pooling update

app.py

@@ -28,10 +28,16 @@ def predict_similarity(question, candidate_answer, ai_answer):
     }
     ort_outputs = ort_session.run(None, ort_inputs)
 
-    # Calculate cosine similarity
+    # Get embeddings (shape: (seq_length, 768))
     embeddings = ort_outputs[0]
-    similarity = np.dot(embeddings[0], embeddings[1]) / \
-        (np.linalg.norm(embeddings[0]) * np.linalg.norm(embeddings[1]))
+
+    # Apply mean pooling to reduce (seq_length, 768) to (768,)
+    candidate_embedding = np.mean(embeddings[0], axis=0)  # Shape (768,)
+    ai_embedding = np.mean(embeddings[1], axis=0)  # Shape (768,)
+
+    # Calculate cosine similarity
+    similarity = np.dot(candidate_embedding, ai_embedding) / \
+        (np.linalg.norm(candidate_embedding) * np.linalg.norm(ai_embedding))
 
     return float(similarity)