pgvector vs Milvus for startups: Which Should You Use?

pgvector is the “keep it in Postgres” choice: simple, familiar, and good enough for a lot of startup workloads. Milvus is the purpose-built vector database: more moving parts, but much better when retrieval becomes a core product feature.

For most startups, start with pgvector unless you already know you need high-scale ANN search, hybrid retrieval at volume, or multi-tenant vector infrastructure from day one.

Quick Comparison

When pgvector Wins

• •

  You already run Postgres as your source of truth

  This is the biggest reason to choose pgvector. If your app stores users, billing, documents, and embeddings in one database, you get fewer systems to manage and fewer failure modes.

  A typical schema looks like this:

  CREATE EXTENSION IF NOT EXISTS vector;
  
  CREATE TABLE documents (
    id bigserial PRIMARY KEY,
    tenant_id uuid NOT NULL,
    content text NOT NULL,
    embedding vector(1536)
  );
  
  CREATE INDEX ON documents USING hnsw (embedding vector_cosine_ops);
  

• •

  You need fast iteration with a small team

  Startups die from complexity more often than from lack of model quality. pgvector lets one backend engineer ship semantic search without introducing a separate distributed system.

  Querying is just SQL:

  SELECT id, content
  FROM documents
  WHERE tenant_id = $1
  ORDER BY embedding <=> $2
  LIMIT 10;
  

• •

  Your dataset is moderate

  If you’re dealing with tens of thousands to low millions of vectors per tenant or per app, pgvector is usually enough. With the right index choice — hnsw for low-latency search or ivfflat for cheaper approximate search — it performs well.

• •

  You need joins and transactional consistency

  This matters more than people admit. If vector search results must be filtered by business data — plan tier, region, permissions, document status — Postgres handles that cleanly in one query.

When Milvus Wins

• •

  Vector search is the product

  If your core feature depends on retrieval speed and recall at scale — not just “nice-to-have” RAG — Milvus is the right tool. It’s designed for ANN search first, not as an extension bolted onto a relational database.

• •

  You expect large-scale growth quickly

  Milvus handles bigger collections and heavier query loads better than pgvector once you move beyond comfortable Postgres territory. If you expect millions to hundreds of millions of vectors soon, don’t force Postgres to pretend it’s a vector platform.

• •

  You need dedicated retrieval infrastructure

  Milvus gives you vector-native primitives like collections, partitions, scalar filtering alongside similarity search, and index choices optimized for retrieval workloads such as HNSW and IVF-based indexing through its own API surface.

  A Python example looks like this:

  from pymilvus import Collection
  
  collection = Collection("documents")
  results = collection.search(
      data=[query_embedding],
      anns_field="embedding",
      param={"metric_type": "COSINE", "params": {"ef": 64}},
      limit=10,
      expr='tenant_id == "abc"'
  )
  

• •

  You’re building multi-tenant AI infrastructure

  If different customers will each have large corpora and strict isolation requirements around performance or scaling, Milvus gives you more room to grow without turning Postgres into an overloaded Swiss army knife.

For startups Specifically

Use pgvector first unless vector retrieval is your entire business model or you already know your workload will be huge in month one. It keeps architecture boring: one database for app data and embeddings, one backup strategy, one operational surface area.

Move to Milvus when pgvector starts fighting you on scale or latency — not before. Startups lose time by overengineering the retrieval layer too early; pgvector buys speed of delivery without boxing you into a dead end immediately.

Keep learning

• •The complete AI Agents Roadmap — my full 8-step breakdown
• •Free: The AI Agent Starter Kit — PDF checklist + starter code
• •Work with me — I build AI for banks and insurance companies

By Cyprian Aarons, AI Consultant at Topiax.