Spaces:
Running
Running
File size: 6,349 Bytes
8485d6e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 | """ํ๋จ ๊ทผ๊ฑฐ ์ค๋ช
์์ฑ๊ธฐ (Rule-explainer-v1)
=================================================
๋ถ๋ฅ๊ธฐ๊ฐ ๋๋ ค์ค ๊ตฌ์กฐํ๋ reasons ๋ฅผ **์์ฐ์ด ์ค๋ช
**์ผ๋ก ๋ณํํ๋ค.
SPEC ยง1 ๊ธฐ๋ฅ 4 โ *"๋ถ๋ฅ ๊ฒฐ๊ณผ ์ค๋ช
"* ์ ๊ฐํ ๋ฒ์ .
LLM ์ ํธ์ถํ์ง ์๊ณ ํ
ํ๋ฆฟ/๋ฌธ์ฅ ์กฐ๋ฆฝ๋ง ์ฌ์ฉ โ ๊ฒฐ์ ์ ์ด๊ณ ๋น ๋ฅด๋ฉฐ ๋น์ฉ 0.
๋ค๋ง ์
๋ ฅ์ด ๊ทธ๋๋ก ์์ฐ์ด ๋ฌธ์ฅ์ผ๋ก ๋งคํ๋๋๋ก ์ถฉ๋ถํ ํ๋ถํ ๋ถ๊ธฐ๋ฅผ ๊ฐ๋๋ค.
"""
from __future__ import annotations
EXPLAINER_VERSION = "rule-explainer-v1"
# entity_type โ ํ๊ตญ์ด ์ค๋ช
ENTITY_DESCRIPTIONS = {
"KR_RRN": "์ฃผ๋ฏผ๋ฑ๋ก๋ฒํธ",
"KR_PASSPORT": "์ฌ๊ถ๋ฒํธ",
"KR_BIZ_NO": "์ฌ์
์๋ฑ๋ก๋ฒํธ",
"KR_PHONE": "ํ๊ตญ ์ ํ๋ฒํธ",
"KR_ADDRESS": "ํ๊ตญ ์ฃผ์",
"PHONE_NUMBER": "์ ํ๋ฒํธ",
"CREDIT_CARD": "์ ์ฉ์นด๋๋ฒํธ",
"US_SSN": "๋ฏธ๊ตญ SSN",
"IBAN_CODE": "IBAN ๊ณ์ข๋ฒํธ",
"AWS_ACCESS_KEY": "AWS ์ก์ธ์ค ํค",
"GENERIC_API_KEY": "API ํค ์ถ์ ํ ํฐ",
"VIP_NAMES": "VIP ๋ช
๋จ ์ด๋ฆ",
"INTERNAL_PROJECTS": "๋ด๋ถ ํ๋ก์ ํธ๋ช
",
"EMAIL_ADDRESS": "์ด๋ฉ์ผ ์ฃผ์",
"IP_ADDRESS": "IP ์ฃผ์",
"URL": "URL",
"PERSON": "์ธ๋ช
",
"LOCATION": "์ง๋ช
/์ฅ์",
"ORGANIZATION": "์กฐ์ง๋ช
",
"DATE_TIME": "๋ ์ง/์๊ฐ",
}
def _grade_label(g: str) -> str:
return {"C": "**์ํ (Critical)**", "S": "**๋ฏผ๊ฐ (Sensitive)**", "O": "**๊ณต๊ฐ (Open)**"}.get(g, g)
def _signal_phrase(reason: dict) -> str:
label = reason.get("label", "?")
cnt = reason.get("count", 1)
contrib = reason.get("contribution", 0)
if reason.get("kind") == "keyword":
return f"๋ฑ๊ธ ๋ผ๋ฒจ '{label}' {cnt}ํ ({contrib:+.2f}์ )"
desc = ENTITY_DESCRIPTIONS.get(label, label)
if cnt > 1:
return f"{desc} {cnt}๊ฑด ({contrib:+.2f}์ )"
return f"{desc} ({contrib:+.2f}์ )"
def explain(classification: dict, findings: list[dict] | None = None) -> dict:
"""classification + findings โ {summary, narrative, bullets, version}.
Returns:
summary: 1์ค ์์ฝ (๋ฑ๊ธ + ์ ์)
narrative: 2~5๋ฌธ์ฅ ์์ฐ์ด ์ค๋ช
(markdown bold ํฌํจ)
bullets: ์ฌ์ฉ์๊ฐ ๋น ๋ฅด๊ฒ ํ์ ์ ์๋ ํค ํฌ์ธํธ ๋ฆฌ์คํธ
version: "rule-explainer-v1"
"""
g = classification.get("grade", "O")
score = classification.get("score", 0.0)
conf = classification.get("confidence", 0.5)
th = classification.get("thresholds", {"C": 5.0, "S": 2.0})
reasons = classification.get("reasons") or []
entity_reasons = [r for r in reasons if r.get("kind") == "entity"]
kw_reasons = [r for r in reasons if r.get("kind") == "keyword"]
top = reasons[:3]
# ---- summary (ํ ์ค) ----
summary = f"{_grade_label(g)} โ score {score} (์ ๋ขฐ๋ {conf*100:.0f}%)"
# ---- narrative (๋ฌธ๋จ) ----
parts: list[str] = []
# 1) ๋ฑ๊ธ ๊ฒฐ์ ์ด์ + ๋ง์ง
if g == "C":
margin = score - th["C"]
parts.append(
f"์ด ๋ฌธ์๋ {_grade_label(g)} ๋ฑ๊ธ์ผ๋ก ๋ถ๋ฅ๋ฉ๋๋ค โ "
f"๋์ ์ ์ {score} ๊ฐ C ์๊ณ๊ฐ {th['C']} ๋ฅผ {margin:.2f}์ ์ด๊ณผํ์ต๋๋ค."
)
elif g == "S":
parts.append(
f"์ด ๋ฌธ์๋ {_grade_label(g)} ๋ฑ๊ธ์ผ๋ก ๋ถ๋ฅ๋ฉ๋๋ค โ "
f"์ ์ {score} ๊ฐ S ์๊ณ๊ฐ {th['S']} ์ C ์๊ณ๊ฐ {th['C']} ์ฌ์ด์ ์์นํฉ๋๋ค."
)
else:
parts.append(
f"์ด ๋ฌธ์๋ {_grade_label(g)} ๋ฑ๊ธ์ผ๋ก ๋ถ๋ฅ๋ฉ๋๋ค โ "
f"์ ์ {score} ๊ฐ S ์๊ณ๊ฐ {th['S']} ๋ฏธ๋ง์ผ๋ก, ๋ฑ๊ธ์ ์ฌ๋ฆด ๋งํ ์ ํธ๊ฐ ๋ถ์กฑํฉ๋๋ค."
)
# 2) ๊ฒฐ์ ์ ์ ํธ
if top:
phrases = [_signal_phrase(r) for r in top]
if len(phrases) == 1:
parts.append(f"๊ฒฐ์ ์ ์ ํธ๋ {phrases[0]} ๋จ ํ ๊ฐ์์ต๋๋ค.")
else:
parts.append("๊ฒฐ์ ์ ์ด์๋ ์ ํธ: " + ", ".join(phrases) + ".")
else:
parts.append("๋งค์นญ๋ ์ ํธ๊ฐ ์์ด ์ ์๊ฐ 0์ ๊ฐ๊น์ต๋๋ค.")
# 3) ์ ํธ ๊ตฌ์ฑ ๋ถ์
if kw_reasons and entity_reasons:
parts.append(
f"๋ฑ๊ธ ๋ผ๋ฒจ ํค์๋ {len(kw_reasons)}์ข
๊ณผ ์๋ณ์ {len(entity_reasons)}์ข
์ด ํจ๊ป ๋งค์นญ๋์ด "
f"๋ฑ๊ธ์ด ๋ ์์ ์ ์ผ๋ก ๊ฒฐ์ ๋์์ต๋๋ค."
)
elif kw_reasons and not entity_reasons:
parts.append(
"๋ณธ๋ฌธ์ ๋ช
์๋ ๋ฑ๊ธ ๋ผ๋ฒจ(์: ๋์ธ๋น/๊ธฐ๋ฐ) ๋ง์ผ๋ก ๊ฒฐ์ ๋์์ต๋๋ค โ "
"์ค์ ์๋ณ์๊ฐ ์์ ์๋ ์์ผ๋ ์ฌ์ฉ์ ๊ฒํ ๋ฅผ ๊ถ์ฅํฉ๋๋ค."
)
elif entity_reasons and not kw_reasons:
if g == "C":
parts.append("๋ฑ๊ธ ๋ผ๋ฒจ ํค์๋ ์์ด ์๋ณ์ ๊ฒ์ถ๋ง์ผ๋ก ์ํ ๋ฑ๊ธ์ด ํ์ ๋์์ต๋๋ค.")
elif g == "S":
parts.append("๊ฐ์ธ์ ๋ณด/๊ณ์ ์๋ณ์ ๊ฒ์ถ๋ก ๋ฏผ๊ฐ ๋ฑ๊ธ์ด ๋ถ์ฌ๋์์ต๋๋ค.")
# 4) ํต์ฌ PII ์์ฝ (์์ ๋๋ง)
pii_high = [r for r in entity_reasons if r.get("contribution", 0) >= 2.0]
if pii_high:
names = ", ".join(ENTITY_DESCRIPTIONS.get(r["label"], r["label"]) for r in pii_high)
parts.append(f"๊ณ ์ํ ์๋ณ์: {names}.")
# 5) ์ ๋ขฐ๋ ์ฝ๋ฉํธ
if conf < 0.62:
parts.append(
f"โ ์ ๋ขฐ๋ {conf*100:.0f}% โ ์๊ณ๊ฐ ๊ฒฝ๊ณ์ ๊ฐ๊น์ ์ฌ์ฉ์ ์ต์ข
ํ์ธ์ ๊ถ์ฅํฉ๋๋ค."
)
elif conf > 0.85:
parts.append(f"์ ๋ขฐ๋ {conf*100:.0f}% โ ๋ฑ๊ธ ๊ฒฝ๊ณ์์ ์ถฉ๋ถํ ๋จ์ด์ง ๋ช
ํํ ๋งค์นญ.")
narrative = " ".join(parts)
# ---- bullets (๋์๋ณด๋์ฉ) ----
bullets: list[str] = []
if top:
for r in top:
bullets.append(_signal_phrase(r))
bullets.append(f"์ ์ {score} (Sโฅ{th['S']} ยท Cโฅ{th['C']})")
bullets.append(f"์ ๋ขฐ๋ {conf*100:.0f}%")
if not entity_reasons and not kw_reasons:
bullets.append("๋งค์นญ๋ ์ ํธ ์์ โ ๊ธฐ๋ณธ๊ฐ(O)")
return {
"summary": summary,
"narrative": narrative,
"bullets": bullets,
"version": EXPLAINER_VERSION,
}
|