Every RAG system needs somewhere to store embeddings. The market offers dedicated vector databases, search engines with vector features, and the quiet contender: PostgreSQL with pgvector.
Start with Postgres more often than not
For corpora under tens of millions of vectors, pgvector delivers excellent recall with the operational simplicity of a database your team already runs — plus joins with your relational data.
When dedicated stores earn their keep
Billion-scale collections, extreme QPS, or heavy metadata filtering at scale justify purpose-built engines. Benchmark with your real queries; vendor benchmarks flatter the vendor.
What matters more than the engine
Chunking strategy, embedding model choice, and hybrid search tuning move retrieval quality far more than switching databases. Datastrel designs and benchmarks retrieval stacks matched to your data and budget.