Add Retrieval-Conditioned Confidence Metric (RCCS) for RAG evaluation

[AIMindCrafter]

Description

This pull request implements the Retrieval-Conditioned Confidence Score (RCCS) metric under metrics/rccs/ as proposed in issue #736.

RCCS provides a lightweight, model-agnostic calibration diagnostic for Retrieval-Augmented Generation (RAG) pipelines. By evaluating the joint alignment between retrieval relevance ($R$), model confidence ($C$), and ground-truth correctness ($A$), this metric helps detect instances of "confident hallucinations" (i.e., high confidence in the output despite poor or missing retrieved context).

Key Components Implemented:

1. Inputs Accepted:

   • retrieval_score: The relevance/quality score ($R$) of retrieved documents.
   • confidence_score: The model's generation confidence ($C$).
   • correctness: Binary ground-truth correctness ($A$, $0$ or $1$).

2. Metrics Returned:

   • rccs_correlation: Pearson correlation between $(R \times C)$ and $A$.
   • confidence_calibration_error: Mean absolute error mean(|R * C - A|).
   • mean_rc: Average of $(R \times C)$.
   • n: Total sample count.

---

Files Added

• metrics/rccs/rccs.py: Core metric implementation inheriting from evaluate.Metric.
• metrics/rccs/README.md: Comprehensive metric card documentation, installation instructions, and examples.
• metrics/rccs/requirements.txt: Dependencies list (scipy, numpy).
• metrics/rccs/app.py: Gradio app launcher script for the metric widget.

---

Verification and Tests

1. Doctests:
   Successfully passed using pytest on the local module test suite:

   pytest tests/test_metric_common.py -k rccs

   Output:

   tests/test_metric_common.py .
   tests/test_metric_common.py .s
   =========== 1 passed, 1 skipped, 133 deselected, 2 warnings in 7.15s ===========
   

2. Manual verification:

   import evaluate
   rccs = evaluate.load("metrics/rccs")
   results = rccs.compute(
       retrieval_score=[0.8, 0.3, 0.5],
       confidence_score=[0.9, 0.2, 0.7],
       correctness=[1, 0, 1]
   )
   print(results)

   Output:

   {
       'rccs_correlation': 0.8290069500014531,
       'confidence_calibration_error': 0.33000001311302185,
       'mean_rc': 0.3766666650772095,
       'n': 3
   }