Update api.py

api.py

@@ -260,7 +260,12 @@ class SimilarityMemory:
 
     def merge_prompt(self, prompt: str, memories: list):
         context = "\n".join(memories)
-        return f"{context}\n\n{prompt}" if context else prompt
+        full_prompt = ""
+        if context:
+            full_prompt += f"기억:\n{context}\n\n"
+        full_prompt += f"현재 질문:\n{prompt}\n\n응답:"
+        return full_prompt
+
 
 memory = SimilarityMemory()
 
@@ -302,11 +307,23 @@ def respond(input_text):
     merged_prompt = memory.merge_prompt(input_text, related_memories)
 
     for _ in range(3):  # 최대 3번 재시도
-        response = generate_text_sample(model, merged_prompt)
+        full_response = generate_text_sample(model, merged_prompt)
+
+        # 여기서 '응답:' 뒤의 텍스트만 뽑기
+        if "응답:" in full_response:
+            response = full_response.split("응답:")[-1].strip()
+        else:
+            response = full_response.strip()
+
         if is_valid_response(response) and not mismatch_tone(input_text, response):
             memory.process_input(response)
             return response
 
+    # 3번 모두 실패 시 fallback
+    fallback_response = "죄송해요, 제대로 답변을 못했어요."
+    memory.process_input(fallback_response)
+    return fallback_response
+
 
 @app.get("/generate", response_class=PlainTextResponse)
 async def generate(request: Request):