fix(metrics): normalize internal whitespace in ExactMatch, as its docstring promises

docs/07_metric_system.md

@@ -238,12 +238,19 @@ class SemanticSimilarity(BaseMetric):
 ```python
 class ExactMatch(BaseMetric):
     name = "exact_match"
-    description = "Whether answer exactly matches ground truth"
+    description = "Binary metric: 1 if answer matches non-empty ground truth; 0 otherwise"
 
     def evaluate(self, answer: str, ground_truth: str) -> MetricResult:
-        match = answer.strip().lower() == ground_truth.strip().lower()
+        if not ground_truth:
+            return MetricResult(score=0.0, reason="No ground truth provided",
+                                metadata={"match": False})
+
+        normalized_answer = " ".join(answer.lower().split())
+        normalized_truth = " ".join(ground_truth.lower().split())
+        match = normalized_answer == normalized_truth
         return MetricResult(score=1.0 if match else 0.0, 
-                          reason="Exact match" if match else "No match")
+                          reason="Exact match" if match else "No match",
+                          metadata={"match": match})
 ```
 
 ### F1 Score

docs/08_plugin_system.md

@@ -190,7 +190,7 @@ OpenAgent Eval includes the following built-in plugins:
 - `answer_relevancy` - LLM-based relevancy evaluation
 - `hallucination` - Hallucination detection
 - `semantic_similarity` - Sentence transformer similarity
-- `exact_match` - Exact string matching
+- `exact_match` - Case-insensitive, whitespace-normalized exact matching
 - `f1_score` - Token-level F1
 - `bleu` - BLEU score
 - `rouge` - ROUGE score

examples/rag_evaluation_tutorial.ipynb

@@ -1860,7 +1860,7 @@
     "\n",
     "### 12. Exact Match (`exact_match`)\n",
     "\n",
-    "**Theory**: Case-insensitive exact string match.\n",
+    "**Theory**: Case-insensitive exact string match after whitespace normalization.\n",
     "\n",
     "**Why it matters**: Strict correctness for factual QA.\n",
     "\n",

openagent_eval/metrics/generation/exact_match.py

@@ -1,6 +1,7 @@
 """Exact Match metric.
 
-Measures whether the generated answer exactly matches the ground truth.
+Measures whether the generated answer exactly matches the ground truth using
+case-insensitive comparison and whitespace normalization.
 """
 
 from __future__ import annotations
@@ -11,13 +12,14 @@
 
 
 class ExactMatch(BaseMetric):
-    """Binary metric: 1 if answer exactly matches ground truth.
+    """Binary metric: 1 if answer matches non-empty ground truth; 0 otherwise.
 
-    Comparison is case-insensitive and whitespace-normalized.
+    Comparison is case-insensitive and whitespace-normalized. An empty ground
+    truth always scores 0.
     """
 
     name = "exact_match"
-    description = "Binary metric: 1 if answer exactly matches ground truth"
+    description = "Binary metric: 1 if answer matches non-empty ground truth; 0 otherwise"
 
     def evaluate(self, **kwargs: Any) -> MetricResult:
         """Evaluate exact match.
@@ -39,8 +41,8 @@ def evaluate(self, **kwargs: Any) -> MetricResult:
                 metadata={"match": False},
             )
 
-        normalized_answer = answer.strip().lower()
-        normalized_truth = ground_truth.strip().lower()
+        normalized_answer = " ".join(answer.lower().split())
+        normalized_truth = " ".join(ground_truth.lower().split())
         match = normalized_answer == normalized_truth
         score = 1.0 if match else 0.0
 

tests/unit/test_metrics/test_generation.py

@@ -41,6 +41,14 @@ def test_case_insensitive(self):
         )
         assert result.score == 1.0
 
+    def test_internal_whitespace_normalized(self):
+        """Match collapses internal whitespace."""
+        result = self.metric.evaluate(
+            answer="hello  world",
+            ground_truth="hello world",
+        )
+        assert result.score == 1.0
+
     def test_no_match(self):
         """Answer does not match ground truth."""
         result = self.metric.evaluate(