I'm building a recommendation engine with embeddings, any database suggestions?

Yes — for an embedding-based recommendation engine, the best database depends on your scale, latency needs, and whether you want a single system or a specialized vector database.

Good options

1. PostgreSQL + pgvector

Best if you want simplicity and already use Postgres.

Pros

• Easy to adopt if you already have relational data
• Supports metadata filtering alongside vector search
• Good for prototypes, MVPs, and moderate scale
• Operationally simpler than adding a new DB

Cons

• Not ideal for very large-scale, high-QPS vector search
• Fewer vector-specific optimizations than dedicated vector DBs

Use if

• You want one database for users/items/embeddings
• Your dataset is up to a few million vectors, depending on latency requirements

---

2. Pinecone

Best managed vector DB for production-grade similarity search.

Pros

• Fully managed
• Low-latency ANN search
• Easy scaling and indexing
• Good metadata filtering

Cons

• Can be expensive at scale
• Vendor lock-in
• Not a general-purpose relational DB

Use if

• You want minimal ops and strong vector-search performance

---

3. Weaviate

Good open-source / managed hybrid option.

Pros

• Vector search + metadata filtering
• Hybrid search support
• Flexible schema
• Can self-host or use managed

Cons

• More moving parts than Postgres
• Requires tuning/ops knowledge for self-hosting

Use if

• You want a vector-native system but also flexibility

---

4. Milvus

Strong choice for large-scale vector search.

Pros

• Designed for large vector workloads
• Scales well
• Open source
• Good ANN performance

Cons

• More operational complexity
• Usually paired with other systems for metadata/transactions

Use if

• You expect large catalogs and high throughput

---

5. OpenSearch / Elasticsearch

Useful if you want search + vectors together.

Pros

• Great for combining keyword and vector search
• Good filtering and ranking pipelines
• Familiar for search teams

Cons

• Vector search is not as specialized as dedicated vector DBs
• Can be heavy/costly

Use if

• Recommendations are part of a broader search/discovery platform

---

Quick recommendation by stage

• Prototype / MVP: PostgreSQL + pgvector
• Managed production with low ops: Pinecone
• Open-source vector-native: Weaviate or Milvus
• Search + recommendations together: OpenSearch/Elasticsearch

What matters most for your choice

1. Dataset size — number of items/vectors
2. Latency target — sub-50ms? sub-200ms?
3. Update frequency — static catalog vs frequent writes
4. Filtering needs — category, locale, price, availability
5. Budget / ops tolerance — managed vs self-hosted
6. Hybrid search — keyword + vector ranking

Practical architecture

For many recommendation systems:

• Store core app data in PostgreSQL
• Store embeddings in pgvector or a dedicated vector DB
• Add Redis for caching hot recommendations
• Use a feature store / event pipeline if you need real-time personalization

If you want, I can also suggest the best database for your specific case if you tell me:

• number of items
• number of users
• expected QPS
• whether you need filters
• cloud/on-prem preference