Update app.py

app.py

@@ -15,7 +15,9 @@ class ChatConfig:
     DEFAULT_TEMP = 0.3
     DEFAULT_TOP_P = 0.95
 
-client = InferenceClient(ChatConfig.MODEL)
+HF_TOKEN = os.getenv("HF_TOKEN")  # Fetch from environment variables
+
+client = InferenceClient(ChatConfig.MODEL, token=HF_TOKEN)
 embed_model = SentenceTransformer("all-MiniLM-L6-v2")  # Lightweight embedding model
 vector_dim = 384  # Embedding size
 index = faiss.IndexFlatL2(vector_dim)  # FAISS index