0002 — Hybrid retrieval (Lesson 2)
Date: 2026-06-22
Taught: Inside stage 3. Vector (dense, meaning) vs BM25 (sparse, exact tokens) fail in opposite directions; hybrid runs both and merges with RRF (rank-based, sidesteps incompatible score scales); reranking previewed as the optional final pass.
Anchoring: grounded all four query examples in DuckyCutz (customer free-text → vector; order ids / barber tables / branch names → BM25). Tied directly to gbrain's stated "vector + BM25 + RRF + graph signals" so 3 of those 4 words are now understood.
Spacing: opened with a no-peek recall of Lesson 1's five stages.
ZPD for Lesson 3: he understands dense+sparse+RRF. Next gap is reranking proper (why first-pass order isn't final, cross-encoder vs bi-encoder, cost/latency tradeoff). Still conceptual. After Lesson 3, retrieval model is complete enough to start the applied phase OR pivot to Lesson 4 (graph signals — the fourth word, and where Graphiti lives).
Decision point coming: after L3, ask RT whether to (a) do Lesson 4 graph/temporal conceptually, or (b) go hands-on now (clone gbrain, inspect his Graphiti graph). He said concepts-first but he's a fast practitioner; offer the fork.