New changes

app.py

@@ -564,7 +564,19 @@ def evaluate_rag_pipeline(domain, q_indices):
         gt_r = dataset[i].get('relevance_score')
         gt_u = dataset[i].get('utilization_score')
         gt_c = dataset[i].get('completeness_score')
-        gt_a = dataset[i].get('gpt3_adherence')
+
+        adherence_raw = dataset[i].get('adherence_score')
+        if isinstance(adherence_raw, bool):
+            gt_a = 1 if adherence_raw is True else 0
+        elif isinstance(adherence_raw, str):
+            if adherence_raw.lower() == "true":
+                gt_a = 1
+            elif adherence_raw.lower() == "false":
+                gt_a = 0
+            else:
+                gt_a = None
+        else:
+            gt_a = None
 
         safe_append(gt_relevance, pred_relevance, gt_r, predicted['Context Relevance'])
         safe_append(gt_utilization, pred_utilization, gt_u, predicted['Context Utilization'])