Summary
A vector database stores embeddings and returns nearest neighbors by similarity. The category is crowded: Pinecone, Weaviate, Qdrant, Milvus, Chroma, pgvector. The choice matters less than the discipline of using one as a database — with backups, migrations, and access control.
Objectives
- 01Name the four axes on which vector databases differ.
- 02Recognize when pgvector suffices.
- 03State one reason to pick each of the leading options.
The Lesson
The axes
1. Managed vs. self-hosted. 2. Standalone vs. embedded (in Postgres). 3. Index algorithm (HNSW, IVF, DiskANN). 4. Hybrid search support (BM25 alongside vectors). Every serious DB supports HNSW; the differences are operational.
Choosing
pgvector if you already run Postgres and have <10M vectors. Qdrant for self-hosted with strong metadata filtering. Pinecone for zero-ops managed. Weaviate for hybrid search and modular design. Milvus for scale beyond 100M vectors.
Key Ideas
- Vector DBs mostly do the same thing; operations differ.
- Start with pgvector; move only when it hurts.