更新

rag_eval/metrics/pipeline.py

@@ -94,6 +94,23 @@ class MetricPipeline:
                 reference=sample.ground_truth,
                 retrieved_contexts=sample.contexts,
             )
+        elif name == "noise_sensitivity":
+            coroutine = metric.ascore(
+                user_input=sample.question,
+                response=sample.answer,
+                reference=sample.ground_truth,
+                retrieved_contexts=sample.contexts,
+            )
+        elif name == "factual_correctness":
+            coroutine = metric.ascore(
+                response=sample.answer,
+                reference=sample.ground_truth,
+            )
+        elif name == "semantic_similarity":
+            coroutine = metric.ascore(
+                reference=sample.ground_truth,
+                response=sample.answer,
+            )
         else:
             raise ValueError(f"Unsupported metric: {name}")
 

rag_eval/metrics/registry.py

@@ -1,8 +1,13 @@
 """Supported metric names recognized by scenario validation and pipeline setup."""
 
 SUPPORTED_METRICS = {
+    # Core retrieval / generation metrics (always available).
     "faithfulness",
     "answer_relevancy",
     "context_recall",
     "context_precision",
+    # Robustness and end-to-end metrics (see 架构设计 §10.2).
+    "noise_sensitivity",      # 鲁棒性：对检索噪声的敏感度
+    "factual_correctness",    # 端到端：回答相对标准答案的事实正确性
+    "semantic_similarity",    # 端到端：回答与标准答案的语义相似度（embedding，无 LLM 调用）
 }