Module 1 · Building the index · Lesson 3
Vector indexes and databases
How HNSW and IVF indexes trade recall for latency, how filters and tenant isolation interact with them, how to choose between pgvector, Turbopuffer, LanceDB and Qdrant, and how to keep an index fresh as documents change.
By the end of this lesson you can
- Explain how IVF and HNSW indexes find approximate neighbours, and tune nprobe and ef_search against a recall target
- Predict when a filtered vector query returns too few results, and choose between exact search, iterative scans, partial indexes and partitioning
- Design tenant and case isolation that is enforced by the database, not by the prompt
- Choose a vector store for a given workload and justify it on latency, cost, filtering and operations
- Keep an index fresh with idempotent upserts, supersession, redaction deletes and blue-green re-embedding