feat: Implement lazy module loading for zero-overhead server startup

vector_search/embed.py

@@ -12,17 +12,10 @@ import os
 from pathlib import Path
 from typing import List, Union
 import numpy as np
-import torch
 import config
 
 logger = logging.getLogger(__name__)
 
-# Optimize PyTorch CPU parallelism
-try:
-    torch.set_num_threads(max(1, torch.get_num_threads()))
-except Exception:
-    pass
-
 _EMBEDDER_INSTANCE = None
 
 

vector_search/rerank.py

@@ -4,9 +4,7 @@ import re
 from pathlib import Path
 from typing import Any, Dict, List
 import numpy as np
-import torch
 from rank_bm25 import BM25Okapi
-from sentence_transformers.cross_encoder import CrossEncoder
 import config
 
 logger = logging.getLogger(__name__)
@@ -180,6 +178,7 @@ class PyTorchCrossEncoderRanker:
     PyTorch fallback wrapper for sentence-transformers CrossEncoder.
     """
     def __init__(self, model_name: str = config.CROSS_ENCODER_MODEL_NAME):
+        from sentence_transformers.cross_encoder import CrossEncoder
         load_path = model_name
         local_cache = getattr(config, "CROSS_ENCODER_LOCAL_CACHE", None)
         if local_cache and Path(local_cache).exists():