Add early-exit and Detoxify call counters to moderation logs

app/services/moderation_service.py

@@ -64,6 +64,12 @@ ACTION_MAP = {
 }
 
 
+def _log_pipeline_counts(early_exit_count: int, detoxify_call_count: int):
+    print(f"Early exit: {early_exit_count}")
+    print(f"Detoxify çağrısı: {detoxify_call_count}")
+    print(f"CUDA aktif: {torch.cuda.is_available()}")
+
+
 def _extract_blacklist_hits(cleaned_text: str, dil: str):
     pure_text = re.sub(r"[^a-zA-ZçğıöşüÇĞİÖŞÜ0-9\s]", "", cleaned_text).lower()
     words_in_pure_text = set(pure_text.split())
@@ -125,8 +131,10 @@ def _tr_off_scores_batched(text_list, state, batch_size: int = 8):
 
 def _tr_result_with_off_score(cleaned_text: str, profanity_hits, insult_hits, off_score: float, state, dil: str):
     # Only run Detoxify on uncertain content to reduce inference cost.
+    detoxify_called = False
     if off_score < 0.60:
         raw_threat_res = state["D_MULTI"].predict(cleaned_text)
+        detoxify_called = True
     else:
         raw_threat_res = {
             "toxicity": off_score,
@@ -158,16 +166,20 @@ def _tr_result_with_off_score(cleaned_text: str, profanity_hits, insult_hits, of
         },
     )
     detail.update({"action": ACTION_MAP.get(verdict["risk_level"], "MONITOR")})
-    return verdict["decision"], verdict["reason"], verdict["risk_level"], dil, cleaned_text, detail
+    return (verdict["decision"], verdict["reason"], verdict["risk_level"], dil, cleaned_text, detail), detoxify_called
 
 
 def run_moderation(text: str, platform_dil: str = "tr"):
     state = _ensure_runtime_ready()
+    early_exit_count = 0
+    detoxify_call_count = 0
 
     temiz = clean_text_nfkc(text)
     dil = "en" if platform_dil == "en" else "tr"
 
     if is_spam(temiz, dil):
+        early_exit_count += 1
+        _log_pipeline_counts(early_exit_count, detoxify_call_count)
         return (
             "🗑️ SPAM/GİBBERİSH",
             "Anlamsız veya tekrarlı içerik.",
@@ -181,6 +193,8 @@ def run_moderation(text: str, platform_dil: str = "tr"):
 
     # Fast path: if blacklist catches profanity/insult, skip all ML inference.
     if profanity_hits or insult_hits:
+        early_exit_count += 1
+        _log_pipeline_counts(early_exit_count, detoxify_call_count)
         return _blacklist_early_result(profanity_hits, insult_hits, dil, temiz)
 
     if dil == "en":
@@ -191,6 +205,8 @@ def run_moderation(text: str, platform_dil: str = "tr"):
                 "score": float(gb_raw.get("score", 0.0)),
             }
             if gb.get("label", "").lower() == "noise" and gb.get("score", 0.0) > 0.98:
+                early_exit_count += 1
+                _log_pipeline_counts(early_exit_count, detoxify_call_count)
                 return (
                     "🗑️ GIBBERISH/SPAM",
                     "Metin anlamsız karakter dizileri içeriyor.",
@@ -201,6 +217,7 @@ def run_moderation(text: str, platform_dil: str = "tr"):
                 )
 
         raw_res = state["D_EN"].predict(temiz)
+        detoxify_call_count += 1
         res = {k: float(v) for k, v in raw_res.items()}
         tox_score = float(res.get("toxicity", 0.0))
         ins_score = float(res.get("insult", 0.0))
@@ -223,15 +240,22 @@ def run_moderation(text: str, platform_dil: str = "tr"):
             },
         )
         detail.update({"action": ACTION_MAP.get(verdict["risk_level"], "MONITOR")})
+        _log_pipeline_counts(early_exit_count, detoxify_call_count)
         return verdict["decision"], verdict["reason"], verdict["risk_level"], dil, temiz, detail
 
     off_score = _tr_off_scores_batched([temiz], state, batch_size=1)[0]
-    return _tr_result_with_off_score(temiz, profanity_hits, insult_hits, off_score, state, dil)
+    result, detoxify_called = _tr_result_with_off_score(temiz, profanity_hits, insult_hits, off_score, state, dil)
+    if detoxify_called:
+        detoxify_call_count += 1
+    _log_pipeline_counts(early_exit_count, detoxify_call_count)
+    return result
 
 
 def run_moderation_batch(texts, platform_dil: str = "tr", batch_size: int = 8):
     state = _ensure_runtime_ready()
     dil = "en" if platform_dil == "en" else "tr"
+    early_exit_count = 0
+    detoxify_call_count = 0
 
     results = [None] * len(texts)
     tr_pending = []
@@ -241,6 +265,7 @@ def run_moderation_batch(texts, platform_dil: str = "tr", batch_size: int = 8):
         temiz = clean_text_nfkc(text)
 
         if is_spam(temiz, dil):
+            early_exit_count += 1
             results[idx] = (
                 "🗑️ SPAM/GİBBERİSH",
                 "Anlamsız veya tekrarlı içerik.",
@@ -253,11 +278,55 @@ def run_moderation_batch(texts, platform_dil: str = "tr", batch_size: int = 8):
 
         profanity_hits, insult_hits = _extract_blacklist_hits(temiz, dil)
         if profanity_hits or insult_hits:
+            early_exit_count += 1
             results[idx] = _blacklist_early_result(profanity_hits, insult_hits, dil, temiz)
             continue
 
         if dil == "en":
-            results[idx] = run_moderation(text, platform_dil="en")
+            gb = None
+            if state["GB_PIPE"] is not None:
+                gb_raw = state["GB_PIPE"](temiz)[0]
+                gb = {
+                    "label": str(gb_raw.get("label", "")),
+                    "score": float(gb_raw.get("score", 0.0)),
+                }
+                if gb.get("label", "").lower() == "noise" and gb.get("score", 0.0) > 0.98:
+                    early_exit_count += 1
+                    results[idx] = (
+                        "🗑️ GIBBERISH/SPAM",
+                        "Metin anlamsız karakter dizileri içeriyor.",
+                        "LOW",
+                        "en",
+                        temiz,
+                        {"gibberish": gb, "action": "MONITOR", "detox": {}},
+                    )
+                    continue
+
+            raw_res = state["D_EN"].predict(temiz)
+            detoxify_call_count += 1
+            res = {k: float(v) for k, v in raw_res.items()}
+            tox_score = float(res.get("toxicity", 0.0))
+            ins_score = float(res.get("insult", 0.0))
+            identity_attack = float(res.get("identity_attack", 0.0))
+            detail = {
+                "detox": res,
+                "insult": ins_score,
+                "toxicity": tox_score,
+                "identity_attack": identity_attack,
+                "hits": profanity_hits,
+                "insult_hits": insult_hits,
+                "gibberish": gb if state["GB_PIPE"] is not None else None,
+            }
+            verdict = calculate_verdict(
+                profanity_hits,
+                insult_hits,
+                {
+                    "off_score": 0.0,
+                    "detox_toxicity": tox_score,
+                },
+            )
+            detail.update({"action": ACTION_MAP.get(verdict["risk_level"], "MONITOR")})
+            results[idx] = (verdict["decision"], verdict["reason"], verdict["risk_level"], dil, temiz, detail)
             continue
 
         tr_pending.append((idx, temiz, profanity_hits, insult_hits))
@@ -267,6 +336,11 @@ def run_moderation_batch(texts, platform_dil: str = "tr", batch_size: int = 8):
         off_scores = _tr_off_scores_batched(tr_pending_texts, state, batch_size=batch_size)
         for pending, off_score in zip(tr_pending, off_scores):
             idx, temiz, profanity_hits, insult_hits = pending
-            results[idx] = _tr_result_with_off_score(temiz, profanity_hits, insult_hits, off_score, state, dil)
+            result, detoxify_called = _tr_result_with_off_score(temiz, profanity_hits, insult_hits, off_score, state, dil)
+            if detoxify_called:
+                detoxify_call_count += 1
+            results[idx] = result
+
+    _log_pipeline_counts(early_exit_count, detoxify_call_count)
 
     return results