fix(legal): add model cascade fallback on 503/429 high demand spikes

services/ai_legal_engine.py

@@ -204,36 +204,50 @@ class AILegalEngine:
             f"{self.get_role_instructions(role)}"
         )
 
-        model_to_use = self.model_name
+        # Каскад моделей при перегрузке или временной недоступности (503/429)
+        models_to_try = [
+            self.model_name,
+            "gemini-flash-latest",
+            "gemini-3-flash-preview"
+        ]
+        # Убираем дубликаты с сохранением порядка
+        seen = set()
+        models_cascade = [m for m in models_to_try if not (m in seen or seen.add(m))]
 
-        try:
-            reply_text = await self._query_rest_api(
-                prompt=prompt,
-                system_instruction=system_instruction,
-                model=model_to_use,
-                audio_path=audio_path
-            )
+        last_err = None
+        for model_to_use in models_cascade:
+            try:
+                reply_text = await self._query_rest_api(
+                    prompt=prompt,
+                    system_instruction=system_instruction,
+                    model=model_to_use,
+                    audio_path=audio_path
+                )
 
-            latency = (datetime.now() - start_time).total_seconds()
+                latency = (datetime.now() - start_time).total_seconds()
 
-            return {
-                "success": True,
-                "text": reply_text.strip(),
-                "answer": reply_text.strip(),
-                "role": role,
-                "latency_sec": round(latency, 2),
-                "model": model_to_use
-            }
+                return {
+                    "success": True,
+                    "text": reply_text.strip(),
+                    "answer": reply_text.strip(),
+                    "role": role,
+                    "latency_sec": round(latency, 2),
+                    "model": model_to_use
+                }
 
-        except Exception as e:
-            latency = (datetime.now() - start_time).total_seconds()
-            log.error(f"Gemini Error: {e}")
-            return {
-                "success": False,
-                "text": f"Ошибка обработки запроса ИИ: {e}",
-                "role": role,
-                "latency_sec": round(latency, 2)
-            }
+            except Exception as e:
+                last_err = e
+                log.warning(f"Model {model_to_use} failed ({e}), trying next fallback...")
+                await asyncio.sleep(0.5)
+
+        latency = (datetime.now() - start_time).total_seconds()
+        log.error(f"All Gemini fallback models failed: {last_err}")
+        return {
+            "success": False,
+            "text": f"Шлюз ИИ временно перегружен. Пожалуйста, повторите вопрос через несколько секунд.",
+            "role": role,
+            "latency_sec": round(latency, 2)
+        }
 
 # Глобальный синглтон юридического ядра для роутеров бота
 legal_engine = AILegalEngine()