ADR 0004: Full-text search first; embeddings and LLMs only after Phase 1

Context

Semantic search, overlap detection and concept tagging could all use embeddings or LLMs. But the right grain (slide, objective or event), the right model, and whether a hosted vendor is allowed all depend on what faculty actually ask. LLM use is also waiting on institutional approval.

Decision

Build BM25 full-text search over pages (title, text, speaker notes), objectives and events first. Phase 1 measures where it fails on a realistic question set and on objective-to-material support, then decides whether embeddings are needed and at what grain. MeSH entry terms are available for synonym expansion without any model.

Consequences

The first faculty tools rest on deterministic, explainable retrieval. Every LLM or embedding output added later must record the model, version, date and prompt/parameters next to the derived value.