Files
hermes-agent/tools
Hermes Agent 4129cc0d0c
Some checks failed
Docker Build and Publish / build-and-push (pull_request) Has been skipped
Contributor Attribution Check / check-attribution (pull_request) Failing after 37s
Supply Chain Audit / Scan PR for supply chain risks (pull_request) Successful in 55s
Tests / test (pull_request) Failing after 55s
Tests / e2e (pull_request) Successful in 2m49s
feat: implement Reader-Guided Reranking — bridge R@5 vs E2E gap (#666)
Resolves #666. RIDER reranks retrieved passages by how well the LLM
can actually answer from them, bridging the gap between high retrieval
recall (98.4% R@5) and low end-to-end accuracy (17%).

agent/rider.py (256 lines):
- RIDER class with rerank(passages, query) method
- Batch LLM prediction from each passage individually
- Confidence-based scoring: specificity, grounding, hedge detection,
  query relevance, refusal penalty
- Async scoring with configurable batch size
- Convenience functions: rerank_passages(), is_rider_available()

tools/session_search_tool.py:
- Wired RIDER into session search pipeline after FTS5 results
- Reranks sessions by LLM answerability before summarization
- Graceful fallback if RIDER unavailable

tests/test_reader_guided_reranking.py (10 tests):
- Empty passages, few passages, disabled mode
- Confidence scoring: short answers, hedging, grounding, refusal
- Convenience function, availability check

Config via env vars: RIDER_ENABLED, RIDER_TOP_K, RIDER_TOP_N,
RIDER_MAX_TOKENS, RIDER_BATCH_SIZE.
2026-04-15 07:40:15 -04:00
..