How to wire hybrid search for RAG

Keyword + vector + re-rank in a minimal pipeline.

1. Run BM25 (or keyword index) for top 20

2. Run vector search for top 20 on same corpus

3. Merge by doc ID, dedupe chunks

4. Re-rank merged set with cross-encoder or LLM scorer

5. Pass top 5 chunks to generation prompt

Add `product`, `version`, or `region` filters before search when docs are tagged.

**Try the lesson:** `hybrid-search-rag` in Lane C