Ultra-Stable: Defer ALL heavy imports to function call to prevent boot crash (500 error)

app.py

@@ -1,3 +1,4 @@
+# 1. 최상단에서는 오직 가벼운 라이브러리만 로드
 import huggingface_hub
 if not hasattr(huggingface_hub, "HfFolder"):
     class MockHfFolder:
@@ -10,22 +11,31 @@ if not hasattr(huggingface_hub, "HfFolder"):
     huggingface_hub.HfFolder = MockHfFolder
 
 import gradio as gr
-import pandas as pd
-import numpy as np
 import os
 import time
 
-# 지연 로딩을 위해 전역 변수로 선언
-_consultant = None
+# 전역 상태 관리 (지연 로딩 용)
+_models = {
+    "predictor": None,
+    "consultant": None
+}
 
-def get_consultant():
-    global _consultant
-    if _consultant is None:
+def load_all_models():
+    """사용자가 버튼을 처음 누를 때만 무거운 모델들을 로드합니다."""
+    global _models
+    if _models["predictor"] is None:
+        print("Loading heavy models...")
+        # 함수 내부에서 임포트하여 부팅 시 부하 방지
+        from predictors.score_prediction import CreditPredictor
         from langchain_google_genai import ChatGoogleGenerativeAI
         from langchain_huggingface import HuggingFaceEmbeddings
         from langchain_community.vectorstores import FAISS
         from config import EMBEDDING_MODEL, FAISS_PATH, RETRIEVER_K, GEMINI_API_KEY
         
+        # 1. 예측기 로드
+        _models["predictor"] = CreditPredictor()
+        
+        # 2. RAG 상담원 로드
         class Consultant:
             def __init__(self):
                 os.environ["GOOGLE_API_KEY"] = os.getenv("GOOGLE_API_KEY", GEMINI_API_KEY)
@@ -36,27 +46,16 @@ def get_consultant():
                 else:
                     self.retriever = None
                 self.llm = ChatGoogleGenerativeAI(model="gemini-1.5-flash", temperature=0.7)
-        _consultant = Consultant()
-    return _consultant
-
-# 예측기 로드
-from predictors.score_prediction import predictor
+        
+        _models["consultant"] = Consultant()
+        print("Models loaded successfully!")
 
 FEATURES_MAP = {
-    'C1Z001386': '1년내 카드 총 이용금액 (만원)',
-    'C1M210000': '신용카드 건수',
-    'C18210000': '체크카드 건수',
-    'C1L120001': '카드 총 한도금액 (만원)',
-    'C1L120004': '카드 개설일수',
-    'L10210000': '은행업종 대출 건수',
-    'L90210100': '대부업종 대출 건수',
-    'L90210200': '저축은행 대출 건수',
-    'L10210B00': '주택담보 대출 건수',
-    'L10216000': '신용대출 총 잔액 (만원)',
-    'L10217000': '담보대출 총 잔액 (만원)',
-    'D10110000': '연체 건수',
-    'D10133000': '연체 잔액 (만원)',
-    'PERF1': '90일 연체 여부'
+    'C1Z001386': '1년내 카드 총 이용금액 (만원)', 'C1M210000': '신용카드 건수', 'C18210000': '체크카드 건수',
+    'C1L120001': '카드 총 한도금액 (만원)', 'C1L120004': '카드 개설일수', 'L10210000': '은행업종 대출 건수',
+    'L90210100': '대부업종 대출 건수', 'L90210200': '저축은행 대출 건수', 'L10210B00': '주택담보 대출 건수',
+    'L10216000': '신용대출 총 잔액 (만원)', 'L10217000': '담보대출 총 잔액 (만원)', 'D10110000': '연체 건수',
+    'D10133000': '연체 잔액 (만원)', 'PERF1': '90일 연체 여부'
 }
 
 ALL_FEATURES_KEYS = [
@@ -65,102 +64,81 @@ ALL_FEATURES_KEYS = [
     'L10217000', 'D10110000', 'D10133000', 'PERF1'
 ]
 
+def handle_predict(*args):
+    try:
+        load_all_models()
+        features_dict = {}
+        for i, key in enumerate(ALL_FEATURES_KEYS):
+            if key == 'PERF1': features_dict[key] = int(args[i])
+            else:
+                val = str(args[i]).strip() if args[i] else "0"
+                features_dict[key] = float(val) if val else 0.0
+        
+        score = _models["predictor"].predict(features_dict)
+        display_score = int(round(score)) if score is not None else 0
+        return {"features": features_dict, "score": display_score}, display_score
+    except Exception as e:
+        return {"error": str(e)}, 0
+
 def generate_response(history, user_message, analysis_report):
     if not user_message: yield history, ""; return
-    if history and isinstance(history[-1], tuple) and history[-1][1] == "생각 중...": history = history[:-1]
-    
-    # 챗 상태 초기화
-    history = history + [[user_message, ""]]
-    t0 = time.time()
+    history = history + [[user_message, "⚡ **시스템 준비 중... (최대 30초 소요)**"]]
+    yield history, ""
     
     try:
-        # 지연 로딩 실행
-        history[-1][1] = "⚡ **시스템 초기화 중...**"
-        yield history, ""
-        cons = get_consultant()
+        load_all_models()
+        cons = _models["consultant"]
         
-        # R
-        history[-1][1] = "📂 **[R] 분석 문서 검색 중...**"
+        # R (Retrieval)
+        history[-1][1] = "📂 **분석 문서 검색 중...**"
         yield history, ""
         docs = cons.retriever.invoke(user_message) if cons.retriever else []
-        t_retrieve = time.time() - t0
         context = "\n\n".join([doc.page_content for doc in docs])
         
-        # A
+        # A (Augmentation)
         from llm.prompt import QA_PROMPT
         if analysis_report:
             score_val = int(analysis_report["score"])
             features_text = "\n".join([f"- {FEATURES_MAP.get(k, k)}: {v}" for k, v in analysis_report['features'].items()])
-            query_text = f"■ 고객 분석 결과: {score_val}점\n{features_text}\n\n■ 질문: {user_message}"
+            query_text = f"■ 고객 정보: {score_val}점\n{features_text}\n\n■ 질문: {user_message}"
         else:
             query_text = f"■ 질문: {user_message}"
 
-        prompt_text = QA_PROMPT.format(context=context, query=query_text)
-        t_augment = time.time() - t0
-
-        # G
+        # G (Generation)
+        history[-1][1] = "💬 **답변 생성 중...**"
+        yield history, ""
+        
         answer_buffer = ""
-        for chunk in cons.llm.stream(prompt_text):
+        for chunk in cons.llm.stream(QA_PROMPT.format(context=context, query=query_text)):
             answer_buffer += chunk.content
-            history[-1][1] = (
-                f"📂 **[R] 완료** ({t_retrieve:.1f}s) | 🔗 **[A] 완료** ({t_augment-t_retrieve:.2f}s)\n\n"
-                f"{answer_buffer}"
-            )
+            history[-1][1] = answer_buffer
             yield history, ""
             
     except Exception as e:
-        history[-1][1] = f"⚠️ 서버 오류가 발생했습니다: {str(e)}"
+        history[-1][1] = f"⚠️ 오류 발생: {str(e)}"
         yield history, ""
 
-def handle_predict(*args):
-    try:
-        features_dict = {}
-        for i, key in enumerate(ALL_FEATURES_KEYS):
-            if key == 'PERF1': features_dict[key] = int(args[i])
-            else:
-                val = str(args[i]).strip() if args[i] else "0"
-                features_dict[key] = float(val) if val else 0.0
-        
-        score = predictor.predict(features_dict)
-        display_score = int(min(max(round(score), 0), 1000)) if not np.isnan(score) else 0
-        return {"features": features_dict, "score": display_score}, display_score
-    except Exception as e:
-        return {"error": str(e)}, 0
-
 with gr.Blocks(title="KCB AI Consultant") as demo:
     analysis_report = gr.State(None)
-    
-    gr.Markdown("# 🛡️ KCB AI 신용 점수 분석 시스템 (LTS)")
+    gr.Markdown("# 🛡️ KCB AI 신용 상담 시스템 (Ultra-Stable)")
     
     with gr.Row():
         with gr.Column(scale=1):
-            input_list = []
-            for key in ALL_FEATURES_KEYS:
-                if key == 'PERF1':
-                    input_list.append(gr.Checkbox(label=FEATURES_MAP[key], value=False))
-                else:
-                    input_list.append(gr.Textbox(label=FEATURES_MAP[key], placeholder="0"))
+            input_list = [gr.Checkbox(label=FEATURES_MAP[k]) if k == 'PERF1' else gr.Textbox(label=FEATURES_MAP[k], placeholder="0") for k in ALL_FEATURES_KEYS]
             predict_btn = gr.Button("📈 점수 분석하기", variant="primary")
             
         with gr.Column(scale=2):
             result_display = gr.Label(label="예측 신용 점수")
             chatbot = gr.Chatbot(label="실시간 상담", height=500)
-            
             with gr.Row():
                 msg = gr.Textbox(placeholder="질문을 입력하세요...", scale=8, container=False)
                 submit_btn = gr.Button("상담하기", variant="primary", scale=1)
 
     predict_btn.click(handle_predict, inputs=input_list, outputs=[analysis_report, result_display])
     
-    def process_chat(message, history):
-        return "", history + [[message, "생각 중..."]]
-
-    msg.submit(process_chat, [msg, chatbot], [msg, chatbot]).then(
-        generate_response, [chatbot, msg, analysis_report], [chatbot, msg]
-    )
-    submit_btn.click(process_chat, [msg, chatbot], [msg, chatbot]).then(
-        generate_response, [chatbot, msg, analysis_report], [chatbot, msg]
-    )
+    def clear_msg(m, h): return "", h + [[m, "생각 중..."]]
+    msg.submit(clear_msg, [msg, chatbot], [msg, chatbot]).then(generate_response, [chatbot, msg, analysis_report], [chatbot, msg])
+    submit_btn.click(clear_msg, [msg, chatbot], [msg, chatbot]).then(generate_response, [chatbot, msg, analysis_report], [chatbot, msg])
 
 if __name__ == "__main__":
     demo.launch()