diff --git a/app/chain.py b/app/chain.py
index de68d9c..0ce9a94 100644
--- a/app/chain.py
+++ b/app/chain.py
@@ -3,13 +3,23 @@ from langchain_core.runnables import RunnablePassthrough
 from langchain_core.output_parsers import StrOutputParser
 from langchain_core.prompts import ChatPromptTemplate
 from langchain_core.runnables import RunnableLambda
-from app.retriever import get_retriever
-from app.llm import TritonLLM
+from retriever import get_retriever
+from llm import TritonLLM
+import json
 
 def build_chain():
     retriever = get_retriever()
     llm = TritonLLM()
 
+    def build_chat_input(x):
+        context = format_docs(retriever.invoke(x["question"]))
+        messages = [
+            {"role": "system", "content": f"다음 정보를 참고하여 사용자의 질문에 답변하세요:\n\n{x["context"]}\n\n답변할 수 없는 내용이면 솔직하게 모른다고 말하세요."},
+            {"role": "user", "content": x["question"]}
+        ]
+
+        return  json.dumps(messages, ensure_ascii=False)
+
     # RAG 프롬프트 템플릿 생성
     prompt_template = ChatPromptTemplate.from_messages([
         ("system", """다음 정보를 참고하여 사용자의 질문에 답변하세요:
@@ -21,15 +31,20 @@ def build_chain():
         ("human", "{question}")
     ])
 
+    llm_with_params = RunnableLambda(
+        lambda x: llm.invoke(x["question"], temperature=x.get("temperature"), max_tokens=x.get("max_tokens"))
+    )
+
     def format_docs(docs):
         return "\n\n".join(f"문서: {i+1}\n{doc.page_content}" for i, doc in enumerate(docs))
 
     rag_chain = (
             {
                 "context": RunnableLambda(lambda x: x["question"]) | retriever | format_docs,
-                "question": RunnablePassthrough()
+                "question": lambda x: x["question"]
             }
-            | prompt_template
+            #| prompt_template
+            | RunnableLambda(build_chat_input)
             | llm
             | StrOutputParser()
     )