backend/src/logic.py

# logic.py
import streamlit as st
import pandas as pd
import numpy as np
from gensim.models import Word2Vec, Doc2Vec
from ast import literal_eval
import pickle
from datetime import datetime, timedelta, timezone
from supabase import create_client
import re
import os
from collections import Counter

# ==========================================
# 0. 환경 설정 및 규칙 정의
# ==========================================
PRICE_KEYWORD_RULES = [
    (5, ['소고기', '한우', '채끝', '등심', '안심', '갈비살', '전복', '장어']),
    (4, ['돼지', '삼겹', '목살', '앞다리', '뒷다리', '갈비', '오리', '낙지', '오징어', '새우', '명란']),
    (3, ['닭', '치킨', '햄', '소시지', '베이컨', '스팸', '참치', '동원', '어묵', '맛살', '버섯', '치즈']),
    (2, ['두부', '순두부', '콩나물', '숙주', '김치', '무', '감자', '고구마', '당근', '호박']),
    (1, ['양파', '대파', '쪽파', '실파', '마늘', '고추', '물', '소금', '설탕', '간장', '소스', '양념', '육수'])
]
PRICE_RULE_EXCEPTIONS = ['돼지감자', '닭의장풀', '새우젓', '멸치액젓', '다시다']

# ==========================================
# 1. Supabase DB 연동 및 데이터 저장/로드
# ==========================================
@st.cache_resource
def init_supabase():
    try:
        url = None
        key = None
        
        # 1. Try Streamlit secrets
        try:
            if hasattr(st, "secrets") and "supabase" in st.secrets:
                url = st.secrets["supabase"]["url"]
                key = st.secrets["supabase"]["key"]
        except:
            pass
            
        # 2. Add fallback to OS environment variables
        if not url or not key:
            url = os.environ.get("SUPABASE_URL")
            key = os.environ.get("SUPABASE_KEY")
            
        if not url or not key:
            # 로컬 개발 중 secrets 없이 실행될 경우를 대비해 스킵하거나 에러 처리
            # API 서버에서는 필수이므로 로그 남김
            return None

        return create_client(url, key)
    except Exception as e:
        print(f"Supabase 연결 경고: {e}")
        return None

def get_kst_now_iso():
    kst_timezone = timezone(timedelta(hours=9))
    now_kst = datetime.now(kst_timezone)
    return now_kst.isoformat()

@st.cache_data(ttl=300)
def load_global_stopwords():
    try:
        supabase = init_supabase()
        response = supabase.table("stopwords").select("word").order("created_at", desc=True).execute()
        if response.data:
            return [item['word'] for item in response.data]
        return []
    except Exception as e:
        print(f"불용어 로드 실패: {e}")
        return []

@st.cache_data(ttl=600)
def get_usage_stats(timeframe='today'):
    try:
        supabase = init_supabase()
        query = supabase.table("usage_log").select("dish, target")

        if timeframe == 'today':
            kst = timezone(timedelta(hours=9))
            now_kst = datetime.now(kst)
            today_start = now_kst.replace(hour=0, minute=0, second=0, microsecond=0)
            tomorrow_start = today_start + timedelta(days=1)
            query = query.gte("created_at", today_start.isoformat()).lt("created_at", tomorrow_start.isoformat())
        
        response = query.execute()
        data = response.data
        
        count = len(data)
        top_dishes = pd.Series(dtype=int)
        top_targets = pd.Series(dtype=int)

        if count > 0:
            df_log = pd.DataFrame(data)
            df_log['clean_dish'] = df_log['dish'].astype(str).str.replace(r'\[Custom\]', '', regex=True).str.strip()
            top_dishes = df_log[df_log['clean_dish'] != '']['clean_dish'].value_counts().head(5)

            all_targets = []
            for t in df_log['target']:
                if t:
                    all_targets.extend([x.strip() for x in str(t).split(',') if x.strip()])
            top_targets = pd.Series(all_targets).value_counts().head(5)

        return count, top_dishes, top_targets

    except Exception as e:
        print(f"통계 데이터 로드 실패 ({timeframe}): {e}")
        return 0, pd.Series(dtype=int), pd.Series(dtype=int)

@st.cache_data(ttl=600)
def get_wordcloud_text(timeframe='today'):
    try:
        supabase = init_supabase()
        query = supabase.table("usage_log").select("target")
        if timeframe == 'today':
            kst = timezone(timedelta(hours=9))
            now_kst = datetime.now(kst)
            today_start = now_kst.replace(hour=0, minute=0, second=0, microsecond=0)
            tomorrow_start = today_start + timedelta(days=1)
            query = query.gte("created_at", today_start.isoformat()).lt("created_at", tomorrow_start.isoformat())
        
        response = query.execute()
        data = response.data
        all_targets = []
        if data:
            for item in data:
                if item['target']:
                    all_targets.extend([x.strip() for x in str(item['target']).split(',') if x.strip()])
        return " ".join(all_targets)
    except Exception as e:
        print(f"워드클라우드 데이터 로드 실패: {e}")
        return ""

def save_stopwords_to_db(words_string):
    words = [w.strip() for w in words_string.split(',') if w.strip()]
    if not words: return False, "저장할 단어가 없습니다."
    supabase = init_supabase()
    success_count, duplicate_count, fail_count = 0, 0, 0
    for word in words:
        try:
            supabase.table("stopwords").insert({"word": word}).execute()
            success_count += 1
        except Exception as e:
            if 'duplicate' in str(e).lower(): duplicate_count += 1
            else: fail_count += 1
    if success_count > 0: st.cache_data.clear()
    msg_parts = []
    if success_count > 0: msg_parts.append(f"✅ {success_count}개 저장")
    if duplicate_count > 0: msg_parts.append(f"⚠️ {duplicate_count}개 중복")
    if fail_count > 0: msg_parts.append(f"❌ {fail_count}개 실패")
    return success_count > 0, ", ".join(msg_parts)

@st.cache_data(ttl=60) 
def get_board_messages():
    try:
        supabase = init_supabase()
        response = supabase.table("board").select("*").order("created_at", desc=True).limit(50).execute()
        if response.data:
            for item in response.data:
                dt = datetime.fromisoformat(item['created_at'])
                dt_kst = dt + timedelta(hours=9) 
                item['display_time'] = dt_kst.strftime("%m/%d %H:%M")
            return response.data
        return []
    except Exception as e:
        print(f"게시판 로드 실패: {e}")
        return []

def save_board_message(nickname, content):
    if not nickname or not content: return False
    try:
        supabase = init_supabase()
        supabase.table("board").insert({"nickname": nickname, "content": content}).execute()
        st.cache_data.clear()
        return True
    except Exception as e:
        print(f"게시판 저장 실패: {e}")
        return False

def save_feedback_to_db(feedback_text):
    try:
        supabase = init_supabase()
        supabase.table("feedback").insert({"content": feedback_text, "created_at": get_kst_now_iso()}).execute()
        return True
    except Exception as e:
        print(f"피드백 저장 에러: {e}")
        return False

def save_log_to_db(dish, target, stops, w1, w2, w3, w4, rec_list=None, is_custom=False):
    try:
        supabase = init_supabase()
        r1 = rec_list[0] if rec_list and len(rec_list) > 0 else None
        r2 = rec_list[1] if rec_list and len(rec_list) > 1 else None
        r3 = rec_list[2] if rec_list and len(rec_list) > 2 else None
        dish_name_to_save = f"[Custom] {dish}" if is_custom else dish
        data = {
            "dish": dish_name_to_save, "target": target, "stops": ", ".join(stops) if stops else "없음",
            "w_w2v": w1, "w_d2v": w2, "w_method": w3, "w_cat": w4, "rec_1": r1, "rec_2": r2, "rec_3": r3,
            "created_at": get_kst_now_iso()
        }
        response = supabase.table("usage_log").insert(data).execute()
        if response.data: return response.data[0]['id']
        return None
    except Exception as e:
        print(f"로그 저장 에러: {e}")
        return None

def update_feedback_in_db(log_id, status):
    try:
        supabase = init_supabase()
        if log_id:
            supabase.table("usage_log").update({"satisfaction": status}).eq("id", log_id).execute()
            return True
        return False
    except Exception as e:
        print(f"만족도 업데이트 에러: {e}")
        return False

# ==========================================
# 2. 데이터 및 모델 로드
# ==========================================
# ==========================================
# 2. 데이터 및 모델 로드 (Lazy Loading 적용)
# ==========================================
w2v_model = None
d2v_model = None
df = None
stats = None
price_map = {}
global_stopwords_set = set()
all_ingredients_set = set()
method_map = {}
recipes_by_ingredient = {}
ing_method_counts = {}
ing_cat_counts = {}
total_method_counts = {}
total_cat_counts = {}
TOTAL_RECIPES = 0

def load_resources():
    global w2v_model, d2v_model, df, stats, price_map, global_stopwords_set, all_ingredients_set
    global method_map, recipes_by_ingredient, ing_method_counts, ing_cat_counts, total_method_counts, total_cat_counts, TOTAL_RECIPES
    
    print("Loading resources... (This may take a while)")
    
    # 기준 경로 설정 (현재 파일 위치 logic.py 기준 상위 폴더)
    base_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
    
    # mmap='r' 옵션으로 메모리 사용량 최소화 (디스크에서 직접 읽음)
    w2v_model = Word2Vec.load(os.path.join(base_dir, "models/w2v.model"), mmap='r')
    d2v_model = Doc2Vec.load(os.path.join(base_dir, "models/d2v.model"), mmap='r')
    df_temp = pd.read_csv(os.path.join(base_dir, "data/final_recipe_data.csv"))
    df_temp['재료토큰'] = df_temp['재료토큰'].apply(literal_eval)
    df = df_temp # Assign to global
    
    with open(os.path.join(base_dir, "data/stats.pkl"), "rb") as f:
        stats = pickle.load(f)
        
    try:
        price_df = pd.read_csv(os.path.join(base_dir, "data/price_rank.csv"), encoding='utf-8-sig')
        price_df.columns = price_df.columns.str.strip()
        price_map = dict(zip(price_df['ingredient'], price_df['rank']))
    except:
        price_map = {}
    
    global_stopwords_set = set() 
    all_ingredients_set = set()
    for ings in df['재료토큰']:
        all_ingredients_set.update(ings)
        
    # Stats unpacking
    method_map = stats["method_map"]
    recipes_by_ingredient = stats["recipes_by_ingredient"]
    ing_method_counts = stats["ing_method_counts"]
    ing_cat_counts = stats["ing_cat_counts"]
    total_method_counts = stats["total_method_counts"]
    total_cat_counts = stats["total_cat_counts"]
    TOTAL_RECIPES = stats["TOTAL_RECIPES"]
    print("Resources loaded successfully!")

def ensure_initialized():
    if df is None:
        load_resources()

# 기존 즉시 실행 코드 제거
# w2v_model, d2v_model, ... = load_resources()

# ==========================================
# 3. 핵심 계산 로직
# ==========================================
def cos_sim(vec_a, vec_b):
    norm = (np.linalg.norm(vec_a) * np.linalg.norm(vec_b) + 1e-9)
    return max(0.0, float(np.dot(vec_a, vec_b) / norm))

def get_stat_score(ingredient, target_key, ing_count_dict, total_count_dict, total_n, min_count=5):
    cnts = ing_count_dict.get(ingredient)
    if not cnts: return 0.0
    ing_target_count = cnts[target_key]
    ing_total_count = sum(cnts.values())
    if ing_total_count < min_count: return 0.0
    prob_ing_context = ing_target_count / ing_total_count
    baseline_prob = total_count_dict[target_key] / total_n
    if baseline_prob == 0: return 0.0
    return prob_ing_context / baseline_prob

def get_estimated_price_rank(ing_name, price_map):
    if ing_name in price_map: return price_map[ing_name]
    if any(exp in ing_name for exp in PRICE_RULE_EXCEPTIONS): return 3
    for rank, keywords in PRICE_KEYWORD_RULES:
        if any(kw in ing_name for kw in keywords): return rank
    return 3

# ==========================================
# 4. 대체 추천 알고리즘 (DB 기반)
# ==========================================
def substitute_single(recipe_id, target_ing, user_stopwords, w_w2v, w_d2v, w_method, w_cat, topn=10):
    row = df[df['레시피일련번호'] == recipe_id].iloc[0]
    current_method = row['요리방법별명']
    current_cat = row['요리종류별명_세분화']
    context_ings = row['재료토큰']
    tag = f"recipe_{recipe_id}"
    if target_ing not in w2v_model.wv: return pd.DataFrame()
    total_weight = w_w2v + w_d2v + w_method + w_cat
    if total_weight == 0: total_weight = 1.0
    vec_recipe = None
    if w_d2v > 0 and tag in d2v_model.dv: vec_recipe = d2v_model.dv[tag]
    target_rank = get_estimated_price_rank(target_ing, price_map)
    candidates_raw = w2v_model.wv.most_similar(target_ing, topn=50)
    temp_results = []
    seen_candidates = set()
    
    # [수정] 실시간 로드
    global_stopwords_set = set(load_global_stopwords())
    final_stopwords = set(user_stopwords) | global_stopwords_set

    for cand, score_w2v in candidates_raw:
        clean_cand = cand
        if final_stopwords:
            for stop in final_stopwords: clean_cand = clean_cand.replace(stop, "")
        clean_cand = clean_cand.strip()
        
        if not clean_cand: continue
        if clean_cand in final_stopwords: continue

        if clean_cand in context_ings: continue
        if clean_cand == target_ing: continue
        if clean_cand not in w2v_model.wv: continue
        if clean_cand in seen_candidates: continue
        seen_candidates.add(clean_cand)
        real_score_w2v = w2v_model.wv.similarity(target_ing, clean_cand)
        s_w2v = max(0.0, real_score_w2v)
        if s_w2v < 0.35: continue
        s_d2v = 0.0
        if w_d2v > 0 and vec_recipe is not None:
            rid_list = recipes_by_ingredient.get(clean_cand, [])
            same_method_ids = [r for r in rid_list if method_map.get(r) == current_method]
            if len(same_method_ids) > 20:
                np.random.seed(42)
                same_method_ids = np.random.choice(same_method_ids, 20, replace=False)
            if same_method_ids is not None and len(same_method_ids) > 0:
                sims = []
                for r in same_method_ids:
                    rt = f"recipe_{r}"
                    if rt in d2v_model.dv: sims.append(cos_sim(vec_recipe, d2v_model.dv[rt]))
                if sims: s_d2v = np.mean(sims)
        s_method = 0.0 if w_method <= 0 else get_stat_score(clean_cand, current_method, ing_method_counts, total_method_counts, TOTAL_RECIPES)
        s_cat = 0.0 if w_cat <= 0 else get_stat_score(clean_cand, current_cat, ing_cat_counts, total_cat_counts, TOTAL_RECIPES)
        cand_rank = get_estimated_price_rank(clean_cand, price_map)
        saving_score = target_rank - cand_rank
        temp_results.append({"대체재료": clean_cand, "raw_W2V": s_w2v, "raw_D2V": s_d2v, "raw_Method": s_method, "raw_Category": s_cat, "saving_score": saving_score})
    if not temp_results: return pd.DataFrame()
    df_res = pd.DataFrame(temp_results)
    cols = ["raw_W2V", "raw_D2V", "raw_Method", "raw_Category"]
    norm_cols = ["W2V", "D2V", "Method", "Category"]
    for raw_col, norm_col in zip(cols, norm_cols):
        min_val = df_res[raw_col].min()
        max_val = df_res[raw_col].max()
        if max_val - min_val == 0: df_res[norm_col] = 0.5
        else: df_res[norm_col] = (df_res[raw_col] - min_val) / (max_val - min_val)
    df_res["최종점수"] = ((df_res["W2V"]*w_w2v) + (df_res["D2V"]*w_d2v) + (df_res["Method"]*w_method) + (df_res["Category"]*w_cat)) / total_weight
    return df_res.sort_values("최종점수", ascending=False).head(topn).reset_index(drop=True)

def substitute_multi(recipe_id, targets, user_stopwords, w_w2v, w_d2v, w_method, w_cat, beam_width=3, result_topn=3):
    row = df[df['레시피일련번호'] == recipe_id].iloc[0]
    current_method = row['요리방법별명']
    current_cat = row['요리종류별명_세분화']
    initial_context = row['재료토큰']
    tag = f"recipe_{recipe_id}"
    vec_recipe = None
    if w_d2v > 0 and tag in d2v_model.dv: vec_recipe = d2v_model.dv[tag]
    total_weight = w_w2v + w_d2v + w_method + w_cat
    if total_weight == 0: total_weight = 1.0
    target_ranks_sum = 0
    for t in targets: target_ranks_sum += get_estimated_price_rank(t, price_map)
    
    # [수정] 실시간 로드
    global_stopwords_set = set(load_global_stopwords())
    final_stopwords = set(user_stopwords) | global_stopwords_set

    beam = [(0.0, [], initial_context)]
    for target_ing in targets:
        next_beam = []
        if target_ing not in w2v_model.wv:
            for score, subs, ctx in beam: next_beam.append((score, subs + [target_ing], ctx))
            beam = next_beam
            continue
        for path_score, path_subs, path_ctx in beam:
            current_ctx_ing = [x for x in path_ctx if x != target_ing]
            candidates = w2v_model.wv.most_similar(target_ing, topn=30)
            temp_candidates = []
            seen_candidates = set()
            for cand, _ in candidates:
                clean_cand = cand
                if final_stopwords:
                    for stop in final_stopwords: clean_cand = clean_cand.replace(stop, "")
                clean_cand = clean_cand.strip()

                if not clean_cand: continue
                if clean_cand in final_stopwords: continue

                if clean_cand in current_ctx_ing or clean_cand in path_subs: continue
                if clean_cand == target_ing: continue
                if clean_cand not in w2v_model.wv: continue
                if clean_cand in seen_candidates: continue
                seen_candidates.add(clean_cand)
                sim_orig = w2v_model.wv.similarity(target_ing, clean_cand)
                sim_orig = max(0.0, sim_orig)
                if sim_orig < 0.3: continue
                harmony_scores = [w2v_model.wv.similarity(clean_cand, c) for c in current_ctx_ing if c in w2v_model.wv]
                sim_harmony = np.mean(harmony_scores) if harmony_scores else 0.0
                s_w2v = 0.5 * sim_orig + 0.5 * max(0.0, sim_harmony)
                s_d2v = 0.0
                if vec_recipe is not None:
                    rid_list = recipes_by_ingredient.get(clean_cand, [])
                    same_method_ids = [r for r in rid_list if method_map.get(r) == current_method]
                    if len(same_method_ids) > 10:
                        np.random.seed(42)
                        same_method_ids = np.random.choice(same_method_ids, 10, replace=False)
                    if same_method_ids is not None and len(same_method_ids) > 0:
                        sims = []
                        for r in same_method_ids:
                            rt = f"recipe_{r}"
                            if rt in d2v_model.dv: sims.append(cos_sim(vec_recipe, d2v_model.dv[rt]))
                        if sims: s_d2v = np.mean(sims)
                s_method = 0.0 if w_method <= 0 else get_stat_score(clean_cand, current_method, ing_method_counts, total_method_counts, TOTAL_RECIPES)
                s_cat = 0.0 if w_cat <= 0 else get_stat_score(clean_cand, current_cat, ing_cat_counts, total_cat_counts, TOTAL_RECIPES)
                temp_candidates.append({"cand": clean_cand, "raw_w2v": s_w2v, "raw_d2v": s_d2v, "raw_method": s_method, "raw_cat": s_cat})
            if not temp_candidates: continue
            df_temp = pd.DataFrame(temp_candidates)
            cols = ["raw_w2v", "raw_d2v", "raw_method", "raw_cat"]
            for col in cols:
                min_val = df_temp[col].min()
                max_val = df_temp[col].max()
                if max_val - min_val == 0: df_temp[col + "_norm"] = 0.5
                else: df_temp[col + "_norm"] = (df_temp[col] - min_val) / (max_val - min_val)
            for _, r in df_temp.iterrows():
                weighted_sum = ((r["raw_w2v_norm"]*w_w2v) + (r["raw_d2v_norm"]*w_d2v) + (r["raw_method_norm"]*w_method) + (r["raw_cat_norm"]*w_cat)) / total_weight
                new_total_score = path_score + weighted_sum
                new_subs = path_subs + [r["cand"]]
                new_ctx = current_ctx_ing + [r["cand"]]
                next_beam.append((new_total_score, new_subs, new_ctx))
        next_beam.sort(key=lambda x: x[0], reverse=True)
        beam = next_beam[:beam_width]
    final_results = []
    for score, subs, _ in beam:
        avg_score = score / len(targets) if targets else 0.0
        cand_ranks_sum = 0
        for sub_ing in subs: cand_ranks_sum += get_estimated_price_rank(sub_ing, price_map)
        total_saving_score = target_ranks_sum - cand_ranks_sum
        final_results.append((subs, avg_score, total_saving_score))
    return final_results[:result_topn]

# ==========================================
# 5. 커스텀 입력 기반 대체 알고리즘 (수정됨)
# ==========================================
def substitute_single_custom(target_ing, context_ings_list, user_stopwords, w_w2v, w_d2v, excluded_ings=None, topn=10):
    if target_ing not in w2v_model.wv: return pd.DataFrame()
    total_weight = w_w2v + w_d2v
    if total_weight == 0: total_weight = 1.0
    vec_custom_context = None
    if w_d2v > 0:
        valid_context = [word for word in context_ings_list if word in d2v_model.wv]
        if valid_context: vec_custom_context = d2v_model.infer_vector(valid_context)
    target_rank = get_estimated_price_rank(target_ing, price_map)
    candidates_raw = w2v_model.wv.most_similar(target_ing, topn=50)
    temp_results = []
    seen_candidates = set()

    # [수정] 실시간 로드
    global_stopwords_set = set(load_global_stopwords())
    final_stopwords = set(user_stopwords) | global_stopwords_set
    excluded_set = set(excluded_ings) if excluded_ings else set()

    for cand, score_w2v in candidates_raw:
        clean_cand = cand
        if final_stopwords:
            for stop in final_stopwords: clean_cand = clean_cand.replace(stop, "")
        clean_cand = clean_cand.strip()

        if not clean_cand: continue
        if clean_cand in final_stopwords: continue
        if clean_cand in excluded_set: continue

        if clean_cand in context_ings_list: continue
        if clean_cand == target_ing: continue
        if clean_cand not in w2v_model.wv: continue
        if clean_cand in seen_candidates: continue
        seen_candidates.add(clean_cand)
        real_score_w2v = w2v_model.wv.similarity(target_ing, clean_cand)
        s_w2v = max(0.0, real_score_w2v)
        if s_w2v < 0.35: continue
        s_d2v = 0.0
        if w_d2v > 0 and vec_custom_context is not None:
            rid_list = recipes_by_ingredient.get(clean_cand, [])
            if len(rid_list) > 20:
                np.random.seed(42)
                rid_list = np.random.choice(rid_list, 20, replace=False)
            if rid_list is not None and len(rid_list) > 0:
                sims = []
                for r in rid_list:
                    rt = f"recipe_{r}"
                    if rt in d2v_model.dv: sims.append(cos_sim(vec_custom_context, d2v_model.dv[rt]))
                if sims: s_d2v = np.mean(sims)
        s_method, s_cat = 0.0, 0.0
        cand_rank = get_estimated_price_rank(clean_cand, price_map)
        saving_score = target_rank - cand_rank
        temp_results.append({"대체재료": clean_cand, "raw_W2V": s_w2v, "raw_D2V": s_d2v, "raw_Method": s_method, "raw_Category": s_cat, "saving_score": saving_score})
    if not temp_results: return pd.DataFrame()
    df_res = pd.DataFrame(temp_results)
    cols = ["raw_W2V", "raw_D2V"]
    norm_cols = ["W2V", "D2V"]
    for raw_col, norm_col in zip(cols, norm_cols):
        min_val = df_res[raw_col].min()
        max_val = df_res[raw_col].max()
        if max_val - min_val == 0: df_res[norm_col] = 0.5
        else: df_res[norm_col] = (df_res[raw_col] - min_val) / (max_val - min_val)
    df_res["최종점수"] = ((df_res["W2V"]*w_w2v) + (df_res["D2V"]*w_d2v)) / total_weight
    return df_res.sort_values("최종점수", ascending=False).head(topn).reset_index(drop=True)

def substitute_multi_custom(targets, context_ings_list, user_stopwords, w_w2v, w_d2v, excluded_ings=None, beam_width=3, result_topn=3):
    total_weight = w_w2v + w_d2v
    if total_weight == 0: total_weight = 1.0
    vec_custom_context = None
    if w_d2v > 0:
        valid_context = [word for word in context_ings_list if word in d2v_model.wv]
        if valid_context: vec_custom_context = d2v_model.infer_vector(valid_context)
    target_ranks_sum = 0
    for t in targets: target_ranks_sum += get_estimated_price_rank(t, price_map)
    
    # [수정] 실시간 로드
    global_stopwords_set = set(load_global_stopwords())
    final_stopwords = set(user_stopwords) | global_stopwords_set
    excluded_set = set(excluded_ings) if excluded_ings else set()

    beam = [(0.0, [], context_ings_list)]
    for target_ing in targets:
        next_beam = []
        if target_ing not in w2v_model.wv:
            for score, subs, ctx in beam: next_beam.append((score, subs + [target_ing], ctx))
            beam = next_beam
            continue
        for path_score, path_subs, path_ctx in beam:
            current_ctx_ing = [x for x in path_ctx if x != target_ing]
            candidates = w2v_model.wv.most_similar(target_ing, topn=30)
            temp_candidates = []
            seen_candidates = set()
            for cand, _ in candidates:
                clean_cand = cand
                if final_stopwords:
                    for stop in final_stopwords: clean_cand = clean_cand.replace(stop, "")
                clean_cand = clean_cand.strip()

                if not clean_cand: continue
                if clean_cand in final_stopwords: continue
                if clean_cand in excluded_set: continue

                if clean_cand in current_ctx_ing or clean_cand in path_subs: continue
                if clean_cand == target_ing: continue
                if clean_cand not in w2v_model.wv: continue
                if clean_cand in seen_candidates: continue
                seen_candidates.add(clean_cand)
                sim_orig = w2v_model.wv.similarity(target_ing, clean_cand)
                sim_orig = max(0.0, sim_orig)
                if sim_orig < 0.3: continue
                harmony_scores = [w2v_model.wv.similarity(clean_cand, c) for c in current_ctx_ing if c in w2v_model.wv]
                sim_harmony = np.mean(harmony_scores) if harmony_scores else 0.0
                s_w2v = 0.5 * sim_orig + 0.5 * max(0.0, sim_harmony)
                s_d2v = 0.0
                if w_d2v > 0:
                    valid_path_ctx = [word for word in current_ctx_ing if word in d2v_model.wv]
                    if valid_path_ctx:
                        vec_path_context = d2v_model.infer_vector(valid_path_ctx)
                        rid_list = recipes_by_ingredient.get(clean_cand, [])
                        if len(rid_list) > 10:
                            np.random.seed(42)
                            rid_list = np.random.choice(rid_list, 10, replace=False)
                        if rid_list is not None and len(rid_list) > 0:
                            sims = []
                            for r in rid_list:
                                rt = f"recipe_{r}"
                                if rt in d2v_model.dv: sims.append(cos_sim(vec_path_context, d2v_model.dv[rt]))
                            if sims: s_d2v = np.mean(sims)
                s_method, s_cat = 0.0, 0.0
                temp_candidates.append({"cand": clean_cand, "raw_w2v": s_w2v, "raw_d2v": s_d2v})
            if not temp_candidates: continue
            df_temp = pd.DataFrame(temp_candidates)
            cols = ["raw_w2v", "raw_d2v"]
            for col in cols:
                min_val = df_temp[col].min()
                max_val = df_temp[col].max()
                if max_val - min_val == 0: df_temp[col + "_norm"] = 0.5
                else: df_temp[col + "_norm"] = (df_temp[col] - min_val) / (max_val - min_val)
            for _, r in df_temp.iterrows():
                weighted_sum = ((r["raw_w2v_norm"]*w_w2v) + (r["raw_d2v_norm"]*w_d2v)) / total_weight
                new_total_score = path_score + weighted_sum
                new_subs = path_subs + [r["cand"]]
                new_ctx = current_ctx_ing + [r["cand"]]
                next_beam.append((new_total_score, new_subs, new_ctx))
        next_beam.sort(key=lambda x: x[0], reverse=True)
        beam = next_beam[:beam_width]
    final_results = []
    for score, subs, _ in beam:
        avg_score = score / len(targets) if targets else 0.0
        cand_ranks_sum = 0
        for sub_ing in subs: cand_ranks_sum += get_estimated_price_rank(sub_ing, price_map)
        total_saving_score = target_ranks_sum - cand_ranks_sum
        final_results.append((subs, avg_score, total_saving_score))
    return final_results[:result_topn]

# ==========================================
# 6. 재료 키워드 기반 레시피 검색 (기존과 동일)
# ==========================================
def find_recipes_by_ingredient_keyword(keyword, topn=5):
    keyword = keyword.strip()
    if not keyword: return []
    matched_dishes = set()
    for _, row in df.iterrows():
        for ing in row['재료토큰']:
            if keyword in ing:
                matched_dishes.add(row['요리명'])
                break 
    return list(matched_dishes)[:topn]