Fix ZeroGPU retrieval: pin jina query embedder to CPU

rag.py

@@ -58,8 +58,11 @@ def _chunks() -> dict[int, dict]:
 @lru_cache(maxsize=1)
 def _embedder():
     # transformers ~=4.45 (pinned) loads jina's remote code without shims.
+    # device="cpu" is REQUIRED on ZeroGPU: query embedding runs in retrieve(),
+    # outside the @spaces.GPU block, so CUDA isn't really allocated there — left
+    # on auto it lands on a phantom cuda device and returns zero vectors.
     from sentence_transformers import SentenceTransformer
-    return SentenceTransformer(EMBED_MODEL, trust_remote_code=True)
+    return SentenceTransformer(EMBED_MODEL, trust_remote_code=True, device="cpu")
 
 
 def _embed_query(query: str) -> np.ndarray: