try

app.py

@@ -15,7 +15,7 @@ chunk_embeddings = model.encode(chunks, convert_to_tensor=True)
 # Hugging Face Inference Client (add token if required)
 client = InferenceClient(
     model='Qwen/Qwen2.5-72B-Instruct',
-    # token=os.getenv("HF_TOKEN")
+    token = 'HF_TOKEN'
 )
 
 # Top similar chunks