Summary
A deployed, indexed, benchmarked vector database. The infrastructure piece behind every RAG system, treated as a production service rather than a demo.
Objectives
- 01Choose a vector DB with a written justification.
- 02Deploy it with authentication and backups.
- 03Load and index at least 100,000 vectors.
- 04Benchmark queries across parameter sweeps.
Key Ideas
- A vector DB is a database; treat it with database hygiene.
- The tradeoff is always recall vs. latency vs. memory — pick two.
Lab
Deploy one vector database (pgvector, Qdrant, Weaviate, or Pinecone) as a real service — authenticated, backed up, monitored. Load at least 100,000 embeddings from a corpus of your choice. Benchmark ANN query latency at k=10 across three index parameter sets (e.g., HNSW ef_construction values). Report recall vs. latency tradeoffs. Document the runbook: how to redeploy, how to reindex, how to roll back.
Deliverables
- — Deployed service with credentials and monitoring.
- — Ingest script and re-run instructions.
- — Benchmark report (recall vs. latency at 3 configs).
- — Runbook covering deploy, reindex, and rollback.
Rubric
- — Realism — deployed, authenticated, not a laptop demo.
- — Scale — at least 100k vectors indexed.
- — Benchmarking — the tradeoff curve is visible and honest.
- — Operability — the runbook works for a stranger.