scripts/42_soft_grid.py

"""카테고리 소프트 부스트 가중치 그리드 — 엔진 1회 로드, 여러 weight 측정."""
import sys, os, json
from pathlib import Path
sys.path.insert(0, str(Path(__file__).parent.parent))
sys.stdout.reconfigure(encoding="utf-8")
os.environ.setdefault("LLM_BACKEND", "openai")
import warnings; warnings.filterwarnings("ignore")
import logging; logging.disable(logging.CRITICAL)

import core.search_engine as se
from core.agent import run_agent

BIX = Path(__file__).parent.parent.parent / "BIX_메뉴검색_20260624" / "BIX_회귀_통합.json"
cases = json.loads(BIX.read_text(encoding="utf-8"))["confirmed_failures"]

# 시장 케이스(해외 명시) vs 의도 케이스 분류 (라벨용)
MARKET = {"REG-002","REG-003","REG-004","REG-017","REG-018","REG-019","REG-020"}
INTENT = {"REG-007","REG-008","REG-009","REG-010","REG-011","REG-012"}

def run_once(use_hyde):
    passed, market_ok, intent_ok, fails = 0, 0, 0, []
    for c in cases:
        r = run_agent(query=c["query"], top_n=3, use_hyde=use_hyde, use_reranker=False)
        hits = r.get("search_results", [])
        top1 = hits[0]["menu_id"] if hits else ""
        ok = top1 in set(c.get("expected_menu_ids", []))
        cid = c["regression_case_id"]
        if ok:
            passed += 1
            if cid in MARKET: market_ok += 1
            if cid in INTENT: intent_ok += 1
        else:
            fails.append(cid)
    return passed, market_ok, intent_ok, fails

configs = [
    ("하드필터(기준)", False, None),
    ("소프트 w=0.02", True, 0.02),
    ("소프트 w=0.01", True, 0.01),
    ("소프트 w=0.005", True, 0.005),
    ("소프트 w=0.0",  True, 0.0),
]
print(f"{'설정':16} | 통과 | 시장{len(MARKET)} | 의도{len(INTENT)} | 실패목록")
print("-"*80)
for label, soft, w in configs:
    se.CATEGORY_SOFT_FILTER = soft
    if w is not None:
        se.CATEGORY_SOFT_BOOST_WEIGHT = w
    p, m, i, fails = run_once(use_hyde=False)
    print(f"{label:16} | {p:2}/29 | {m:2}/{len(MARKET)}  | {i:2}/{len(INTENT)}  | {','.join(fails)}")