AppHost logoAppHost

Devii · AI & ML · 2026-05-27 · 8 min read

Share

Vector Databases: Embeddings, ANN Indexes, And Retrieval Workloads

What dedicated vector stores optimize, and when Postgres or OpenSearch is enough.

**Vector databases** optimize approximate nearest neighbor (**ANN**) search over high-dimensional embeddings. Products include Pinecone, Weaviate, Qdrant, Milvus, and pgvector-backed Postgres.

Indexes like **HNSW** and **IVF** trade memory, build time, and recall. Metadata filters (tenant ID, date range) should apply before or during graph traversal depending on engine.

RAG pipeline design
RAG pipeline design

Dedicated stores help when QPS, billion-scale corpora, or multi-tenant isolation exceed comfortable Postgres limits. Smaller RAG products often start on pgvector with a migration path documented upfront.

Version embedding models explicitly; changing model dimensions invalidates indexes. Evaluate hybrid retrieval (BM25 plus vectors) before scaling infrastructure.