Update app/search.py

app/search.py

@@ -1,37 +1,77 @@
 import json
 from pathlib import Path
-from typing import List, Dict
-from sentence_transformers import SentenceTransformer
+from typing import List, Dict, Optional
 import numpy as np
 
+from sentence_transformers import SentenceTransformer
+
+# ---------- Global embedder (loaded once, CPU-safe) ----------
+_EMBEDDER: Optional[SentenceTransformer] = None
+
+def _get_embedder() -> SentenceTransformer:
+    global _EMBEDDER
+    if _EMBEDDER is None:
+        # Explicit device="cpu" avoids any device_map/meta init paths.
+        # Use the canonical model id to avoid redirect surprises.
+        _EMBEDDER = SentenceTransformer(
+            "sentence-transformers/all-MiniLM-L6-v2",
+            device="cpu"
+        )
+        # Optional: shorten for speed on Spaces; keep accuracy reasonable
+        _EMBEDDER.max_seq_length = 256
+    return _EMBEDDER
+
 def load_index(env: Dict):
-    import faiss, json
+    import faiss
     index_path = Path(env["INDEX_DIR"]) / "faiss.index"
     meta_path = Path(env["INDEX_DIR"]) / "meta.json"
     if not index_path.exists():
         raise RuntimeError("Index not found. Run ingest first.")
     index = faiss.read_index(str(index_path))
-    metas = json.load(open(meta_path))
+    with open(meta_path, "r") as f:
+        metas = json.load(f)
     return index, metas
 
-def embed(texts: List[str]):
-    model = SentenceTransformer("all-MiniLM-L6-v2")
-    return model.encode(texts, convert_to_numpy=True, normalize_embeddings=True)
+def embed(texts: List[str]) -> np.ndarray:
+    emb = _get_embedder()
+    vecs = emb.encode(
+        texts,
+        convert_to_numpy=True,
+        normalize_embeddings=True,
+        show_progress_bar=False,
+        batch_size=32,
+    )
+    # FAISS expects float32
+    if vecs.dtype != np.float32:
+        vecs = vecs.astype(np.float32, copy=False)
+    return vecs
 
 def search(q: str, env: Dict, top_k: int = 15, filters: Dict = None) -> List[Dict]:
+    import faiss
     index, metas = load_index(env)
-    qv = embed([q])
-    scores, idxs = index.search(qv, top_k)
+
+    qv = embed([q])  # shape (1, d) float32
+    # Defensive: ensure index dim matches query dim
+    if hasattr(index, "d") and index.d != qv.shape[1]:
+        raise RuntimeError(f"FAISS index dim {getattr(index, 'd', '?')} "
+                           f"!= embedding dim {qv.shape[1]}")
+
+    scores, idxs = index.search(qv, top_k)  # scores shape (1, k), idxs shape (1, k)
+
     results = []
+    f_geo = (filters or {}).get("geo")
+    f_cats = (filters or {}).get("categories")
+
     for score, idx in zip(scores[0], idxs[0]):
-        if idx == -1: continue
-        m = metas[idx]
-        if filters:
-            if "geo" in filters and filters["geo"] and m.get("geo") not in filters["geo"]:
+        if idx == -1:
+            continue
+        m = dict(metas[idx])  # copy so we don’t mutate the cached list
+        if f_geo and m.get("geo") not in f_geo:
+            continue
+        if f_cats:
+            if not set(f_cats).intersection(set(m.get("categories", []))):
                 continue
-            if "categories" in filters and filters["categories"]:
-                if not set(filters["categories"]).intersection(set(m.get("categories",[]))):
-                    continue
         m["score"] = float(score)
         results.append(m)
+
     return results