Module 7 - Vector Database Engineering
Modern AI applications are defined by their ability to retrieve meaning, not just keywords. Vector databases are the infrastructure layer that makes semantic search, RAG pipelines, recommendation systems, and multimodal retrieval possible at scale.
This module teaches you how vector search actually works - from similarity metrics to approximate nearest neighbor algorithms - and how to operate vector databases reliably in production.
What You Will Learn
Lessons in This Module
| # | Lesson | Key Concepts |
|---|---|---|
| 01 | Vector Similarity Search | Cosine, dot product, L2, recall@K, exact vs approximate |
| 02 | ANN Algorithms | HNSW, IVF, PQ, IVFPQ, LSH, DiskANN |
| 03 | Vector Databases Compared | Pinecone, Weaviate, Qdrant, Chroma, pgvector |
| 04 | Embedding Pipelines | Model selection, batching, re-indexing, drift |
| 05 | Hybrid Search | BM25 + dense, SPLADE, Reciprocal Rank Fusion |
| 06 | Filtering and Metadata | Pre/post filter, ACORN, multi-tenancy, sharding |
| 07 | Scalability and Sharding | Horizontal scale, hot-cold tiering, distributed HNSW |
| 08 | Production Vector DB | Monitoring, capacity planning, disaster recovery |
Key Mental Models
Recall vs Latency is the central tradeoff. Exact search guarantees 100% recall but scales as . Approximate search trades a few percent of recall for orders-of-magnitude speedup. The right operating point depends on your application.
Embeddings are not static. When you upgrade your embedding model, every vector in your index becomes stale. Re-indexing 50M documents takes time, compute, and a zero-downtime migration strategy.
Filtering changes everything. Adding a metadata filter to a vector query sounds trivial but can destroy index efficiency. Pre-filtering, post-filtering, and hybrid approaches (ACORN) each have different performance characteristics.
The database is not the hard part. Choosing between Pinecone and Qdrant is less important than getting your embedding pipeline right, your recall evaluation correct, and your metadata schema designed for the queries you'll actually run.
Prerequisites
- Familiarity with embeddings and similarity search concepts (Module 05)
- Basic Python and NumPy
- Understanding of database indexing fundamentals