I'm planning a recommendation engine with millions of user and item embeddings; what vector store fits best?

For a recommendation engine with millions of user and item embeddings, the “best” vector store depends mainly on your latency, freshness, filtering, and ops requirements. In practice, there are a few strong choices:

Best general-purpose choices

1) Milvus

Best if you want:

• Large scale (millions to billions of vectors)
• Good ANN performance
• Hybrid search with metadata filters
• Self-hosted control and cost efficiency

Why it fits:

• Designed for high-dimensional similarity search at scale
• Supports multiple index types
• Good if you expect to grow a lot and need tuning flexibility

2) Pinecone

Best if you want:

• Managed service
• Fast time-to-production
• Low operational overhead
• Simple scaling and good reliability

Why it fits:

• Easier to operate than self-hosted systems
• Great if your team wants to focus on product rather than infrastructure
• Often a strong choice for production recommender systems if budget is less constrained

3) Qdrant

Best if you want:

• Strong metadata filtering
• Simple deployment
• Good performance at moderate-to-large scale
• Self-hosted or managed options

Why it fits:

• Very practical for recommendation use cases
• Good balance of speed, filtering, and usability
• Often easier to manage than more complex distributed systems

Other options

4) Weaviate

Good if you want:

• Vector search plus richer schema/search features
• Easier developer experience
• Hybrid keyword + vector search

5) FAISS

Good if you want:

• Maximum raw performance in a custom setup
• You’re okay building your own service around it

But:

• It’s a library, not a full vector database
• You’ll need to handle persistence, scaling, filtering, and replication yourself

---

My practical recommendation

If you want the safest default:

• Pinecone if you want managed simplicity
• Milvus if you want self-hosted scale and cost control
• Qdrant if you want a simpler self-hosted system with strong filtering

---

For recommendation engines specifically

Important features to prioritize:

• Fast top-K ANN search
• Metadata filtering like region, category, price tier, or availability
• High ingest/update throughput if embeddings change often
• Low-latency retrieval for online serving
• Hybrid ranking pipeline: vector store for candidate generation, separate ranker for final scoring

A common architecture is:

1. Use vector store to retrieve top 100–1000 candidates
2. Use a ranking model to reorder them
3. Apply business rules/filters afterward

---

Quick pick by scenario

• Startup / easiest ops → Pinecone
• Self-hosted / best value at scale → Milvus
• Need strong filtering and clean API → Qdrant
• Custom in-house system → FAISS + your own service layer

If you want, I can also give you a decision matrix comparing Milvus vs Pinecone vs Qdrant for your specific constraints like QPS, latency, filtering, and budget.