What should I use for embeddings if I already run PostgreSQL?

If you already run PostgreSQL, the simplest and most practical choice is usually:

Use pgvector inside PostgreSQL.

Why this is a good default

• Keeps everything in one system: data, metadata, and embeddings together.
• Easier ops: no extra vector database to run or sync.
• Good enough for many apps: semantic search, RAG, recommendation, deduping, similarity matching.
• Works well with SQL: filter by user, tenant, time, tags, permissions, then do vector search.

Typical setup

1. Store your embedding as a vector column.
2. Index it with an ANN index:
   • HNSW for better query performance / general use
   • IVFFlat if you want a simpler/lighter index and can tune it
3. Combine vector similarity with normal SQL filters.

When PostgreSQL + pgvector is a strong fit

• You have up to moderate scale.
• You want transactional consistency.
• Your application already depends heavily on PostgreSQL.
• You need hybrid search: text filters + embeddings.
• You want to avoid operational complexity.

When to consider a dedicated vector DB instead

If you need:

• Very large scale embeddings
• Very low-latency similarity search at high QPS
• Advanced vector-specific features
• Heavy write rates + large ANN workloads
• Multi-node vector search at significant scale

Then consider:

• Pinecone
• Weaviate
• Qdrant
• Milvus
• OpenSearch/Elasticsearch if you also need text search

Practical recommendation

• Start with PostgreSQL + pgvector.
• Only move to a dedicated vector store if you hit clear performance or scale limits.

If you want, I can also give you:

1. a recommended pgvector schema,
2. index choice guidance (HNSW vs IVFFlat), or
3. a decision matrix comparing pgvector vs vector DBs.