fix: remove score validation bounds to support raw logits from cross-encoder

[fparrav]

Summary

• Fixes validation error when using cross-encoder models that return raw logits (e.g., ms-marco-MiniLM-L-6-v2)
• Removes artificial -1.0 to 1.0 constraint on rerank scores that was blocking legitimate model outputs

Problem

The current code enforces score bounds (ge=-1.0, le=1.0) in RerankResult, which causes validation errors with standard cross-encoder models that output raw logits (e.g., scores like 7.45 or -2.3). This breaks reranking functionality with industry-standard models.

Solution

• Remove ge=-1.0, le=1.0 constraints from RerankResult.score field
• Allow models to return their natural output format (raw logits)
• Normalization (sigmoid) is already handled at the router layer when OPENAI_RERANK_AUTO_SIGMOID=true

Testing

Tested with cross-encoder/ms-marco-MiniLM-L-6-v2 on macOS Apple Silicon:

• Before: Validation error on scores > 1.0
• After: Scores pass validation, normalization applied correctly at router level

Impact

• Unblocks compatibility with most sentence-transformers cross-encoder models
• Restores documented behavior of auto-sigmoid feature
• No breaking changes (only removes overly restrictive validation)

[Copilot]

Pull request overview

This PR fixes a validation error that prevented legitimate cross-encoder reranking models from functioning correctly. The previous code enforced artificial score bounds (-1.0 to 1.0) on the RerankResult.score field, which caused validation errors when cross-encoder models (like cross-encoder/ms-marco-MiniLM-L-6-v2) returned raw logits outside this range (e.g., 7.45 or -2.3).

Key Changes:

• Removes the ge=-1.0, le=1.0 constraints from the RerankResult.score field to allow raw logits
• Updates the field description to clarify that scores can be raw logits or normalized values
• Minor formatting improvements to ConfigDict definitions (multi-line format)

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[joonsoome]

@fparrav Thanks for the PR! I reviewed the response model changes—removing the ge/le constraint on RerankResult.score makes sense now that scores can be raw logits or normalized differently, and it should prevent Pydantic validation errors when values fall outside [-1, 1]. Since this is limited to the response schema/description, the risk looks low. LGTM — approved. If it’s not already covered, a quick note in the API docs about whether scores are normalized would be helpful too.