Update retriever.py

retriever.py

@@ -1,7 +1,8 @@
 import os
 import logging
+from pathlib import Path
 
-# Optional FAISS (keeps your original behavior)
+# Optional FAISS (keeps original behavior)
 try:
     import faiss
     _HAS_FAISS = True
@@ -12,14 +13,11 @@ except ImportError:
 from sentence_transformers import SentenceTransformer
 
 # ---- Writable cache + stable repo id for Spaces ----
-_ST_CACHE = os.getenv("SENTENCE_TRANSFORMERS_HOME", "/data/.cache/sentence-transformers")
+_HOME = Path.home()
+_ST_CACHE = os.getenv("SENTENCE_TRANSFORMERS_HOME", str(_HOME / ".cache" / "sentence-transformers"))
 _ST_MODEL_ID = "sentence-transformers/all-MiniLM-L6-v2"  # canonical repo id
 
 def _load_st_model():
-    """
-    Load SentenceTransformer using an explicit cache folder to avoid
-    Hugging Face 'xet' transport / permission issues on Spaces.
-    """
     # Ensure cache dir exists
     try:
         os.makedirs(_ST_CACHE, exist_ok=True)
@@ -37,12 +35,12 @@ def _load_st_model():
             return SentenceTransformer(_ST_MODEL_ID, cache_folder=_ST_CACHE, trust_remote_code=True)
         except Exception as e2:
             logging.exception("Failed loading SentenceTransformer model on both attempts.")
-            raise RuntimeError(
-                f"Failed loading SentenceTransformer '{_ST_MODEL_ID}'.\n"
-                f"First error: {e1}\nSecond error: {e2}\n"
-                f"Check cache dir permissions at: {_ST_CACHE}\n"
-                f"Tip: ensure app.py sets HF_HUB_ENABLE_XET=0 and uses writable caches under /data."
+            # Soft-fail: disable retrieval rather than crashing the whole app
+            logging.error(
+                "Disabling retrieval due to model load failure. "
+                f"Check permissions for {_ST_CACHE} and HF_* env vars."
             )
+            return None
 
 # Load embedding model (works even if FAISS missing)
 _model = _load_st_model()
@@ -56,12 +54,16 @@ def init_retriever(docs=None):
     docs: list[str] to index
     """
     global _index, _docs
+    if _model is None:
+        _docs = docs or []
+        return
     if not _HAS_FAISS:
         _docs = docs or []
         return
 
     if docs:
         _docs = docs
+        import numpy as np
         embeddings = _model.encode(docs, convert_to_numpy=True, normalize_embeddings=False)
         d = embeddings.shape[1]
         _index = faiss.IndexFlatL2(d)
@@ -72,6 +74,8 @@ def retrieve_context(query: str, k: int = 5):
     Retrieve top-k docs matching query.
     Falls back to empty list if FAISS unavailable or not initialized.
     """
+    if _model is None:
+        return []
     if not _HAS_FAISS or _index is None or not _docs:
         return []