Update src/rag.py

src/rag.py

@@ -314,7 +314,7 @@ def load_generation_model():
         model = AutoModelForCausalLM.from_pretrained(
             model_name,
             device_map="cpu",
-            token = token
+            token = token,
             torch_dtype=torch.float16,
             low_cpu_mem_usage=True,