Problem

External API rerankers (SiliconFlow, Cohere, etc.) return pre-normalized relevance_score in [0, 1] with very small absolute values. The current code applies sigmoid to all scores, assuming they are logits. This compresses everything to ~0.5, destroying the ranking signal and making recency the sole sorting factor.

Example with SiliconFlow BAAI/bge-reranker-v2-m3

With sigmoid, all scores are ~0.5 and recency becomes the only ranking signal. With rank-based normalization, the CE signal correctly dominates.

Fix

Detect the score range in CrossEncoderReranker.rerank():

• If all scores are in [0, 1]: Use rank-based normalization with tie handling (equal scores get equal ranks). This preserves relative ordering without depending on absolute score magnitudes.
• Otherwise (logits): Use sigmoid as before. This maintains backward compatibility with local models (e.g. cross-encoder/ms-marco-MiniLM-L-6-v2).

Testing

Verified with real SiliconFlow API scores:

• SiliconFlow [0,1] small values: correctly rank-normalized (1.0 > 0.5 > 0.0)
• Local model logits: correctly uses sigmoid (0.92 > 0.86 > 0.57)
• All identical scores: correctly assigned equal ranks (no artificial separation)
• Tied scores: correctly grouped (e.g. two tied at top → 1.0, distinct lower → 0.0)
• Empty input: returns []
• Passthrough reranker: unaffected (apply_combined_scoring overrides with RRF rank)

Unit tests added in tests/test_reranker_score_normalization.py.

Related

• Closes the ranking quality issue for SiliconFlow users (related to Support OpenAI-compatible reranker endpoint (e.g. SiliconFlow) via base_url override #859)
• Does not affect Bug: recall endpoint TypeError: bad operand type for unary -: list in cross-encoder sigmoid #1369 fix (PR fix(reranker): handle 2-D cross-encoder scores so recall does not crash (closes #1369) #1416) which handles 2D score flattening