Learning path · Embeddings & Representation · 37
Vector Databases
Storage engines optimized for approximate nearest-neighbor search over millions of embeddings with metadata filters.
Why it matters
- Operational home for RAG indexes and agent memory.
- Filter support (tenant, ACL, date) is as important as raw speed.
- Index parameters affect recall and rebuild cost.
Key ideas
- ANN indexes
- Metadata filtering
- Sharding and replication
Vector databases—Pinecone, Weaviate, pgvector, and others—host embedding indexes with hybrid query APIs. Choose based on filter complexity, ops model, and consistency needs. Plan for reindex jobs when embedding models change. Treat collections as part of your data plane: backups, ACL sync, and deletion workflows must match source systems. Practice disaster recovery drills including full reindex from source systems; vector snapshots alone are useless without ingestion pipelines. Ship only after eval gates pass on representative production failures.
Updated 2026-08-09 · Full learning path