lets see

app.py

@@ -10,7 +10,6 @@ from langchain.llms.huggingface_text_gen_inference import HuggingFaceTextGenInfe
 from langchain.prompts import PromptTemplate
 from langchain.chains import RetrievalQA
 
-headers = {"Authorization": f"Bearer {os.environ.get('API_KEY')}"}
 embed_model_id = 'sentence-transformers/all-MiniLM-L6-v2'
 # device = f'cuda:{cuda.current_device()}' if cuda.is_available() else 'cpu'
 
@@ -56,6 +55,7 @@ vectorstore = Pinecone(
 
 
 API_URL = "https://api-inference.huggingface.co/models/HuggingFaceH4/zephyr-7b-beta"
+headers = {"Authorization": f"Bearer {os.environ.get('API_KEY')}"}
 llm = HuggingFaceTextGenInference(
     inference_server_url=API_URL,
     max_new_tokens=512,