Add top-score magnitude gating to hybrid fusion alpha

[Destrayon]

Description

Entropy-based adaptive alpha alone misses the failure mode where a retrieval leg has a peaked distribution but absolutely low scores. Add magnitude gating: multiply the entropy-derived weight by clamp(max_raw_score / threshold, 0, 1) so that legs with low absolute scores are downweighted regardless of distribution shape. This approach was validated at the ICML 2025 Workshop on Vector Databases (entropy-based dynamic hybrid retrieval paper).

Evidence: exp-020 showed +2.77% retrieval quality improvement with no speed cost.

Acceptance Criteria

• ComputeEffectiveAlpha incorporates magnitude factor alongside entropy
• VectorMagnitudeThreshold and KeywordMagnitudeThreshold settings added
• EnableMagnitudeGating toggle added (default: true)
• Magnitude gating can be disabled without affecting existing entropy logic
• Tests pass (dotnet test)

Implementation Notes

• Key files: src/Connapse.Search/Hybrid/HybridSearchService.cs, src/Connapse.Core/Models/SettingsModels.cs
• Pattern: Extend existing ComputeEffectiveAlpha method
• Reference: auto-rag-eval/results/exp-020-query_adaptive_alpha-e0a55f-changespec.json

Size Estimate

S — ~60 lines across 2 files

Related

none