Stabilize: Downgrade to Gradio 4.44.1 and implement Lazy Loading to fix 500 error

README.md

@@ -4,11 +4,11 @@ emoji: 🛡️
 colorFrom: blue
 colorTo: indigo
 sdk: gradio
-sdk_version: 5.4.0
+sdk_version: 4.44.1
 python_version: "3.10"
 app_file: app.py
 pinned: false
 ---
 
-# 🛡️ KCB AI 신용 점수 분석 서비스
-KCB 가이드라인 기반의 신용 상담 AI입니다.
+# 🛡️ KCB AI 신용 점수 상담 서비스
+가장 안정적인 최신 배포 버전입니다.

app.py

@@ -13,21 +13,35 @@ import gradio as gr
 import pandas as pd
 import numpy as np
 import os
-import json
 import time
 
-from langchain_google_genai import ChatGoogleGenerativeAI
-from langchain_huggingface import HuggingFaceEmbeddings
-from langchain_community.vectorstores import FAISS
+# 지연 로딩을 위해 전역 변수로 선언
+_consultant = None
 
-from config import EMBEDDING_MODEL, FAISS_PATH, RETRIEVER_K, GEMINI_API_KEY
-from llm.prompt import QA_PROMPT
-from predictors.score_prediction import predictor
+def get_consultant():
+    global _consultant
+    if _consultant is None:
+        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
+        
+        class Consultant:
+            def __init__(self):
+                os.environ["GOOGLE_API_KEY"] = os.getenv("GOOGLE_API_KEY", GEMINI_API_KEY)
+                self.embedding_model = HuggingFaceEmbeddings(model_name=EMBEDDING_MODEL)
+                if os.path.exists(FAISS_PATH):
+                    self.vectorstore = FAISS.load_local(FAISS_PATH, self.embedding_model, allow_dangerous_deserialization=True)
+                    self.retriever = self.vectorstore.as_retriever(search_kwargs={"k": RETRIEVER_K})
+                else:
+                    self.retriever = None
+                self.llm = ChatGoogleGenerativeAI(model="gemini-1.5-flash", temperature=0.7)
+        _consultant = Consultant()
+    return _consultant
 
-# API 키 설정
-os.environ["GOOGLE_API_KEY"] = os.getenv("GOOGLE_API_KEY", GEMINI_API_KEY)
+# 예측기 로드
+from predictors.score_prediction import predictor
 
-# 14개 피처에 대한 상세 설명 매핑
 FEATURES_MAP = {
     'C1Z001386': '1년내 카드 총 이용금액 (만원)',
     'C1M210000': '신용카드 건수',
@@ -45,37 +59,35 @@ FEATURES_MAP = {
     'PERF1': '90일 연체 여부'
 }
 
-# 모델이 요구하는 정확한 순서
 ALL_FEATURES_KEYS = [
     'C1Z001386', 'C1M210000', 'C18210000', 'C1L120001', 'C1L120004', 
     'L10210000', 'L90210100', 'L90210200', 'L10210B00', 'L10216000', 
     'L10217000', 'D10110000', 'D10133000', 'PERF1'
 ]
 
-class CreditRAGConsultant:
-    def __init__(self):
-        self.embedding_model = HuggingFaceEmbeddings(model_name=EMBEDDING_MODEL)
-        if os.path.exists(FAISS_PATH):
-            self.vectorstore = FAISS.load_local(FAISS_PATH, self.embedding_model, allow_dangerous_deserialization=True)
-            self.retriever = self.vectorstore.as_retriever(search_kwargs={"k": RETRIEVER_K})
-        else:
-            self.vectorstore = self.retriever = None
-
-        self.llm = ChatGoogleGenerativeAI(model="gemini-1.5-flash", temperature=0.7)
-
-    def generate_response(self, history, user_message, analysis_report):
-        if not user_message: yield history, ""; return
-        if history and history[-1].get("content") == "생각 중...": history = history[:-1]
-        
-        history = history + [{"role": "assistant", "content": ""}]
-        t0 = time.time()
+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()
+    
+    try:
+        # 지연 로딩 실행
+        history[-1][1] = "⚡ **시스템 초기화 중...**"
+        yield history, ""
+        cons = get_consultant()
         
         # R
-        docs = self.retriever.invoke(user_message) if self.retriever else []
+        history[-1][1] = "📂 **[R] 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
+        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()])
@@ -88,29 +100,40 @@ class CreditRAGConsultant:
 
         # G
         answer_buffer = ""
-        try:
-            for chunk in self.llm.stream(prompt_text):
-                answer_buffer += chunk.content
-                history[-1]["content"] = (
-                    f"📂 **[R] 완료** ({t_retrieve:.1f}s) | 🔗 **[A] 완료** ({t_augment-t_retrieve:.2f}s)\n\n"
-                    f"{answer_buffer}"
-                )
-                yield history, ""
-        except Exception as e:
-            history[-1]["content"] = f"⚠️ 에러: {str(e)}"
+        for chunk in cons.llm.stream(prompt_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}"
+            )
             yield history, ""
+            
+    except Exception as e:
+        history[-1][1] = f"⚠️ 서버 오류가 발생했습니다: {str(e)}"
+        yield history, ""
 
-consultant = CreditRAGConsultant()
+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 Credit AI", theme=gr.themes.Soft()) as demo:
+with gr.Blocks(title="KCB AI Consultant") as demo:
     analysis_report = gr.State(None)
-    pending_msg = gr.State("")
-
-    gr.Markdown("# 🛡️ KCB AI 신용 점수 분석 시스템")
+    
+    gr.Markdown("# 🛡️ KCB AI 신용 점수 분석 시스템 (LTS)")
     
     with gr.Row():
         with gr.Column(scale=1):
-            gr.Markdown("### 📊 금융 지표 입력")
             input_list = []
             for key in ALL_FEATURES_KEYS:
                 if key == 'PERF1':
@@ -121,37 +144,22 @@ with gr.Blocks(title="KCB Credit AI", theme=gr.themes.Soft()) as demo:
             
         with gr.Column(scale=2):
             result_display = gr.Label(label="예측 신용 점수")
-            chatbot = gr.Chatbot(label="실시간 상담", height=550, type="messages")
+            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)
 
-    def handle_predict(*args):
-        features_dict = {}
-        # 입력 리스트와 키 순서가 동일함
-        for i, key in enumerate(ALL_FEATURES_KEYS):
-            if key == 'PERF1':
-                features_dict[key] = int(args[i])
-            else:
-                val = args[i] if args[i] else "0"
-                features_dict[key] = float(val)
-        
-        score = predictor.predict(features_dict)
-        display_score = int(min(max(round(score), 0), 1000))
-        return {"features": features_dict, "score": display_score}, display_score
-
-    def user_msg(user_message, history):
-        if not user_message: return history, "", ""
-        return history + [{"role": "user", "content": user_message}, {"role": "assistant", "content": "생각 중..."}], "", user_message
-
     predict_btn.click(handle_predict, inputs=input_list, outputs=[analysis_report, result_display])
     
-    msg.submit(user_msg, [msg, chatbot], [chatbot, msg, pending_msg]).then(
-        consultant.generate_response, [chatbot, pending_msg, analysis_report], [chatbot, msg]
+    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(user_msg, [msg, chatbot], [chatbot, msg, pending_msg]).then(
-        consultant.generate_response, [chatbot, pending_msg, analysis_report], [chatbot, msg]
+    submit_btn.click(process_chat, [msg, chatbot], [msg, chatbot]).then(
+        generate_response, [chatbot, msg, analysis_report], [chatbot, msg]
     )
 
 if __name__ == "__main__":

requirements.txt

@@ -1,4 +1,4 @@
-gradio==5.4.0
+gradio==4.44.1
 pandas
 numpy<2.0.0
 tensorflow-cpu==2.15.0