it better
• Metadata: store in same table and filter in SQL
• Embeddings: use one consistent model and normalize if using cosine similarity

---

Rule of thumb

• < 1M vectors and want simplicity → Postgres + pgvector
• > 1M–5M vectors or heavy traffic → consider Qdrant/Pinecone
• Need only semantic search in a simple app → pgvector almost always first choice

If you want, I can recommend a setup more precisely based on:

1. expected vector count,
2. query latency target,
3. whether you need metadata filters, and
4. your current stack.