Update app.py

app.py

@@ -18,6 +18,57 @@ os.environ["MKL_NUM_THREADS"] = str(psutil.cpu_count(logical=False))
 # Set device globally
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 
+@st.cache_data
+def preprocess_text(text):
+    # Add any text cleaning or normalization here
+    return text.strip()
+
+# Optimized function for evidence extraction and classification with better CPU performance
+def perform_verification(claim, context, model_qatc, tokenizer_qatc, model_tc, tokenizer_tc, 
+                          model_bc, tokenizer_bc, tfidf_threshold, length_ratio_threshold):
+    # Extract evidence
+    evidence_start_time = time.time()
+    evidence = extract_evidence_tfidf_qatc(
+        claim, context, model_qatc, tokenizer_qatc,
+        DEVICE,
+        confidence_threshold=tfidf_threshold,
+        length_ratio_threshold=length_ratio_threshold
+    )
+    evidence_time = time.time() - evidence_start_time
+    
+    # Explicit garbage collection after evidence extraction
+    gc.collect()
+    
+    # Classify the claim
+    verdict_start_time = time.time()
+    with torch.no_grad():
+        verdict = "NEI"
+        prob3class, pred_tc = classify_claim(
+            claim, evidence, model_tc, tokenizer_tc, DEVICE
+        )
+        
+        # Only run binary classifier if needed
+        prob2class, pred_bc = 0, 0
+        if pred_tc != 0:
+            prob2class, pred_bc = classify_claim(
+                claim, evidence, model_bc, tokenizer_bc, DEVICE
+            )
+            verdict = "SUPPORTED" if pred_bc == 0 else "REFUTED" if prob2class > prob3class else ["NEI", "SUPPORTED", "REFUTED"][pred_tc]
+            
+    verdict_time = time.time() - verdict_start_time
+
+    return {
+        "evidence": evidence,
+        "verdict": verdict,
+        "evidence_time": evidence_time,
+        "verdict_time": verdict_time,
+        "prob3class": prob3class.item() if isinstance(prob3class, torch.Tensor) else prob3class,
+        "pred_tc": pred_tc,
+        "prob2class": prob2class.item() if isinstance(prob2class, torch.Tensor) else prob2class,
+        "pred_bc": pred_bc
+    }
+
+
 # Set page configuration
 st.set_page_config(
     page_title="SemViQA - Hệ thống Kiểm chứng Thông tin Tiếng Việt",