adjust parallel search and rerank

app/gemini_client.py

@@ -9,6 +9,7 @@ from .request_limit_manager import RequestLimitManager
 from .utils import (
     _safe_truncate
 )
+from .config import get_settings
 
 class GeminiResponseError(Exception):
     """Custom exception for non-retriable Gemini response issues like safety or token limits."""
@@ -24,6 +25,10 @@ class GeminiResponseError(Exception):
 class GeminiClient:
     def __init__(self):
         self.limit_manager = RequestLimitManager("gemini")
+        settings = get_settings()
+        num_keys = len(settings.gemini_api_keys.split(',')) if settings.gemini_api_keys else 0
+        num_models = len(settings.gemini_models.split(',')) if settings.gemini_models else 0
+        logger.info(f"[GEMINI_INIT] Limiter is considering {num_keys} API keys and {num_models} models.")
         self._cached_model = None
         self._cached_key = None
         self._cached_model_instance = None