fix vector python embedding

python/vector/vector_service.py

@@ -1,6 +1,7 @@
 #!/usr/bin/env python3
 
 import json
+import os
 import logging
 from logging.handlers import RotatingFileHandler
 import threading
@@ -13,6 +14,12 @@ import numpy as np
 from fastapi import FastAPI, HTTPException, Request
 from fastapi.responses import JSONResponse
 from pydantic import BaseModel
+
+# Keep HuggingFace/SentenceTransformer model loading deterministic.
+os.environ.setdefault("HF_HUB_DISABLE_XET", "1")
+os.environ.setdefault("HF_HUB_ETAG_TIMEOUT", "10")
+os.environ.setdefault("HF_HUB_DOWNLOAD_TIMEOUT", "30")
+
 from sentence_transformers import SentenceTransformer
 
 
@@ -147,6 +154,15 @@ def _safe_read_json(path: Path) -> Optional[Any]:
         return None
 
 
+
+def _resolve_embedding_model_name(configured_model_name: str) -> str:
+    # A local model directory avoids implicit network/cache lookups in production.
+    model_override = os.environ.get("RETRIEX_EMBEDDING_MODEL_PATH", "").strip()
+    if model_override:
+        return model_override
+
+    return configured_model_name.strip()
+
 def _as_key(value: Any) -> Optional[str]:
     if value is None:
         return None
@@ -362,10 +378,12 @@ def load_all() -> None:
         if not embedding_model_name:
             raise RuntimeError("embedding_model missing in index_meta.json")
 
-        if model is None or embedding_model_name != loaded_embedding_model_name:
-            logger.info("[Reload] Loading embedding model: %s", embedding_model_name)
-            model = SentenceTransformer(embedding_model_name)
-            loaded_embedding_model_name = embedding_model_name
+        resolved_embedding_model_name = _resolve_embedding_model_name(str(embedding_model_name))
+
+        if model is None or resolved_embedding_model_name != loaded_embedding_model_name:
+            logger.info("[Reload] Loading embedding model: %s", resolved_embedding_model_name)
+            model = SentenceTransformer(resolved_embedding_model_name)
+            loaded_embedding_model_name = resolved_embedding_model_name
 
         runtime = _safe_read_json(INDEX_RUNTIME_PATH)
         chunk_runtime_stamp, tags_runtime_stamp, tags_index_present = _extract_runtime_state(runtime)