do i need a separate database for embeddings and metadata?

Not necessarily.

You have a few common options:

1) Store embeddings and metadata in the same database

This is often the simplest choice.

• Relational DBs like Postgres can store:
  • the embedding vector
  • metadata columns
  • the original text or document reference
• With the right extension/indexing, you can do vector search there too.

Good when:

• you want simpler architecture
• your scale is moderate
• metadata filtering is important
• you want one source of truth

2) Use a dedicated vector database + separate metadata store

Sometimes people keep:

• embeddings in a vector DB
• metadata in Postgres / MongoDB / etc.

Good when:

• you have very large scale
• vector search performance is the main concern
• metadata is complex or managed by another system
• you want to optimize each store for its job

3) Store everything in a vector database that also supports metadata

Many vector databases allow you to attach metadata fields directly to each vector.

Good when:

• you want fast similarity search
• metadata is mostly for filtering and display
• you don’t need advanced relational queries

---

Practical recommendation

If you’re just starting out, don’t use a separate database unless you have a reason.

A very common setup is:

• Postgres + pgvector for embeddings and metadata together

That gives you:

• one database
• vector search
• SQL filtering
• simpler maintenance

---

When to separate them

Consider a separate DB if:

• embeddings are extremely large in volume
• you need very high vector search throughput
• metadata needs frequent joins/transactions
• different teams own the systems
• you expect to migrate or reindex embeddings often

If you want, I can suggest the best setup for your specific use case (RAG app, semantic search, recommendation system, etc.).