세션 평가와 교수자 분석 보강

apps/api/app/services/guardrail.py

@@ -44,6 +44,7 @@ _KOREAN_SURNAME_CHARS = (
 )
 _KOREAN_FULL_NAME = rf"[{_KOREAN_SURNAME_CHARS}][가-힣]{{1,3}}"
 _KOREAN_FULL_NAME_BEFORE_SUFFIX = rf"[{_KOREAN_SURNAME_CHARS}][가-힣]{{1,3}}?"
+_KOREAN_CONTEXTLESS_NAME = rf"[{_KOREAN_SURNAME_CHARS}][가-힣]{{2,3}}"
 _KOREAN_NAME_STOPWORDS = {
     "연락",
     "연락처",
@@ -126,7 +127,7 @@ _PII_PATTERNS: list[tuple[str, re.Pattern[str]]] = [
     (
         "NAME",
         re.compile(
-            rf"(?<![가-힣])(?P<value>{_KOREAN_FULL_NAME_BEFORE_SUFFIX})"
+            rf"(?<![가-힣])(?P<value>{_KOREAN_CONTEXTLESS_NAME})"
             r"(?P<suffix>(?:은|는|이|가|을|를|와|과|에게|한테|라고|이라는))"
         ),
     ),
@@ -372,6 +373,71 @@ _MEANS_TERMS = [
     "약을 얼마나", "몇 알", "몇 정", "투신", "목을 매", "목 매", "번개탄", "농약",
     "손목을", "칼로", "뛰어내리", "수면제 ", "독", "방법은", "어떻게 죽",
 ]
+_DISPLAY_PLACEHOLDER_LABELS = {
+    "NAME": "그 이름",
+    "ORG": "그 기관",
+    "PHONE": "연락처",
+    "EMAIL": "이메일",
+    "RRN": "식별번호",
+    "NUMID": "식별번호",
+    "DATE": "그 날짜",
+    "MONEY": "그 금액",
+    "ADDR": "그 주소",
+}
+_DISPLAY_PLACEHOLDER_RE = re.compile(
+    r"\[(?P<label>NAME|ORG|PHONE|EMAIL|RRN|NUMID|DATE|MONEY|ADDR)\]"
+    r"(?P<particle>[은는이가을를와과])?"
+)
+_DISPLAY_PLACEHOLDER_STREAM_TAIL = 16
+
+
+def _adjust_particle(value: str, particle: str | None) -> str:
+    if particle is None:
+        return value
+    if particle in ("은", "는"):
+        return value + "은"
+    if particle in ("이", "가"):
+        return value + "이"
+    if particle in ("을", "를"):
+        return value + "을"
+    if particle in ("와", "과"):
+        return value + "과"
+    return value + particle
+
+
+def humanize_pii_placeholders(text: str) -> str:
+    """사용자에게 보이는 내담자 응답에서 PII placeholder 토큰을 자연어로 낮춘다."""
+    if not text:
+        return text
+
+    def _replace(match: re.Match[str]) -> str:
+        label = match.group("label")
+        replacement = _DISPLAY_PLACEHOLDER_LABELS.get(label, "그 정보")
+        return _adjust_particle(replacement, match.group("particle"))
+
+    return _DISPLAY_PLACEHOLDER_RE.sub(_replace, text)
+
+
+class PiiPlaceholderStreamSanitizer:
+    """SSE 토큰 경계를 가로질러 나온 PII placeholder를 사용자 표시 전에 치환한다."""
+
+    def __init__(self) -> None:
+        self._tail = ""
+
+    def feed(self, chunk: str) -> str:
+        if not chunk:
+            return ""
+        self._tail += chunk
+        if len(self._tail) <= _DISPLAY_PLACEHOLDER_STREAM_TAIL:
+            return ""
+        ready = self._tail[:-_DISPLAY_PLACEHOLDER_STREAM_TAIL]
+        self._tail = self._tail[-_DISPLAY_PLACEHOLDER_STREAM_TAIL:]
+        return humanize_pii_placeholders(ready)
+
+    def flush(self) -> str:
+        tail = self._tail
+        self._tail = ""
+        return humanize_pii_placeholders(tail)
 
 
 @dataclass(slots=True)
@@ -406,7 +472,7 @@ def sanitize_client_reply(text: str, *, ideation_stage: int) -> OutputGuardResul
         blocked = True
 
     return OutputGuardResult(
-        text=text,
+        text=humanize_pii_placeholders(text),
         blocked=blocked,
         needs_regeneration=needs_regen,
         reasons=reasons,
@@ -440,6 +506,8 @@ __all__ = [
     "CrisisResult",
     "classify_crisis",
     "OutputGuardResult",
+    "PiiPlaceholderStreamSanitizer",
+    "humanize_pii_placeholders",
     "sanitize_client_reply",
     "clamp_ideation",
     "crisis_resource",

apps/api/app/services/orchestrator.py

@@ -237,6 +237,7 @@ async def run_turn_generate(
 
     # 5) 출력 가드레일 — 수단 차단 + ideation 상한
     guard = guardrail.sanitize_client_reply(reply, ideation_stage=st.ideation_stage)
+    reply = guard.text
     safety_flagged = guard.blocked or (ctx.crisis is not None and ctx.crisis.escalate)
     if guard.needs_regeneration:
         # 수단정보 누출 → 안전 대체 응답으로 치환(1차). 재생성 루프는 후속.
@@ -320,6 +321,7 @@ async def run_turn_stream(
     accumulated = ""
     flagged = False
     stream_meta: dict[str, Any] = {}
+    display_sanitizer = guardrail.PiiPlaceholderStreamSanitizer()
     if ctx.crisis is not None and ctx.crisis.escalate:
         flagged = True
         resource = guardrail.crisis_resource()
@@ -353,6 +355,9 @@ async def run_turn_stream(
             if packet.event == ENGINE_GATEWAY_SSE_ERROR:
                 payload = packet.payload
                 detail = payload.detail if isinstance(payload, StreamErrorEvent) else "engine stream error"
+                display_tail = display_sanitizer.flush()
+                if display_tail:
+                    yield StreamEvent("token", {"text": display_tail})
                 yield StreamEvent("error", {"detail": detail})
                 return
             if packet.event == ENGINE_GATEWAY_SSE_DONE:
@@ -376,7 +381,14 @@ async def run_turn_stream(
                 accumulated = "…(말을 잇지 못하고 잠시 침묵한다)"
                 break
 
-            yield StreamEvent("token", {"text": text_piece})
+            display_piece = display_sanitizer.feed(text_piece)
+            if display_piece:
+                yield StreamEvent("token", {"text": display_piece})
+
+        if not flagged:
+            display_tail = display_sanitizer.flush()
+            if display_tail:
+                yield StreamEvent("token", {"text": display_tail})
 
         latency_ms = int((time.perf_counter() - started) * 1000)
         await _record_llm_audit(
@@ -415,8 +427,14 @@ async def run_turn_stream(
             },
         )
     except EngineGatewaySseDecodeError as e:
+        display_tail = display_sanitizer.flush()
+        if display_tail:
+            yield StreamEvent("token", {"text": display_tail})
         yield StreamEvent("error", {"detail": str(e)})
     except EngineError as e:
+        display_tail = display_sanitizer.flush()
+        if display_tail:
+            yield StreamEvent("token", {"text": display_tail})
         yield StreamEvent("error", {"detail": str(e)})
 
 

apps/api/app/services/session_metrics.py

@@ -164,6 +164,7 @@ def build_learner_growth(
     *,
     learner_label: Callable[[str], str],
     limit: int | None = None,
+    point_limit: int | None = 6,
 ) -> list[LearnerGrowthMetrics]:
     grouped: dict[str, list[InProcSession]] = {}
     for sess in sessions:
@@ -222,7 +223,7 @@ def build_learner_growth(
                 avg_rapport=avg([value for value in rapport_values if value is not None]),
                 trend=trend,
                 top_techniques=top_techniques,
-                points=points[-6:],
+                points=points if point_limit is None else points[-point_limit:],
             )
         )
     sorted_result = sorted(result, key=lambda item: item.latest_at, reverse=True)