Initial commit

app.py

@@ -4,11 +4,10 @@ from fastapi.middleware.cors import CORSMiddleware
 from manager.dialogue_manager import handle_dialogue
 from rag.rag_manager import chroma_initialized, load_game_docs_from_disk, add_docs, set_embedder
 from contextlib import asynccontextmanager
-from models.model_loader import load_emotion_model, load_fallback_model, load_embedder
+from models.model_loader import load_fallback_model, load_embedder
 from schemas import AskReq, AskRes
 from pathlib import Path
 from config import (
-    EMOTION_MODEL_NAME, EMOTION_MODEL_DIR,
     FALLBACK_MODEL_NAME, FALLBACK_MODEL_DIR,
     EMBEDDER_MODEL_NAME, EMBEDDER_MODEL_DIR,
     HF_TOKEN, BASE_DIR
@@ -17,10 +16,7 @@ from config import (
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
-    # Emotion
-    emo_tokenizer, emo_model = load_emotion_model(EMOTION_MODEL_NAME, EMOTION_MODEL_DIR, token=HF_TOKEN)
-    app.state.emotion_tokenizer = emo_tokenizer
-    app.state.emotion_model = emo_model
+    print("🚀 서버 시작 중... 모델 로딩 중...")
 
     # Fallback
     fb_tokenizer, fb_model = load_fallback_model(FALLBACK_MODEL_NAME, FALLBACK_MODEL_DIR, token=HF_TOKEN)

config.py

@@ -14,12 +14,10 @@ HF_TIMEOUT = float(os.getenv("HF_TIMEOUT", "25"))
 
 
 # 모델 이름
-EMOTION_MODEL_NAME = os.getenv("EMOTION_MODEL_NAME", "tae898/emoberta-base-ko")
 FALLBACK_MODEL_NAME = os.getenv("FALLBACK_MODEL_NAME", "skt/ko-gpt-trinity-1.2B-v0.5")
 EMBEDDER_MODEL_NAME = os.getenv("EMBEDDER_MODEL_NAME", "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")
 
 # 모델 디렉토리
-EMOTION_MODEL_DIR = Path(os.getenv("EMOTION_MODEL_DIR", BASE_DIR / "models" / "emotion-classification-model"))
 FALLBACK_MODEL_DIR = Path(os.getenv("FALLBACK_MODEL_DIR", BASE_DIR / "models" / "fallback-npc-model"))
 EMBEDDER_MODEL_DIR = Path(os.getenv("EMBEDDER_MODEL_DIR", BASE_DIR / "models" / "sentence-embedder"))
 

manager/dialogue_manager.py

@@ -3,7 +3,7 @@ from pipeline.preprocess import preprocess_input
 from pipeline.generator import generate_response
 from pipeline.postprocess import postprocess_fallback, postprocess_main
 from models.fallback_model import generate_fallback_response
-from .prompt_builder import build_main_prompt, build_fallback_prompt  # 수정된 prompt 빌더 사용
+from .prompt_builder import build_main_prompt, build_fallback_prompt
 
 async def handle_dialogue(
     request: Request,

models/model_loader.py

@@ -1,24 +1,11 @@
 from pathlib import Path
 from transformers import (
     AutoTokenizer,
-    AutoModelForSequenceClassification,
     AutoModelForCausalLM
 )
 from sentence_transformers import SentenceTransformer
 
 
-def load_emotion_model(model_name: str, model_dir: Path, token: str = None):
-    if not (model_dir / "config.json").exists():
-        tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True, use_auth_token=token)
-        model = AutoModelForSequenceClassification.from_pretrained(model_name, trust_remote_code=True, use_auth_token=token)
-        tokenizer.save_pretrained(model_dir)
-        model.save_pretrained(model_dir)
-
-    tokenizer = AutoTokenizer.from_pretrained(str(model_dir), trust_remote_code=True, local_files_only=True)
-    model = AutoModelForSequenceClassification.from_pretrained(str(model_dir), trust_remote_code=True, local_files_only=True)
-    return tokenizer, model
-
-
 def load_fallback_model(model_name: str, model_dir: Path, token: str = None):
     if not (model_dir / "config.json").exists():
         tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True, use_auth_token=token)

pipeline/preprocess.py

@@ -1,7 +1,6 @@
 import json, torch
 from fastapi import Request
 from manager.agent_manager import agent_manager
-from models.emotion_model import detect_emotion
 from models.fallback_model import generate_fallback_response
 from utils.context_parser import ContextParser
 from sentence_transformers import util