Update app.py

app.py

@@ -38,11 +38,13 @@ os.makedirs(os.environ["NUMBA_CACHE_DIR"], exist_ok=True)
 os.environ["NUMBA_DISABLE_CACHE"] = "1"
 
 # 2) Config from ENV
-MODEL_NAME = os.getenv("EMBED_MODEL", "seznam/simcse-small-e-czech")
+# Read model name from env and normalize to lowercase to match HF repo ID
+env_model = os.getenv("EMBED_MODEL", "Seznam/simcse-small-e-czech")
+MODEL_NAME = env_model
 MIN_TOPIC  = int(os.getenv("MIN_TOPIC_SIZE", "10"))
 MAX_DOCS   = int(os.getenv("MAX_DOCS", "5000"))
 
-# 3) Set HF cache envs to a writeable folder (once at startup)
+# 3) Set HF cache envs to a writeable folder (once at startup) envs to a writeable folder (once at startup)
 cache_dir = "/tmp/hfcache"
 os.makedirs(cache_dir, exist_ok=True)
 import stat
@@ -52,11 +54,17 @@ os.environ["TRANSFORMERS_CACHE"] = cache_dir
 os.environ["SENTENCE_TRANSFORMERS_HOME"] = cache_dir
 
 # 4) Initialise embeddings once
-embeddings = SentenceTransformer(MODEL_NAME, cache_folder=cache_dir)
+# Use huggingface_hub to snapshot-download the model locally
+from huggingface_hub import snapshot_download
+print(f"Downloading model {MODEL_NAME} to {cache_dir}...")
+sys.stdout.flush()
+local_model_path = snapshot_download(repo_id=MODEL_NAME, cache_dir=cache_dir)
+
+# Load SentenceTransformer from local path
+embeddings = SentenceTransformer(local_model_path, cache_folder=cache_dir)
 
 # Pre-initialize fallback global models for small-batch debugging
-# Global UMAP: 2-neighbors, cosine space, random init
-global_umap = UMAP(
+# Global UMAP: 2-neighbors, cosine space, random init(
     n_neighbors=2,
     metric="cosine",
     init="random",