Update app.py

app.py

@@ -35,6 +35,10 @@ model = AutoModelForCausalLM.from_pretrained(model_name)
 tokenizer.pad_token = tokenizer.eos_token
 
 
+# FastAPI setup
+app = FastAPI(title="Apollo RAG Medical Chatbot")
+
+
 # Generation settings
 generation_config = GenerationConfig(
     max_new_tokens=150,
@@ -75,9 +79,6 @@ qa_chain = RetrievalQA.from_chain_type(
     chain_type="stuff"
 )
 
-# FastAPI setup
-app = FastAPI(title="Apollo RAG Medical Chatbot")
-
 class Query(BaseModel):
     question: str = Field(..., example="ما هي اسباب تساقط الشعر ؟", min_length=3)