Reference · Cheatsheet
Re-rank patterns
Retrieve wide, re-rank narrow.
Retrieve wide, re-rank narrow.
Pipeline
1. **Keyword** BM25 top-20 (SKUs, codes, exact phrases)
2. **Vector** top-20 (paraphrases, concepts)
3. **Merge** dedupe by chunk ID
4. **Cross-encoder** score pairs → top-5
5. **LLM** generate with citations
When to skip re-rank
- Tiny corpus (<200 chunks) — vector-only may suffice
- Strict latency budget <200ms total
Tuning
**Related:** glossary `cross-encoder`, `re-ranking`, `hybrid-search`