NoSQLAdvanced
Dedicated vector database or Postgres + pgvector?
“How do you decide where embeddings live?”
What this tests
- Vector storage tradeoffs
- Simplest-thing-first
Answers by level
Read the beginner answer first and notice what is missing.
Start with Postgres + pgvector: the vectors sit next to the rows they describe, filters are SQL, one transaction covers both, one system to run. Good to tens of millions of vectors. Move to a dedicated store when you have hundreds of millions to billions, vector search is the core workload, and you have measured that pgvector cannot keep up.
A dedicated store is a second system with its own consistency lag from the source of truth.
Green flags · Red flags
Strong green flag · Treats the dedicated store as a consistency boundary to justify.
Green flags
- pgvector first
- Moves on a measured number
- Filtered-ANN awareness
Red flags
- Dedicated store by default
- Unaware of the filter/recall interaction
Follow-up questions
F1
What breaks when you add a metadata filter to an ANN query?
Scenario
A RAG feature over 500k chunks is being built. The team wants Pinecone. What do you suggest and why?