Course outline for Vector and AI Data Infrastructure

Module 1: Define representations and index choices

• Vector representations: Record model, dimensions, and preprocessing contracts so document and query vectors remain compatible without implying factual truth.
• Vector indexes: Compare retrieval quality, query latency, and index size against fixed judgments while recording the active corpus generation.
• Vector databases: Define storage, similarity-query, metadata, and lifecycle responsibilities separately from embedding generation and relevance evaluation.

Practical outcome: Create an embedding/index manifest with model, dimensions, preprocessing, corpus version, and relevance-test assumptions.

Module 2: Exercise storage and search

• pgvector: Inspect PostgreSQL-based vector storage and query responsibilities through the supported local path or supplied equivalent backend fixtures.
• Weaviate: Compare its storage and similarity-query responsibilities with pgvector using supplied fixtures without requiring deployment of both backends.
• Similarity search: Retrieve top-five matches from compatible vectors and assess judged relevance rather than treating closeness as evidence of correctness.

Practical outcome: Run one supported local backend and inspect equivalent supplied query fixtures for the other, comparing index and similarity behavior.

Module 3: Benchmark controls and lifecycle changes

• Metadata filtering: Enforce supplied metadata restrictions on every returned result and verify that otherwise similar documents cannot bypass selection rules.
• Hybrid retrieval: Compare combined lexical and vector results with vector-only retrieval using fixed questions, relevance judgments, and separate latency measurements.
• Index generation and lifecycle contracts: Reject incompatible queries and stale cutovers, propagate deletions, and activate rebuilt generations only after fixed benchmark checks pass.

Practical outcome: Benchmark filtered and hybrid results, reject incompatible query vectors, and verify deletion and new-generation cutover against fixed checks.

Practical exercise

• Build a small retrieval benchmark and version manifest using supplied documents, compatible vectors, and judged questions.
• Run valid queries and reject a model-version or dimension mismatch before search; test metadata restrictions on every returned result.
• Compare vector-only and hybrid top-five results using relevance judgments, recording recall, query latency, and index size separately.
• Delete a document and rebuild a generation; verify it is absent from active results and reject a stale-generation cutover that fails the benchmark.