Cross-encoder reranking

[markgewhite]

Summary

• Added Reranker module using sentence-transformers CrossEncoder with configurable model (default: bge-reranker-v2-m3)
• Integrated reranker into pipeline: hybrid search (30 candidates) → cross-encoder rerank → top 10 → LLM
• 7 unit tests covering scoring, reordering, top_k, edge cases

Closes #8

Test plan

• Reranker module loads cross-encoder model
• rerank() scores query-chunk pairs and returns top_k sorted by relevance
• Model name configurable via config.yaml
• Integrated into pipeline between HybridRetriever and Answerer
• Full pipeline: hybrid (30 candidates) → rerank → top 10 → LLM
• Tests pass (reordering verified, top_k respected, empty input handled)
• Full test suite: 77 tests pass

🤖 Generated with Claude Code